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 */
156 # define ESECT __attribute__ ((section("__TEXT,text_env")))
158 # define ESECT __attribute__ ((section("text_env")))
164 /** @defgroup internal LMDB Internals
167 /** @defgroup compat Compatibility Macros
168 * A bunch of macros to minimize the amount of platform-specific ifdefs
169 * needed throughout the rest of the code. When the features this library
170 * needs are similar enough to POSIX to be hidden in a one-or-two line
171 * replacement, this macro approach is used.
175 /* Features under development */
180 /** Wrapper around __func__, which is a C99 feature */
181 #if __STDC_VERSION__ >= 199901L
182 # define mdb_func_ __func__
183 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
184 # define mdb_func_ __FUNCTION__
186 /* If a debug message says <mdb_unknown>(), update the #if statements above */
187 # define mdb_func_ "<mdb_unknown>"
191 #define MDB_USE_HASH 1
192 #define MDB_PIDLOCK 0
193 #define THREAD_RET DWORD
194 #define pthread_t HANDLE
195 #define pthread_mutex_t HANDLE
196 #define pthread_cond_t HANDLE
197 #define pthread_key_t DWORD
198 #define pthread_self() GetCurrentThreadId()
199 #define pthread_key_create(x,y) \
200 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
201 #define pthread_key_delete(x) TlsFree(x)
202 #define pthread_getspecific(x) TlsGetValue(x)
203 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
204 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
205 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
206 #define pthread_cond_signal(x) SetEvent(*x)
207 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
208 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
209 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
210 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
211 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
212 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
213 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
214 #define getpid() GetCurrentProcessId()
215 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
216 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
217 #define ErrCode() GetLastError()
218 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
219 #define close(fd) (CloseHandle(fd) ? 0 : -1)
220 #define munmap(ptr,len) UnmapViewOfFile(ptr)
221 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
222 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
224 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
228 #define THREAD_RET void *
229 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
230 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
231 #define Z "z" /**< printf format modifier for size_t */
233 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
234 #define MDB_PIDLOCK 1
236 #ifdef MDB_USE_POSIX_SEM
238 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
239 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
240 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
241 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
244 mdb_sem_wait(sem_t *sem)
247 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
252 /** Lock the reader mutex.
254 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
255 /** Unlock the reader mutex.
257 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
259 /** Lock the writer mutex.
260 * Only a single write transaction is allowed at a time. Other writers
261 * will block waiting for this mutex.
263 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
264 /** Unlock the writer mutex.
266 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
267 #endif /* MDB_USE_POSIX_SEM */
269 /** Get the error code for the last failed system function.
271 #define ErrCode() errno
273 /** An abstraction for a file handle.
274 * On POSIX systems file handles are small integers. On Windows
275 * they're opaque pointers.
279 /** A value for an invalid file handle.
280 * Mainly used to initialize file variables and signify that they are
283 #define INVALID_HANDLE_VALUE (-1)
285 /** Get the size of a memory page for the system.
286 * This is the basic size that the platform's memory manager uses, and is
287 * fundamental to the use of memory-mapped files.
289 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
292 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
295 #define MNAME_LEN (sizeof(pthread_mutex_t))
301 /** A flag for opening a file and requesting synchronous data writes.
302 * This is only used when writing a meta page. It's not strictly needed;
303 * we could just do a normal write and then immediately perform a flush.
304 * But if this flag is available it saves us an extra system call.
306 * @note If O_DSYNC is undefined but exists in /usr/include,
307 * preferably set some compiler flag to get the definition.
308 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
311 # define MDB_DSYNC O_DSYNC
315 /** Function for flushing the data of a file. Define this to fsync
316 * if fdatasync() is not supported.
318 #ifndef MDB_FDATASYNC
319 # define MDB_FDATASYNC fdatasync
323 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
334 /** A page number in the database.
335 * Note that 64 bit page numbers are overkill, since pages themselves
336 * already represent 12-13 bits of addressable memory, and the OS will
337 * always limit applications to a maximum of 63 bits of address space.
339 * @note In the #MDB_node structure, we only store 48 bits of this value,
340 * which thus limits us to only 60 bits of addressable data.
342 typedef MDB_ID pgno_t;
344 /** A transaction ID.
345 * See struct MDB_txn.mt_txnid for details.
347 typedef MDB_ID txnid_t;
349 /** @defgroup debug Debug Macros
353 /** Enable debug output. Needs variable argument macros (a C99 feature).
354 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
355 * read from and written to the database (used for free space management).
361 static int mdb_debug;
362 static txnid_t mdb_debug_start;
364 /** Print a debug message with printf formatting.
365 * Requires double parenthesis around 2 or more args.
367 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
368 # define DPRINTF0(fmt, ...) \
369 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
371 # define DPRINTF(args) ((void) 0)
373 /** Print a debug string.
374 * The string is printed literally, with no format processing.
376 #define DPUTS(arg) DPRINTF(("%s", arg))
377 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
379 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
382 /** @brief The maximum size of a database page.
384 * It is 32k or 64k, since value-PAGEBASE must fit in
385 * #MDB_page.%mp_upper.
387 * LMDB will use database pages < OS pages if needed.
388 * That causes more I/O in write transactions: The OS must
389 * know (read) the whole page before writing a partial page.
391 * Note that we don't currently support Huge pages. On Linux,
392 * regular data files cannot use Huge pages, and in general
393 * Huge pages aren't actually pageable. We rely on the OS
394 * demand-pager to read our data and page it out when memory
395 * pressure from other processes is high. So until OSs have
396 * actual paging support for Huge pages, they're not viable.
398 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
400 /** The minimum number of keys required in a database page.
401 * Setting this to a larger value will place a smaller bound on the
402 * maximum size of a data item. Data items larger than this size will
403 * be pushed into overflow pages instead of being stored directly in
404 * the B-tree node. This value used to default to 4. With a page size
405 * of 4096 bytes that meant that any item larger than 1024 bytes would
406 * go into an overflow page. That also meant that on average 2-3KB of
407 * each overflow page was wasted space. The value cannot be lower than
408 * 2 because then there would no longer be a tree structure. With this
409 * value, items larger than 2KB will go into overflow pages, and on
410 * average only 1KB will be wasted.
412 #define MDB_MINKEYS 2
414 /** A stamp that identifies a file as an LMDB file.
415 * There's nothing special about this value other than that it is easily
416 * recognizable, and it will reflect any byte order mismatches.
418 #define MDB_MAGIC 0xBEEFC0DE
420 /** The version number for a database's datafile format. */
421 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
422 /** The version number for a database's lockfile format. */
423 #define MDB_LOCK_VERSION 1
425 /** @brief The max size of a key we can write, or 0 for dynamic max.
427 * Define this as 0 to compute the max from the page size. 511
428 * is default for backwards compat: liblmdb <= 0.9.10 can break
429 * when modifying a DB with keys/dupsort data bigger than its max.
431 * Data items in an #MDB_DUPSORT database are also limited to
432 * this size, since they're actually keys of a sub-DB. Keys and
433 * #MDB_DUPSORT data items must fit on a node in a regular page.
435 #ifndef MDB_MAXKEYSIZE
436 #define MDB_MAXKEYSIZE 511
439 /** The maximum size of a key we can write to the environment. */
441 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
443 #define ENV_MAXKEY(env) ((env)->me_maxkey)
446 /** @brief The maximum size of a data item.
448 * We only store a 32 bit value for node sizes.
450 #define MAXDATASIZE 0xffffffffUL
453 /** Key size which fits in a #DKBUF.
456 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
459 * This is used for printing a hex dump of a key's contents.
461 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
462 /** Display a key in hex.
464 * Invoke a function to display a key in hex.
466 #define DKEY(x) mdb_dkey(x, kbuf)
472 /** An invalid page number.
473 * Mainly used to denote an empty tree.
475 #define P_INVALID (~(pgno_t)0)
477 /** Test if the flags \b f are set in a flag word \b w. */
478 #define F_ISSET(w, f) (((w) & (f)) == (f))
480 /** Round \b n up to an even number. */
481 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
483 /** Used for offsets within a single page.
484 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
487 typedef uint16_t indx_t;
489 /** Default size of memory map.
490 * This is certainly too small for any actual applications. Apps should always set
491 * the size explicitly using #mdb_env_set_mapsize().
493 #define DEFAULT_MAPSIZE 1048576
495 /** @defgroup readers Reader Lock Table
496 * Readers don't acquire any locks for their data access. Instead, they
497 * simply record their transaction ID in the reader table. The reader
498 * mutex is needed just to find an empty slot in the reader table. The
499 * slot's address is saved in thread-specific data so that subsequent read
500 * transactions started by the same thread need no further locking to proceed.
502 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
504 * No reader table is used if the database is on a read-only filesystem, or
505 * if #MDB_NOLOCK is set.
507 * Since the database uses multi-version concurrency control, readers don't
508 * actually need any locking. This table is used to keep track of which
509 * readers are using data from which old transactions, so that we'll know
510 * when a particular old transaction is no longer in use. Old transactions
511 * that have discarded any data pages can then have those pages reclaimed
512 * for use by a later write transaction.
514 * The lock table is constructed such that reader slots are aligned with the
515 * processor's cache line size. Any slot is only ever used by one thread.
516 * This alignment guarantees that there will be no contention or cache
517 * thrashing as threads update their own slot info, and also eliminates
518 * any need for locking when accessing a slot.
520 * A writer thread will scan every slot in the table to determine the oldest
521 * outstanding reader transaction. Any freed pages older than this will be
522 * reclaimed by the writer. The writer doesn't use any locks when scanning
523 * this table. This means that there's no guarantee that the writer will
524 * see the most up-to-date reader info, but that's not required for correct
525 * operation - all we need is to know the upper bound on the oldest reader,
526 * we don't care at all about the newest reader. So the only consequence of
527 * reading stale information here is that old pages might hang around a
528 * while longer before being reclaimed. That's actually good anyway, because
529 * the longer we delay reclaiming old pages, the more likely it is that a
530 * string of contiguous pages can be found after coalescing old pages from
531 * many old transactions together.
534 /** Number of slots in the reader table.
535 * This value was chosen somewhat arbitrarily. 126 readers plus a
536 * couple mutexes fit exactly into 8KB on my development machine.
537 * Applications should set the table size using #mdb_env_set_maxreaders().
539 #define DEFAULT_READERS 126
541 /** The size of a CPU cache line in bytes. We want our lock structures
542 * aligned to this size to avoid false cache line sharing in the
544 * This value works for most CPUs. For Itanium this should be 128.
550 /** The information we store in a single slot of the reader table.
551 * In addition to a transaction ID, we also record the process and
552 * thread ID that owns a slot, so that we can detect stale information,
553 * e.g. threads or processes that went away without cleaning up.
554 * @note We currently don't check for stale records. We simply re-init
555 * the table when we know that we're the only process opening the
558 typedef struct MDB_rxbody {
559 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
560 * Multiple readers that start at the same time will probably have the
561 * same ID here. Again, it's not important to exclude them from
562 * anything; all we need to know is which version of the DB they
563 * started from so we can avoid overwriting any data used in that
564 * particular version.
567 /** The process ID of the process owning this reader txn. */
569 /** The thread ID of the thread owning this txn. */
573 /** The actual reader record, with cacheline padding. */
574 typedef struct MDB_reader {
577 /** shorthand for mrb_txnid */
578 #define mr_txnid mru.mrx.mrb_txnid
579 #define mr_pid mru.mrx.mrb_pid
580 #define mr_tid mru.mrx.mrb_tid
581 /** cache line alignment */
582 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
586 /** The header for the reader table.
587 * The table resides in a memory-mapped file. (This is a different file
588 * than is used for the main database.)
590 * For POSIX the actual mutexes reside in the shared memory of this
591 * mapped file. On Windows, mutexes are named objects allocated by the
592 * kernel; we store the mutex names in this mapped file so that other
593 * processes can grab them. This same approach is also used on
594 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
595 * process-shared POSIX mutexes. For these cases where a named object
596 * is used, the object name is derived from a 64 bit FNV hash of the
597 * environment pathname. As such, naming collisions are extremely
598 * unlikely. If a collision occurs, the results are unpredictable.
600 typedef struct MDB_txbody {
601 /** Stamp identifying this as an LMDB file. It must be set
604 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
606 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
607 char mtb_rmname[MNAME_LEN];
609 /** Mutex protecting access to this table.
610 * This is the reader lock that #LOCK_MUTEX_R acquires.
612 pthread_mutex_t mtb_mutex;
614 /** The ID of the last transaction committed to the database.
615 * This is recorded here only for convenience; the value can always
616 * be determined by reading the main database meta pages.
619 /** The number of slots that have been used in the reader table.
620 * This always records the maximum count, it is not decremented
621 * when readers release their slots.
623 unsigned mtb_numreaders;
626 /** The actual reader table definition. */
627 typedef struct MDB_txninfo {
630 #define mti_magic mt1.mtb.mtb_magic
631 #define mti_format mt1.mtb.mtb_format
632 #define mti_mutex mt1.mtb.mtb_mutex
633 #define mti_rmname mt1.mtb.mtb_rmname
634 #define mti_txnid mt1.mtb.mtb_txnid
635 #define mti_numreaders mt1.mtb.mtb_numreaders
636 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
639 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
640 char mt2_wmname[MNAME_LEN];
641 #define mti_wmname mt2.mt2_wmname
643 pthread_mutex_t mt2_wmutex;
644 #define mti_wmutex mt2.mt2_wmutex
646 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
648 MDB_reader mti_readers[1];
651 /** Lockfile format signature: version, features and field layout */
652 #define MDB_LOCK_FORMAT \
654 ((MDB_LOCK_VERSION) \
655 /* Flags which describe functionality */ \
656 + (((MDB_PIDLOCK) != 0) << 16)))
659 /** Common header for all page types.
660 * Overflow records occupy a number of contiguous pages with no
661 * headers on any page after the first.
663 typedef struct MDB_page {
664 #define mp_pgno mp_p.p_pgno
665 #define mp_next mp_p.p_next
667 pgno_t p_pgno; /**< page number */
668 struct MDB_page *p_next; /**< for in-memory list of freed pages */
671 /** @defgroup mdb_page Page Flags
673 * Flags for the page headers.
676 #define P_BRANCH 0x01 /**< branch page */
677 #define P_LEAF 0x02 /**< leaf page */
678 #define P_OVERFLOW 0x04 /**< overflow page */
679 #define P_META 0x08 /**< meta page */
680 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
681 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
682 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
683 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
684 #define P_KEEP 0x8000 /**< leave this page alone during spill */
686 uint16_t mp_flags; /**< @ref mdb_page */
687 #define mp_lower mp_pb.pb.pb_lower
688 #define mp_upper mp_pb.pb.pb_upper
689 #define mp_pages mp_pb.pb_pages
692 indx_t pb_lower; /**< lower bound of free space */
693 indx_t pb_upper; /**< upper bound of free space */
695 uint32_t pb_pages; /**< number of overflow pages */
697 indx_t mp_ptrs[1]; /**< dynamic size */
700 /** Size of the page header, excluding dynamic data at the end */
701 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
703 /** Address of first usable data byte in a page, after the header */
704 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
706 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
707 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
709 /** Number of nodes on a page */
710 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
712 /** The amount of space remaining in the page */
713 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
715 /** The percentage of space used in the page, in tenths of a percent. */
716 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
717 ((env)->me_psize - PAGEHDRSZ))
718 /** The minimum page fill factor, in tenths of a percent.
719 * Pages emptier than this are candidates for merging.
721 #define FILL_THRESHOLD 250
723 /** Test if a page is a leaf page */
724 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
725 /** Test if a page is a LEAF2 page */
726 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
727 /** Test if a page is a branch page */
728 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
729 /** Test if a page is an overflow page */
730 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
731 /** Test if a page is a sub page */
732 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
734 /** The number of overflow pages needed to store the given size. */
735 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
737 /** Link in #MDB_txn.%mt_loose_pages list */
738 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
740 /** Header for a single key/data pair within a page.
741 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
742 * We guarantee 2-byte alignment for 'MDB_node's.
744 typedef struct MDB_node {
745 /** lo and hi are used for data size on leaf nodes and for
746 * child pgno on branch nodes. On 64 bit platforms, flags
747 * is also used for pgno. (Branch nodes have no flags).
748 * They are in host byte order in case that lets some
749 * accesses be optimized into a 32-bit word access.
751 #if BYTE_ORDER == LITTLE_ENDIAN
752 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
754 unsigned short mn_hi, mn_lo;
756 /** @defgroup mdb_node Node Flags
758 * Flags for node headers.
761 #define F_BIGDATA 0x01 /**< data put on overflow page */
762 #define F_SUBDATA 0x02 /**< data is a sub-database */
763 #define F_DUPDATA 0x04 /**< data has duplicates */
765 /** valid flags for #mdb_node_add() */
766 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
769 unsigned short mn_flags; /**< @ref mdb_node */
770 unsigned short mn_ksize; /**< key size */
771 char mn_data[1]; /**< key and data are appended here */
774 /** Size of the node header, excluding dynamic data at the end */
775 #define NODESIZE offsetof(MDB_node, mn_data)
777 /** Bit position of top word in page number, for shifting mn_flags */
778 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
780 /** Size of a node in a branch page with a given key.
781 * This is just the node header plus the key, there is no data.
783 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
785 /** Size of a node in a leaf page with a given key and data.
786 * This is node header plus key plus data size.
788 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
790 /** Address of node \b i in page \b p */
791 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
793 /** Address of the key for the node */
794 #define NODEKEY(node) (void *)((node)->mn_data)
796 /** Address of the data for a node */
797 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
799 /** Get the page number pointed to by a branch node */
800 #define NODEPGNO(node) \
801 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
802 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
803 /** Set the page number in a branch node */
804 #define SETPGNO(node,pgno) do { \
805 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
806 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
808 /** Get the size of the data in a leaf node */
809 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
810 /** Set the size of the data for a leaf node */
811 #define SETDSZ(node,size) do { \
812 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
813 /** The size of a key in a node */
814 #define NODEKSZ(node) ((node)->mn_ksize)
816 /** Copy a page number from src to dst */
818 #define COPY_PGNO(dst,src) dst = src
820 #if SIZE_MAX > 4294967295UL
821 #define COPY_PGNO(dst,src) do { \
822 unsigned short *s, *d; \
823 s = (unsigned short *)&(src); \
824 d = (unsigned short *)&(dst); \
831 #define COPY_PGNO(dst,src) do { \
832 unsigned short *s, *d; \
833 s = (unsigned short *)&(src); \
834 d = (unsigned short *)&(dst); \
840 /** The address of a key in a LEAF2 page.
841 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
842 * There are no node headers, keys are stored contiguously.
844 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
846 /** Set the \b node's key into \b keyptr, if requested. */
847 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
848 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
850 /** Set the \b node's key into \b key. */
851 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
853 /** Information about a single database in the environment. */
854 typedef struct MDB_db {
855 uint32_t md_pad; /**< also ksize for LEAF2 pages */
856 uint16_t md_flags; /**< @ref mdb_dbi_open */
857 uint16_t md_depth; /**< depth of this tree */
858 pgno_t md_branch_pages; /**< number of internal pages */
859 pgno_t md_leaf_pages; /**< number of leaf pages */
860 pgno_t md_overflow_pages; /**< number of overflow pages */
861 size_t md_entries; /**< number of data items */
862 pgno_t md_root; /**< the root page of this tree */
865 /** mdb_dbi_open flags */
866 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
867 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
868 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
869 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
871 /** Handle for the DB used to track free pages. */
873 /** Handle for the default DB. */
876 /** Meta page content.
877 * A meta page is the start point for accessing a database snapshot.
878 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
880 typedef struct MDB_meta {
881 /** Stamp identifying this as an LMDB file. It must be set
884 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
886 void *mm_address; /**< address for fixed mapping */
887 size_t mm_mapsize; /**< size of mmap region */
888 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
889 /** The size of pages used in this DB */
890 #define mm_psize mm_dbs[0].md_pad
891 /** Any persistent environment flags. @ref mdb_env */
892 #define mm_flags mm_dbs[0].md_flags
893 pgno_t mm_last_pg; /**< last used page in file */
894 txnid_t mm_txnid; /**< txnid that committed this page */
897 /** Buffer for a stack-allocated meta page.
898 * The members define size and alignment, and silence type
899 * aliasing warnings. They are not used directly; that could
900 * mean incorrectly using several union members in parallel.
902 typedef union MDB_metabuf {
905 char mm_pad[PAGEHDRSZ];
910 /** Auxiliary DB info.
911 * The information here is mostly static/read-only. There is
912 * only a single copy of this record in the environment.
914 typedef struct MDB_dbx {
915 MDB_val md_name; /**< name of the database */
916 MDB_cmp_func *md_cmp; /**< function for comparing keys */
917 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
918 MDB_rel_func *md_rel; /**< user relocate function */
919 void *md_relctx; /**< user-provided context for md_rel */
922 /** A database transaction.
923 * Every operation requires a transaction handle.
926 MDB_txn *mt_parent; /**< parent of a nested txn */
927 MDB_txn *mt_child; /**< nested txn under this txn */
928 pgno_t mt_next_pgno; /**< next unallocated page */
929 /** The ID of this transaction. IDs are integers incrementing from 1.
930 * Only committed write transactions increment the ID. If a transaction
931 * aborts, the ID may be re-used by the next writer.
934 MDB_env *mt_env; /**< the DB environment */
935 /** The list of pages that became unused during this transaction.
938 /** The list of loose pages that became unused and may be reused
939 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
941 MDB_page *mt_loose_pgs;
942 /** The sorted list of dirty pages we temporarily wrote to disk
943 * because the dirty list was full. page numbers in here are
944 * shifted left by 1, deleted slots have the LSB set.
946 MDB_IDL mt_spill_pgs;
948 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
950 /** For read txns: This thread/txn's reader table slot, or NULL. */
953 /** Array of records for each DB known in the environment. */
955 /** Array of MDB_db records for each known DB */
957 /** Array of sequence numbers for each DB handle */
958 unsigned int *mt_dbiseqs;
959 /** @defgroup mt_dbflag Transaction DB Flags
963 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
964 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
965 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
966 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
968 /** In write txns, array of cursors for each DB */
969 MDB_cursor **mt_cursors;
970 /** Array of flags for each DB */
971 unsigned char *mt_dbflags;
972 /** Number of DB records in use. This number only ever increments;
973 * we don't decrement it when individual DB handles are closed.
977 /** @defgroup mdb_txn Transaction Flags
981 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
982 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
983 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
984 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
986 unsigned int mt_flags; /**< @ref mdb_txn */
987 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
988 * Includes ancestor txns' dirty pages not hidden by other txns'
989 * dirty/spilled pages. Thus commit(nested txn) has room to merge
990 * dirty_list into mt_parent after freeing hidden mt_parent pages.
992 unsigned int mt_dirty_room;
995 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
996 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
997 * raise this on a 64 bit machine.
999 #define CURSOR_STACK 32
1003 /** Cursors are used for all DB operations.
1004 * A cursor holds a path of (page pointer, key index) from the DB
1005 * root to a position in the DB, plus other state. #MDB_DUPSORT
1006 * cursors include an xcursor to the current data item. Write txns
1007 * track their cursors and keep them up to date when data moves.
1008 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1009 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1012 /** Next cursor on this DB in this txn */
1013 MDB_cursor *mc_next;
1014 /** Backup of the original cursor if this cursor is a shadow */
1015 MDB_cursor *mc_backup;
1016 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1017 struct MDB_xcursor *mc_xcursor;
1018 /** The transaction that owns this cursor */
1020 /** The database handle this cursor operates on */
1022 /** The database record for this cursor */
1024 /** The database auxiliary record for this cursor */
1026 /** The @ref mt_dbflag for this database */
1027 unsigned char *mc_dbflag;
1028 unsigned short mc_snum; /**< number of pushed pages */
1029 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1030 /** @defgroup mdb_cursor Cursor Flags
1032 * Cursor state flags.
1035 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1036 #define C_EOF 0x02 /**< No more data */
1037 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1038 #define C_DEL 0x08 /**< last op was a cursor_del */
1039 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1040 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1042 unsigned int mc_flags; /**< @ref mdb_cursor */
1043 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1044 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1047 /** Context for sorted-dup records.
1048 * We could have gone to a fully recursive design, with arbitrarily
1049 * deep nesting of sub-databases. But for now we only handle these
1050 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1052 typedef struct MDB_xcursor {
1053 /** A sub-cursor for traversing the Dup DB */
1054 MDB_cursor mx_cursor;
1055 /** The database record for this Dup DB */
1057 /** The auxiliary DB record for this Dup DB */
1059 /** The @ref mt_dbflag for this Dup DB */
1060 unsigned char mx_dbflag;
1063 /** State of FreeDB old pages, stored in the MDB_env */
1064 typedef struct MDB_pgstate {
1065 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1066 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1069 /** The database environment. */
1071 HANDLE me_fd; /**< The main data file */
1072 HANDLE me_lfd; /**< The lock file */
1073 HANDLE me_mfd; /**< just for writing the meta pages */
1074 /** Failed to update the meta page. Probably an I/O error. */
1075 #define MDB_FATAL_ERROR 0x80000000U
1076 /** We're explicitly changing the mapsize. */
1077 #define MDB_RESIZING 0x40000000U
1078 /** Some fields are initialized. */
1079 #define MDB_ENV_ACTIVE 0x20000000U
1080 /** me_txkey is set */
1081 #define MDB_ENV_TXKEY 0x10000000U
1082 uint32_t me_flags; /**< @ref mdb_env */
1083 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1084 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1085 unsigned int me_maxreaders; /**< size of the reader table */
1086 unsigned int me_numreaders; /**< max numreaders set by this env */
1087 MDB_dbi me_numdbs; /**< number of DBs opened */
1088 MDB_dbi me_maxdbs; /**< size of the DB table */
1089 MDB_PID_T me_pid; /**< process ID of this env */
1090 char *me_path; /**< path to the DB files */
1091 char *me_map; /**< the memory map of the data file */
1092 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1093 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1094 void *me_pbuf; /**< scratch area for DUPSORT put() */
1095 MDB_txn *me_txn; /**< current write transaction */
1096 size_t me_mapsize; /**< size of the data memory map */
1097 off_t me_size; /**< current file size */
1098 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1099 MDB_dbx *me_dbxs; /**< array of static DB info */
1100 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1101 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1102 pthread_key_t me_txkey; /**< thread-key for readers */
1103 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1104 # define me_pglast me_pgstate.mf_pglast
1105 # define me_pghead me_pgstate.mf_pghead
1106 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1107 /** IDL of pages that became unused in a write txn */
1108 MDB_IDL me_free_pgs;
1109 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1110 MDB_ID2L me_dirty_list;
1111 /** Max number of freelist items that can fit in a single overflow page */
1113 /** Max size of a node on a page */
1114 unsigned int me_nodemax;
1115 #if !(MDB_MAXKEYSIZE)
1116 unsigned int me_maxkey; /**< max size of a key */
1118 int me_live_reader; /**< have liveness lock in reader table */
1120 int me_pidquery; /**< Used in OpenProcess */
1121 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1123 #elif defined(MDB_USE_POSIX_SEM)
1124 sem_t *me_rmutex; /* Shared mutexes are not supported */
1127 void *me_userctx; /**< User-settable context */
1128 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1131 /** Nested transaction */
1132 typedef struct MDB_ntxn {
1133 MDB_txn mnt_txn; /**< the transaction */
1134 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1137 /** max number of pages to commit in one writev() call */
1138 #define MDB_COMMIT_PAGES 64
1139 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1140 #undef MDB_COMMIT_PAGES
1141 #define MDB_COMMIT_PAGES IOV_MAX
1144 /** max bytes to write in one call */
1145 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1147 /** Check \b txn and \b dbi arguments to a function */
1148 #define TXN_DBI_EXIST(txn, dbi) \
1149 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1151 /** Check for misused \b dbi handles */
1152 #define TXN_DBI_CHANGED(txn, dbi) \
1153 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1155 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1156 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1157 static int mdb_page_touch(MDB_cursor *mc);
1159 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1160 static int mdb_page_search_root(MDB_cursor *mc,
1161 MDB_val *key, int modify);
1162 #define MDB_PS_MODIFY 1
1163 #define MDB_PS_ROOTONLY 2
1164 #define MDB_PS_FIRST 4
1165 #define MDB_PS_LAST 8
1166 static int mdb_page_search(MDB_cursor *mc,
1167 MDB_val *key, int flags);
1168 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1170 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1171 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1172 pgno_t newpgno, unsigned int nflags);
1174 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1175 static int mdb_env_pick_meta(const MDB_env *env);
1176 static int mdb_env_write_meta(MDB_txn *txn);
1177 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1178 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1180 static void mdb_env_close0(MDB_env *env, int excl);
1182 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1183 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1184 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1185 static void mdb_node_del(MDB_cursor *mc, int ksize);
1186 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1187 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1188 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1189 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1190 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1192 static int mdb_rebalance(MDB_cursor *mc);
1193 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1195 static void mdb_cursor_pop(MDB_cursor *mc);
1196 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1198 static int mdb_cursor_del0(MDB_cursor *mc);
1199 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1200 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1201 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1202 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1203 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1205 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1206 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1208 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1209 static void mdb_xcursor_init0(MDB_cursor *mc);
1210 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1212 static int mdb_drop0(MDB_cursor *mc, int subs);
1213 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1216 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1220 static SECURITY_DESCRIPTOR mdb_null_sd;
1221 static SECURITY_ATTRIBUTES mdb_all_sa;
1222 static int mdb_sec_inited;
1225 /** Return the library version info. */
1227 mdb_version(int *major, int *minor, int *patch)
1229 if (major) *major = MDB_VERSION_MAJOR;
1230 if (minor) *minor = MDB_VERSION_MINOR;
1231 if (patch) *patch = MDB_VERSION_PATCH;
1232 return MDB_VERSION_STRING;
1235 /** Table of descriptions for LMDB @ref errors */
1236 static char *const mdb_errstr[] = {
1237 "MDB_KEYEXIST: Key/data pair already exists",
1238 "MDB_NOTFOUND: No matching key/data pair found",
1239 "MDB_PAGE_NOTFOUND: Requested page not found",
1240 "MDB_CORRUPTED: Located page was wrong type",
1241 "MDB_PANIC: Update of meta page failed",
1242 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1243 "MDB_INVALID: File is not an LMDB file",
1244 "MDB_MAP_FULL: Environment mapsize limit reached",
1245 "MDB_DBS_FULL: Environment maxdbs limit reached",
1246 "MDB_READERS_FULL: Environment maxreaders limit reached",
1247 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1248 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1249 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1250 "MDB_PAGE_FULL: Internal error - page has no more space",
1251 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1252 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1253 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1254 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1255 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1256 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1260 mdb_strerror(int err)
1264 return ("Successful return: 0");
1266 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1267 i = err - MDB_KEYEXIST;
1268 return mdb_errstr[i];
1271 return strerror(err);
1274 /** assert(3) variant in cursor context */
1275 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1276 /** assert(3) variant in transaction context */
1277 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1278 /** assert(3) variant in environment context */
1279 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1282 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1283 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1286 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1287 const char *func, const char *file, int line)
1290 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1291 file, line, expr_txt, func);
1292 if (env->me_assert_func)
1293 env->me_assert_func(env, buf);
1294 fprintf(stderr, "%s\n", buf);
1298 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1302 /** Return the page number of \b mp which may be sub-page, for debug output */
1304 mdb_dbg_pgno(MDB_page *mp)
1307 COPY_PGNO(ret, mp->mp_pgno);
1311 /** Display a key in hexadecimal and return the address of the result.
1312 * @param[in] key the key to display
1313 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1314 * @return The key in hexadecimal form.
1317 mdb_dkey(MDB_val *key, char *buf)
1320 unsigned char *c = key->mv_data;
1326 if (key->mv_size > DKBUF_MAXKEYSIZE)
1327 return "MDB_MAXKEYSIZE";
1328 /* may want to make this a dynamic check: if the key is mostly
1329 * printable characters, print it as-is instead of converting to hex.
1333 for (i=0; i<key->mv_size; i++)
1334 ptr += sprintf(ptr, "%02x", *c++);
1336 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1342 mdb_leafnode_type(MDB_node *n)
1344 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1345 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1346 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1349 /** Display all the keys in the page. */
1351 mdb_page_list(MDB_page *mp)
1353 pgno_t pgno = mdb_dbg_pgno(mp);
1354 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1356 unsigned int i, nkeys, nsize, total = 0;
1360 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1361 case P_BRANCH: type = "Branch page"; break;
1362 case P_LEAF: type = "Leaf page"; break;
1363 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1364 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1365 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1367 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1368 pgno, mp->mp_pages, state);
1371 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1372 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1375 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1379 nkeys = NUMKEYS(mp);
1380 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1382 for (i=0; i<nkeys; i++) {
1383 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1384 key.mv_size = nsize = mp->mp_pad;
1385 key.mv_data = LEAF2KEY(mp, i, nsize);
1387 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1390 node = NODEPTR(mp, i);
1391 key.mv_size = node->mn_ksize;
1392 key.mv_data = node->mn_data;
1393 nsize = NODESIZE + key.mv_size;
1394 if (IS_BRANCH(mp)) {
1395 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1399 if (F_ISSET(node->mn_flags, F_BIGDATA))
1400 nsize += sizeof(pgno_t);
1402 nsize += NODEDSZ(node);
1404 nsize += sizeof(indx_t);
1405 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1406 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1408 total = EVEN(total);
1410 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1411 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1415 mdb_cursor_chk(MDB_cursor *mc)
1421 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1422 for (i=0; i<mc->mc_top; i++) {
1424 node = NODEPTR(mp, mc->mc_ki[i]);
1425 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1428 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1434 /** Count all the pages in each DB and in the freelist
1435 * and make sure it matches the actual number of pages
1437 * All named DBs must be open for a correct count.
1439 static void mdb_audit(MDB_txn *txn)
1443 MDB_ID freecount, count;
1448 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1449 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1450 freecount += *(MDB_ID *)data.mv_data;
1451 mdb_tassert(txn, rc == MDB_NOTFOUND);
1454 for (i = 0; i<txn->mt_numdbs; i++) {
1456 if (!(txn->mt_dbflags[i] & DB_VALID))
1458 mdb_cursor_init(&mc, txn, i, &mx);
1459 if (txn->mt_dbs[i].md_root == P_INVALID)
1461 count += txn->mt_dbs[i].md_branch_pages +
1462 txn->mt_dbs[i].md_leaf_pages +
1463 txn->mt_dbs[i].md_overflow_pages;
1464 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1465 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1466 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1469 mp = mc.mc_pg[mc.mc_top];
1470 for (j=0; j<NUMKEYS(mp); j++) {
1471 MDB_node *leaf = NODEPTR(mp, j);
1472 if (leaf->mn_flags & F_SUBDATA) {
1474 memcpy(&db, NODEDATA(leaf), sizeof(db));
1475 count += db.md_branch_pages + db.md_leaf_pages +
1476 db.md_overflow_pages;
1480 mdb_tassert(txn, rc == MDB_NOTFOUND);
1483 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1484 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1485 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1491 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1493 return txn->mt_dbxs[dbi].md_cmp(a, b);
1497 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1499 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1502 /** Allocate memory for a page.
1503 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1506 mdb_page_malloc(MDB_txn *txn, unsigned num)
1508 MDB_env *env = txn->mt_env;
1509 MDB_page *ret = env->me_dpages;
1510 size_t psize = env->me_psize, sz = psize, off;
1511 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1512 * For a single page alloc, we init everything after the page header.
1513 * For multi-page, we init the final page; if the caller needed that
1514 * many pages they will be filling in at least up to the last page.
1518 VGMEMP_ALLOC(env, ret, sz);
1519 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1520 env->me_dpages = ret->mp_next;
1523 psize -= off = PAGEHDRSZ;
1528 if ((ret = malloc(sz)) != NULL) {
1529 VGMEMP_ALLOC(env, ret, sz);
1530 if (!(env->me_flags & MDB_NOMEMINIT)) {
1531 memset((char *)ret + off, 0, psize);
1535 txn->mt_flags |= MDB_TXN_ERROR;
1539 /** Free a single page.
1540 * Saves single pages to a list, for future reuse.
1541 * (This is not used for multi-page overflow pages.)
1544 mdb_page_free(MDB_env *env, MDB_page *mp)
1546 mp->mp_next = env->me_dpages;
1547 VGMEMP_FREE(env, mp);
1548 env->me_dpages = mp;
1551 /** Free a dirty page */
1553 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1555 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1556 mdb_page_free(env, dp);
1558 /* large pages just get freed directly */
1559 VGMEMP_FREE(env, dp);
1564 /** Return all dirty pages to dpage list */
1566 mdb_dlist_free(MDB_txn *txn)
1568 MDB_env *env = txn->mt_env;
1569 MDB_ID2L dl = txn->mt_u.dirty_list;
1570 unsigned i, n = dl[0].mid;
1572 for (i = 1; i <= n; i++) {
1573 mdb_dpage_free(env, dl[i].mptr);
1578 /** Loosen or free a single page.
1579 * Saves single pages to a list for future reuse
1580 * in this same txn. It has been pulled from the freeDB
1581 * and already resides on the dirty list, but has been
1582 * deleted. Use these pages first before pulling again
1585 * If the page wasn't dirtied in this txn, just add it
1586 * to this txn's free list.
1589 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1592 pgno_t pgno = mp->mp_pgno;
1594 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1595 if (mc->mc_txn->mt_parent) {
1596 MDB_ID2 *dl = mc->mc_txn->mt_u.dirty_list;
1597 /* If txn has a parent, make sure the page is in our
1601 unsigned x = mdb_mid2l_search(dl, pgno);
1602 if (x <= dl[0].mid && dl[x].mid == pgno) {
1603 if (mp != dl[x].mptr) { /* bad cursor? */
1604 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1605 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
1606 return MDB_CORRUPTED;
1613 /* no parent txn, so it's just ours */
1618 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1620 NEXT_LOOSE_PAGE(mp) = mc->mc_txn->mt_loose_pgs;
1621 mc->mc_txn->mt_loose_pgs = mp;
1622 mp->mp_flags |= P_LOOSE;
1624 int rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, pgno);
1632 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1633 * @param[in] mc A cursor handle for the current operation.
1634 * @param[in] pflags Flags of the pages to update:
1635 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1636 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1637 * @return 0 on success, non-zero on failure.
1640 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1642 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1643 MDB_txn *txn = mc->mc_txn;
1649 int rc = MDB_SUCCESS, level;
1651 /* Mark pages seen by cursors */
1652 if (mc->mc_flags & C_UNTRACK)
1653 mc = NULL; /* will find mc in mt_cursors */
1654 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1655 for (; mc; mc=mc->mc_next) {
1656 if (!(mc->mc_flags & C_INITIALIZED))
1658 for (m3 = mc;; m3 = &mx->mx_cursor) {
1660 for (j=0; j<m3->mc_snum; j++) {
1662 if ((mp->mp_flags & Mask) == pflags)
1663 mp->mp_flags ^= P_KEEP;
1665 mx = m3->mc_xcursor;
1666 /* Proceed to mx if it is at a sub-database */
1667 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1669 if (! (mp && (mp->mp_flags & P_LEAF)))
1671 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1672 if (!(leaf->mn_flags & F_SUBDATA))
1681 /* Mark dirty root pages */
1682 for (i=0; i<txn->mt_numdbs; i++) {
1683 if (txn->mt_dbflags[i] & DB_DIRTY) {
1684 pgno_t pgno = txn->mt_dbs[i].md_root;
1685 if (pgno == P_INVALID)
1687 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1689 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1690 dp->mp_flags ^= P_KEEP;
1698 static int mdb_page_flush(MDB_txn *txn, int keep);
1700 /** Spill pages from the dirty list back to disk.
1701 * This is intended to prevent running into #MDB_TXN_FULL situations,
1702 * but note that they may still occur in a few cases:
1703 * 1) our estimate of the txn size could be too small. Currently this
1704 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1705 * 2) child txns may run out of space if their parents dirtied a
1706 * lot of pages and never spilled them. TODO: we probably should do
1707 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1708 * the parent's dirty_room is below a given threshold.
1710 * Otherwise, if not using nested txns, it is expected that apps will
1711 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1712 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1713 * If the txn never references them again, they can be left alone.
1714 * If the txn only reads them, they can be used without any fuss.
1715 * If the txn writes them again, they can be dirtied immediately without
1716 * going thru all of the work of #mdb_page_touch(). Such references are
1717 * handled by #mdb_page_unspill().
1719 * Also note, we never spill DB root pages, nor pages of active cursors,
1720 * because we'll need these back again soon anyway. And in nested txns,
1721 * we can't spill a page in a child txn if it was already spilled in a
1722 * parent txn. That would alter the parent txns' data even though
1723 * the child hasn't committed yet, and we'd have no way to undo it if
1724 * the child aborted.
1726 * @param[in] m0 cursor A cursor handle identifying the transaction and
1727 * database for which we are checking space.
1728 * @param[in] key For a put operation, the key being stored.
1729 * @param[in] data For a put operation, the data being stored.
1730 * @return 0 on success, non-zero on failure.
1733 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1735 MDB_txn *txn = m0->mc_txn;
1737 MDB_ID2L dl = txn->mt_u.dirty_list;
1738 unsigned int i, j, need;
1741 if (m0->mc_flags & C_SUB)
1744 /* Estimate how much space this op will take */
1745 i = m0->mc_db->md_depth;
1746 /* Named DBs also dirty the main DB */
1747 if (m0->mc_dbi > MAIN_DBI)
1748 i += txn->mt_dbs[MAIN_DBI].md_depth;
1749 /* For puts, roughly factor in the key+data size */
1751 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1752 i += i; /* double it for good measure */
1755 if (txn->mt_dirty_room > i)
1758 if (!txn->mt_spill_pgs) {
1759 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1760 if (!txn->mt_spill_pgs)
1763 /* purge deleted slots */
1764 MDB_IDL sl = txn->mt_spill_pgs;
1765 unsigned int num = sl[0];
1767 for (i=1; i<=num; i++) {
1774 /* Preserve pages which may soon be dirtied again */
1775 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1778 /* Less aggressive spill - we originally spilled the entire dirty list,
1779 * with a few exceptions for cursor pages and DB root pages. But this
1780 * turns out to be a lot of wasted effort because in a large txn many
1781 * of those pages will need to be used again. So now we spill only 1/8th
1782 * of the dirty pages. Testing revealed this to be a good tradeoff,
1783 * better than 1/2, 1/4, or 1/10.
1785 if (need < MDB_IDL_UM_MAX / 8)
1786 need = MDB_IDL_UM_MAX / 8;
1788 /* Save the page IDs of all the pages we're flushing */
1789 /* flush from the tail forward, this saves a lot of shifting later on. */
1790 for (i=dl[0].mid; i && need; i--) {
1791 MDB_ID pn = dl[i].mid << 1;
1793 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1795 /* Can't spill twice, make sure it's not already in a parent's
1798 if (txn->mt_parent) {
1800 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1801 if (tx2->mt_spill_pgs) {
1802 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1803 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1804 dp->mp_flags |= P_KEEP;
1812 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1816 mdb_midl_sort(txn->mt_spill_pgs);
1818 /* Flush the spilled part of dirty list */
1819 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1822 /* Reset any dirty pages we kept that page_flush didn't see */
1823 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1826 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1830 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1832 mdb_find_oldest(MDB_txn *txn)
1835 txnid_t mr, oldest = txn->mt_txnid - 1;
1836 if (txn->mt_env->me_txns) {
1837 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1838 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1849 /** Add a page to the txn's dirty list */
1851 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1854 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1856 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1857 insert = mdb_mid2l_append;
1859 insert = mdb_mid2l_insert;
1861 mid.mid = mp->mp_pgno;
1863 rc = insert(txn->mt_u.dirty_list, &mid);
1864 mdb_tassert(txn, rc == 0);
1865 txn->mt_dirty_room--;
1868 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1869 * me_pghead and mt_next_pgno.
1871 * If there are free pages available from older transactions, they
1872 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1873 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1874 * and move me_pglast to say which records were consumed. Only this
1875 * function can create me_pghead and move me_pglast/mt_next_pgno.
1876 * @param[in] mc cursor A cursor handle identifying the transaction and
1877 * database for which we are allocating.
1878 * @param[in] num the number of pages to allocate.
1879 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1880 * will always be satisfied by a single contiguous chunk of memory.
1881 * @return 0 on success, non-zero on failure.
1884 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1886 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1887 /* Get at most <Max_retries> more freeDB records once me_pghead
1888 * has enough pages. If not enough, use new pages from the map.
1889 * If <Paranoid> and mc is updating the freeDB, only get new
1890 * records if me_pghead is empty. Then the freelist cannot play
1891 * catch-up with itself by growing while trying to save it.
1893 enum { Paranoid = 1, Max_retries = 500 };
1895 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1897 int rc, retry = num * 20;
1898 MDB_txn *txn = mc->mc_txn;
1899 MDB_env *env = txn->mt_env;
1900 pgno_t pgno, *mop = env->me_pghead;
1901 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1903 txnid_t oldest = 0, last;
1907 /* If there are any loose pages, just use them */
1908 if (num == 1 && txn->mt_loose_pgs) {
1909 np = txn->mt_loose_pgs;
1910 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1911 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
1919 /* If our dirty list is already full, we can't do anything */
1920 if (txn->mt_dirty_room == 0) {
1925 for (op = MDB_FIRST;; op = MDB_NEXT) {
1928 pgno_t *idl, old_id, new_id;
1930 /* Seek a big enough contiguous page range. Prefer
1931 * pages at the tail, just truncating the list.
1937 if (mop[i-n2] == pgno+n2)
1944 if (op == MDB_FIRST) { /* 1st iteration */
1945 /* Prepare to fetch more and coalesce */
1946 oldest = mdb_find_oldest(txn);
1947 last = env->me_pglast;
1948 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1951 key.mv_data = &last; /* will look up last+1 */
1952 key.mv_size = sizeof(last);
1954 if (Paranoid && mc->mc_dbi == FREE_DBI)
1957 if (Paranoid && retry < 0 && mop_len)
1961 /* Do not fetch more if the record will be too recent */
1964 rc = mdb_cursor_get(&m2, &key, NULL, op);
1966 if (rc == MDB_NOTFOUND)
1970 last = *(txnid_t*)key.mv_data;
1973 np = m2.mc_pg[m2.mc_top];
1974 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1975 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1978 idl = (MDB_ID *) data.mv_data;
1981 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
1986 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1988 mop = env->me_pghead;
1990 env->me_pglast = last;
1992 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1993 last, txn->mt_dbs[FREE_DBI].md_root, i));
1995 DPRINTF(("IDL %"Z"u", idl[k]));
1997 /* Merge in descending sorted order */
2000 mop[0] = (pgno_t)-1;
2004 for (; old_id < new_id; old_id = mop[--j])
2011 /* Use new pages from the map when nothing suitable in the freeDB */
2013 pgno = txn->mt_next_pgno;
2014 if (pgno + num >= env->me_maxpg) {
2015 DPUTS("DB size maxed out");
2021 if (env->me_flags & MDB_WRITEMAP) {
2022 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2024 if (!(np = mdb_page_malloc(txn, num))) {
2030 mop[0] = mop_len -= num;
2031 /* Move any stragglers down */
2032 for (j = i-num; j < mop_len; )
2033 mop[++j] = mop[++i];
2035 txn->mt_next_pgno = pgno + num;
2038 mdb_page_dirty(txn, np);
2044 txn->mt_flags |= MDB_TXN_ERROR;
2048 /** Copy the used portions of a non-overflow page.
2049 * @param[in] dst page to copy into
2050 * @param[in] src page to copy from
2051 * @param[in] psize size of a page
2054 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2056 enum { Align = sizeof(pgno_t) };
2057 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2059 /* If page isn't full, just copy the used portion. Adjust
2060 * alignment so memcpy may copy words instead of bytes.
2062 if ((unused &= -Align) && !IS_LEAF2(src)) {
2063 upper = (upper + PAGEBASE) & -Align;
2064 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2065 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2068 memcpy(dst, src, psize - unused);
2072 /** Pull a page off the txn's spill list, if present.
2073 * If a page being referenced was spilled to disk in this txn, bring
2074 * it back and make it dirty/writable again.
2075 * @param[in] txn the transaction handle.
2076 * @param[in] mp the page being referenced. It must not be dirty.
2077 * @param[out] ret the writable page, if any. ret is unchanged if
2078 * mp wasn't spilled.
2081 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2083 MDB_env *env = txn->mt_env;
2086 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2088 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2089 if (!tx2->mt_spill_pgs)
2091 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2092 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2095 if (txn->mt_dirty_room == 0)
2096 return MDB_TXN_FULL;
2097 if (IS_OVERFLOW(mp))
2101 if (env->me_flags & MDB_WRITEMAP) {
2104 np = mdb_page_malloc(txn, num);
2108 memcpy(np, mp, num * env->me_psize);
2110 mdb_page_copy(np, mp, env->me_psize);
2113 /* If in current txn, this page is no longer spilled.
2114 * If it happens to be the last page, truncate the spill list.
2115 * Otherwise mark it as deleted by setting the LSB.
2117 if (x == txn->mt_spill_pgs[0])
2118 txn->mt_spill_pgs[0]--;
2120 txn->mt_spill_pgs[x] |= 1;
2121 } /* otherwise, if belonging to a parent txn, the
2122 * page remains spilled until child commits
2125 mdb_page_dirty(txn, np);
2126 np->mp_flags |= P_DIRTY;
2134 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2135 * @param[in] mc cursor pointing to the page to be touched
2136 * @return 0 on success, non-zero on failure.
2139 mdb_page_touch(MDB_cursor *mc)
2141 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2142 MDB_txn *txn = mc->mc_txn;
2143 MDB_cursor *m2, *m3;
2147 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2148 if (txn->mt_flags & MDB_TXN_SPILLS) {
2150 rc = mdb_page_unspill(txn, mp, &np);
2156 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2157 (rc = mdb_page_alloc(mc, 1, &np)))
2160 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2161 mp->mp_pgno, pgno));
2162 mdb_cassert(mc, mp->mp_pgno != pgno);
2163 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2164 /* Update the parent page, if any, to point to the new page */
2166 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2167 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2168 SETPGNO(node, pgno);
2170 mc->mc_db->md_root = pgno;
2172 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2173 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2175 /* If txn has a parent, make sure the page is in our
2179 unsigned x = mdb_mid2l_search(dl, pgno);
2180 if (x <= dl[0].mid && dl[x].mid == pgno) {
2181 if (mp != dl[x].mptr) { /* bad cursor? */
2182 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2183 txn->mt_flags |= MDB_TXN_ERROR;
2184 return MDB_CORRUPTED;
2189 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2191 np = mdb_page_malloc(txn, 1);
2196 rc = mdb_mid2l_insert(dl, &mid);
2197 mdb_cassert(mc, rc == 0);
2202 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2204 np->mp_flags |= P_DIRTY;
2207 /* Adjust cursors pointing to mp */
2208 mc->mc_pg[mc->mc_top] = np;
2209 m2 = txn->mt_cursors[mc->mc_dbi];
2210 if (mc->mc_flags & C_SUB) {
2211 for (; m2; m2=m2->mc_next) {
2212 m3 = &m2->mc_xcursor->mx_cursor;
2213 if (m3->mc_snum < mc->mc_snum) continue;
2214 if (m3->mc_pg[mc->mc_top] == mp)
2215 m3->mc_pg[mc->mc_top] = np;
2218 for (; m2; m2=m2->mc_next) {
2219 if (m2->mc_snum < mc->mc_snum) continue;
2220 if (m2->mc_pg[mc->mc_top] == mp) {
2221 m2->mc_pg[mc->mc_top] = np;
2222 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2224 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2226 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2227 if (!(leaf->mn_flags & F_SUBDATA))
2228 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2236 txn->mt_flags |= MDB_TXN_ERROR;
2241 mdb_env_sync(MDB_env *env, int force)
2244 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2245 if (env->me_flags & MDB_WRITEMAP) {
2246 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2247 ? MS_ASYNC : MS_SYNC;
2248 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2251 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2255 if (MDB_FDATASYNC(env->me_fd))
2262 /** Back up parent txn's cursors, then grab the originals for tracking */
2264 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2266 MDB_cursor *mc, *bk;
2271 for (i = src->mt_numdbs; --i >= 0; ) {
2272 if ((mc = src->mt_cursors[i]) != NULL) {
2273 size = sizeof(MDB_cursor);
2275 size += sizeof(MDB_xcursor);
2276 for (; mc; mc = bk->mc_next) {
2282 mc->mc_db = &dst->mt_dbs[i];
2283 /* Kill pointers into src - and dst to reduce abuse: The
2284 * user may not use mc until dst ends. Otherwise we'd...
2286 mc->mc_txn = NULL; /* ...set this to dst */
2287 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2288 if ((mx = mc->mc_xcursor) != NULL) {
2289 *(MDB_xcursor *)(bk+1) = *mx;
2290 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2292 mc->mc_next = dst->mt_cursors[i];
2293 dst->mt_cursors[i] = mc;
2300 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2301 * @param[in] txn the transaction handle.
2302 * @param[in] merge true to keep changes to parent cursors, false to revert.
2303 * @return 0 on success, non-zero on failure.
2306 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2308 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2312 for (i = txn->mt_numdbs; --i >= 0; ) {
2313 for (mc = cursors[i]; mc; mc = next) {
2315 if ((bk = mc->mc_backup) != NULL) {
2317 /* Commit changes to parent txn */
2318 mc->mc_next = bk->mc_next;
2319 mc->mc_backup = bk->mc_backup;
2320 mc->mc_txn = bk->mc_txn;
2321 mc->mc_db = bk->mc_db;
2322 mc->mc_dbflag = bk->mc_dbflag;
2323 if ((mx = mc->mc_xcursor) != NULL)
2324 mx->mx_cursor.mc_txn = bk->mc_txn;
2326 /* Abort nested txn */
2328 if ((mx = mc->mc_xcursor) != NULL)
2329 *mx = *(MDB_xcursor *)(bk+1);
2333 /* Only malloced cursors are permanently tracked. */
2341 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2344 mdb_txn_reset0(MDB_txn *txn, const char *act);
2346 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2352 Pidset = F_SETLK, Pidcheck = F_GETLK
2356 /** Set or check a pid lock. Set returns 0 on success.
2357 * Check returns 0 if the process is certainly dead, nonzero if it may
2358 * be alive (the lock exists or an error happened so we do not know).
2360 * On Windows Pidset is a no-op, we merely check for the existence
2361 * of the process with the given pid. On POSIX we use a single byte
2362 * lock on the lockfile, set at an offset equal to the pid.
2365 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2367 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2370 if (op == Pidcheck) {
2371 h = OpenProcess(env->me_pidquery, FALSE, pid);
2372 /* No documented "no such process" code, but other program use this: */
2374 return ErrCode() != ERROR_INVALID_PARAMETER;
2375 /* A process exists until all handles to it close. Has it exited? */
2376 ret = WaitForSingleObject(h, 0) != 0;
2383 struct flock lock_info;
2384 memset(&lock_info, 0, sizeof(lock_info));
2385 lock_info.l_type = F_WRLCK;
2386 lock_info.l_whence = SEEK_SET;
2387 lock_info.l_start = pid;
2388 lock_info.l_len = 1;
2389 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2390 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2392 } else if ((rc = ErrCode()) == EINTR) {
2400 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2401 * @param[in] txn the transaction handle to initialize
2402 * @return 0 on success, non-zero on failure.
2405 mdb_txn_renew0(MDB_txn *txn)
2407 MDB_env *env = txn->mt_env;
2408 MDB_txninfo *ti = env->me_txns;
2412 int rc, new_notls = 0;
2415 txn->mt_numdbs = env->me_numdbs;
2416 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2418 if (txn->mt_flags & MDB_TXN_RDONLY) {
2420 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2421 txn->mt_txnid = meta->mm_txnid;
2422 txn->mt_u.reader = NULL;
2424 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2425 pthread_getspecific(env->me_txkey);
2427 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2428 return MDB_BAD_RSLOT;
2430 MDB_PID_T pid = env->me_pid;
2431 MDB_THR_T tid = pthread_self();
2433 if (!env->me_live_reader) {
2434 rc = mdb_reader_pid(env, Pidset, pid);
2437 env->me_live_reader = 1;
2441 nr = ti->mti_numreaders;
2442 for (i=0; i<nr; i++)
2443 if (ti->mti_readers[i].mr_pid == 0)
2445 if (i == env->me_maxreaders) {
2446 UNLOCK_MUTEX_R(env);
2447 return MDB_READERS_FULL;
2449 ti->mti_readers[i].mr_pid = pid;
2450 ti->mti_readers[i].mr_tid = tid;
2452 ti->mti_numreaders = ++nr;
2453 /* Save numreaders for un-mutexed mdb_env_close() */
2454 env->me_numreaders = nr;
2455 UNLOCK_MUTEX_R(env);
2457 r = &ti->mti_readers[i];
2458 new_notls = (env->me_flags & MDB_NOTLS);
2459 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2464 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2465 txn->mt_u.reader = r;
2466 meta = env->me_metas[txn->mt_txnid & 1];
2472 txn->mt_txnid = ti->mti_txnid;
2473 meta = env->me_metas[txn->mt_txnid & 1];
2475 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2476 txn->mt_txnid = meta->mm_txnid;
2480 if (txn->mt_txnid == mdb_debug_start)
2483 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2484 txn->mt_u.dirty_list = env->me_dirty_list;
2485 txn->mt_u.dirty_list[0].mid = 0;
2486 txn->mt_free_pgs = env->me_free_pgs;
2487 txn->mt_free_pgs[0] = 0;
2488 txn->mt_spill_pgs = NULL;
2490 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2493 /* Copy the DB info and flags */
2494 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2496 /* Moved to here to avoid a data race in read TXNs */
2497 txn->mt_next_pgno = meta->mm_last_pg+1;
2499 for (i=2; i<txn->mt_numdbs; i++) {
2500 x = env->me_dbflags[i];
2501 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2502 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2504 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2506 /* If we didn't ask for a resize, but the size grew, fail */
2507 if (!(env->me_flags & MDB_RESIZING)
2508 && env->me_mapsize < meta->mm_mapsize) {
2509 mdb_txn_reset0(txn, "renew0-mapfail");
2511 txn->mt_u.reader->mr_pid = 0;
2512 txn->mt_u.reader = NULL;
2514 return MDB_MAP_RESIZED;
2521 mdb_txn_renew(MDB_txn *txn)
2525 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2528 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2529 DPUTS("environment had fatal error, must shutdown!");
2533 rc = mdb_txn_renew0(txn);
2534 if (rc == MDB_SUCCESS) {
2535 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2536 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2537 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2543 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2547 int rc, size, tsize = sizeof(MDB_txn);
2549 if (env->me_flags & MDB_FATAL_ERROR) {
2550 DPUTS("environment had fatal error, must shutdown!");
2553 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2556 /* Nested transactions: Max 1 child, write txns only, no writemap */
2557 if (parent->mt_child ||
2558 (flags & MDB_RDONLY) ||
2559 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2560 (env->me_flags & MDB_WRITEMAP))
2562 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2564 tsize = sizeof(MDB_ntxn);
2566 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2567 if (!(flags & MDB_RDONLY)) {
2568 size += env->me_maxdbs * sizeof(MDB_cursor *);
2569 /* child txns use parent's dbiseqs */
2571 size += env->me_maxdbs * sizeof(unsigned int);
2574 if ((txn = calloc(1, size)) == NULL) {
2575 DPRINTF(("calloc: %s", strerror(ErrCode())));
2578 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2579 if (flags & MDB_RDONLY) {
2580 txn->mt_flags |= MDB_TXN_RDONLY;
2581 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2582 txn->mt_dbiseqs = env->me_dbiseqs;
2584 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2586 txn->mt_dbiseqs = parent->mt_dbiseqs;
2587 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2589 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2590 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2597 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2598 if (!txn->mt_u.dirty_list ||
2599 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2601 free(txn->mt_u.dirty_list);
2605 txn->mt_txnid = parent->mt_txnid;
2606 txn->mt_dirty_room = parent->mt_dirty_room;
2607 txn->mt_u.dirty_list[0].mid = 0;
2608 txn->mt_spill_pgs = NULL;
2609 txn->mt_next_pgno = parent->mt_next_pgno;
2610 parent->mt_child = txn;
2611 txn->mt_parent = parent;
2612 txn->mt_numdbs = parent->mt_numdbs;
2613 txn->mt_flags = parent->mt_flags;
2614 txn->mt_dbxs = parent->mt_dbxs;
2615 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2616 /* Copy parent's mt_dbflags, but clear DB_NEW */
2617 for (i=0; i<txn->mt_numdbs; i++)
2618 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2620 ntxn = (MDB_ntxn *)txn;
2621 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2622 if (env->me_pghead) {
2623 size = MDB_IDL_SIZEOF(env->me_pghead);
2624 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2626 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2631 rc = mdb_cursor_shadow(parent, txn);
2633 mdb_txn_reset0(txn, "beginchild-fail");
2635 rc = mdb_txn_renew0(txn);
2641 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2642 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2643 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2650 mdb_txn_env(MDB_txn *txn)
2652 if(!txn) return NULL;
2656 /** Export or close DBI handles opened in this txn. */
2658 mdb_dbis_update(MDB_txn *txn, int keep)
2661 MDB_dbi n = txn->mt_numdbs;
2662 MDB_env *env = txn->mt_env;
2663 unsigned char *tdbflags = txn->mt_dbflags;
2665 for (i = n; --i >= 2;) {
2666 if (tdbflags[i] & DB_NEW) {
2668 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2670 char *ptr = env->me_dbxs[i].md_name.mv_data;
2672 env->me_dbxs[i].md_name.mv_data = NULL;
2673 env->me_dbxs[i].md_name.mv_size = 0;
2674 env->me_dbflags[i] = 0;
2675 env->me_dbiseqs[i]++;
2681 if (keep && env->me_numdbs < n)
2685 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2686 * May be called twice for readonly txns: First reset it, then abort.
2687 * @param[in] txn the transaction handle to reset
2688 * @param[in] act why the transaction is being reset
2691 mdb_txn_reset0(MDB_txn *txn, const char *act)
2693 MDB_env *env = txn->mt_env;
2695 /* Close any DBI handles opened in this txn */
2696 mdb_dbis_update(txn, 0);
2698 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2699 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2700 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2702 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2703 if (txn->mt_u.reader) {
2704 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2705 if (!(env->me_flags & MDB_NOTLS))
2706 txn->mt_u.reader = NULL; /* txn does not own reader */
2708 txn->mt_numdbs = 0; /* close nothing if called again */
2709 txn->mt_dbxs = NULL; /* mark txn as reset */
2711 mdb_cursors_close(txn, 0);
2713 if (!(env->me_flags & MDB_WRITEMAP)) {
2714 mdb_dlist_free(txn);
2716 mdb_midl_free(env->me_pghead);
2718 if (txn->mt_parent) {
2719 txn->mt_parent->mt_child = NULL;
2720 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2721 mdb_midl_free(txn->mt_free_pgs);
2722 mdb_midl_free(txn->mt_spill_pgs);
2723 free(txn->mt_u.dirty_list);
2727 if (mdb_midl_shrink(&txn->mt_free_pgs))
2728 env->me_free_pgs = txn->mt_free_pgs;
2729 env->me_pghead = NULL;
2733 /* The writer mutex was locked in mdb_txn_begin. */
2735 UNLOCK_MUTEX_W(env);
2740 mdb_txn_reset(MDB_txn *txn)
2745 /* This call is only valid for read-only txns */
2746 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2749 mdb_txn_reset0(txn, "reset");
2753 mdb_txn_abort(MDB_txn *txn)
2759 mdb_txn_abort(txn->mt_child);
2761 mdb_txn_reset0(txn, "abort");
2762 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2763 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2764 txn->mt_u.reader->mr_pid = 0;
2769 /** Save the freelist as of this transaction to the freeDB.
2770 * This changes the freelist. Keep trying until it stabilizes.
2773 mdb_freelist_save(MDB_txn *txn)
2775 /* env->me_pghead[] can grow and shrink during this call.
2776 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2777 * Page numbers cannot disappear from txn->mt_free_pgs[].
2780 MDB_env *env = txn->mt_env;
2781 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2782 txnid_t pglast = 0, head_id = 0;
2783 pgno_t freecnt = 0, *free_pgs, *mop;
2784 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2786 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2788 if (env->me_pghead) {
2789 /* Make sure first page of freeDB is touched and on freelist */
2790 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2791 if (rc && rc != MDB_NOTFOUND)
2795 /* Dispose of loose pages. Usually they will have all
2796 * been used up by the time we get here.
2798 if (txn->mt_loose_pgs) {
2799 MDB_page *mp = txn->mt_loose_pgs;
2800 /* Just return them to freeDB */
2801 if (env->me_pghead) {
2803 mop = env->me_pghead;
2804 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2805 pgno_t pg = mp->mp_pgno;
2807 for (i = mop[0]; i && mop[i] < pg; i--)
2813 /* Oh well, they were wasted. Put on freelist */
2814 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2815 mdb_midl_append(&txn->mt_free_pgs, mp->mp_pgno);
2818 txn->mt_loose_pgs = NULL;
2821 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2822 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2823 ? SSIZE_MAX : maxfree_1pg;
2826 /* Come back here after each Put() in case freelist changed */
2831 /* If using records from freeDB which we have not yet
2832 * deleted, delete them and any we reserved for me_pghead.
2834 while (pglast < env->me_pglast) {
2835 rc = mdb_cursor_first(&mc, &key, NULL);
2838 pglast = head_id = *(txnid_t *)key.mv_data;
2839 total_room = head_room = 0;
2840 mdb_tassert(txn, pglast <= env->me_pglast);
2841 rc = mdb_cursor_del(&mc, 0);
2846 /* Save the IDL of pages freed by this txn, to a single record */
2847 if (freecnt < txn->mt_free_pgs[0]) {
2849 /* Make sure last page of freeDB is touched and on freelist */
2850 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2851 if (rc && rc != MDB_NOTFOUND)
2854 free_pgs = txn->mt_free_pgs;
2855 /* Write to last page of freeDB */
2856 key.mv_size = sizeof(txn->mt_txnid);
2857 key.mv_data = &txn->mt_txnid;
2859 freecnt = free_pgs[0];
2860 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2861 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2864 /* Retry if mt_free_pgs[] grew during the Put() */
2865 free_pgs = txn->mt_free_pgs;
2866 } while (freecnt < free_pgs[0]);
2867 mdb_midl_sort(free_pgs);
2868 memcpy(data.mv_data, free_pgs, data.mv_size);
2871 unsigned int i = free_pgs[0];
2872 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2873 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2875 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2881 mop = env->me_pghead;
2882 mop_len = mop ? mop[0] : 0;
2884 /* Reserve records for me_pghead[]. Split it if multi-page,
2885 * to avoid searching freeDB for a page range. Use keys in
2886 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2888 if (total_room >= mop_len) {
2889 if (total_room == mop_len || --more < 0)
2891 } else if (head_room >= maxfree_1pg && head_id > 1) {
2892 /* Keep current record (overflow page), add a new one */
2896 /* (Re)write {key = head_id, IDL length = head_room} */
2897 total_room -= head_room;
2898 head_room = mop_len - total_room;
2899 if (head_room > maxfree_1pg && head_id > 1) {
2900 /* Overflow multi-page for part of me_pghead */
2901 head_room /= head_id; /* amortize page sizes */
2902 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2903 } else if (head_room < 0) {
2904 /* Rare case, not bothering to delete this record */
2907 key.mv_size = sizeof(head_id);
2908 key.mv_data = &head_id;
2909 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2910 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2913 /* IDL is initially empty, zero out at least the length */
2914 pgs = (pgno_t *)data.mv_data;
2915 j = head_room > clean_limit ? head_room : 0;
2919 total_room += head_room;
2922 /* Fill in the reserved me_pghead records */
2928 rc = mdb_cursor_first(&mc, &key, &data);
2929 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2930 txnid_t id = *(txnid_t *)key.mv_data;
2931 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2934 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2936 if (len > mop_len) {
2938 data.mv_size = (len + 1) * sizeof(MDB_ID);
2940 data.mv_data = mop -= len;
2943 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2945 if (rc || !(mop_len -= len))
2952 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2953 * @param[in] txn the transaction that's being committed
2954 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2955 * @return 0 on success, non-zero on failure.
2958 mdb_page_flush(MDB_txn *txn, int keep)
2960 MDB_env *env = txn->mt_env;
2961 MDB_ID2L dl = txn->mt_u.dirty_list;
2962 unsigned psize = env->me_psize, j;
2963 int i, pagecount = dl[0].mid, rc;
2964 size_t size = 0, pos = 0;
2966 MDB_page *dp = NULL;
2970 struct iovec iov[MDB_COMMIT_PAGES];
2971 ssize_t wpos = 0, wsize = 0, wres;
2972 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2978 if (env->me_flags & MDB_WRITEMAP) {
2979 /* Clear dirty flags */
2980 while (++i <= pagecount) {
2982 /* Don't flush this page yet */
2983 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
2984 dp->mp_flags &= ~P_KEEP;
2988 dp->mp_flags &= ~P_DIRTY;
2993 /* Write the pages */
2995 if (++i <= pagecount) {
2997 /* Don't flush this page yet */
2998 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
2999 dp->mp_flags &= ~P_KEEP;
3004 /* clear dirty flag */
3005 dp->mp_flags &= ~P_DIRTY;
3008 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3013 /* Windows actually supports scatter/gather I/O, but only on
3014 * unbuffered file handles. Since we're relying on the OS page
3015 * cache for all our data, that's self-defeating. So we just
3016 * write pages one at a time. We use the ov structure to set
3017 * the write offset, to at least save the overhead of a Seek
3020 DPRINTF(("committing page %"Z"u", pgno));
3021 memset(&ov, 0, sizeof(ov));
3022 ov.Offset = pos & 0xffffffff;
3023 ov.OffsetHigh = pos >> 16 >> 16;
3024 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3026 DPRINTF(("WriteFile: %d", rc));
3030 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3031 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3033 /* Write previous page(s) */
3034 #ifdef MDB_USE_PWRITEV
3035 wres = pwritev(env->me_fd, iov, n, wpos);
3038 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3040 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3042 DPRINTF(("lseek: %s", strerror(rc)));
3045 wres = writev(env->me_fd, iov, n);
3048 if (wres != wsize) {
3051 DPRINTF(("Write error: %s", strerror(rc)));
3053 rc = EIO; /* TODO: Use which error code? */
3054 DPUTS("short write, filesystem full?");
3065 DPRINTF(("committing page %"Z"u", pgno));
3066 next_pos = pos + size;
3067 iov[n].iov_len = size;
3068 iov[n].iov_base = (char *)dp;
3074 for (i = keep; ++i <= pagecount; ) {
3076 /* This is a page we skipped above */
3079 dl[j].mid = dp->mp_pgno;
3082 mdb_dpage_free(env, dp);
3087 txn->mt_dirty_room += i - j;
3093 mdb_txn_commit(MDB_txn *txn)
3099 if (txn == NULL || txn->mt_env == NULL)
3102 if (txn->mt_child) {
3103 rc = mdb_txn_commit(txn->mt_child);
3104 txn->mt_child = NULL;
3111 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3112 mdb_dbis_update(txn, 1);
3113 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3118 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3119 DPUTS("error flag is set, can't commit");
3121 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3126 if (txn->mt_parent) {
3127 MDB_txn *parent = txn->mt_parent;
3131 unsigned x, y, len, ps_len;
3133 /* Append our free list to parent's */
3134 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3137 mdb_midl_free(txn->mt_free_pgs);
3138 /* Failures after this must either undo the changes
3139 * to the parent or set MDB_TXN_ERROR in the parent.
3142 parent->mt_next_pgno = txn->mt_next_pgno;
3143 parent->mt_flags = txn->mt_flags;
3145 /* Merge our cursors into parent's and close them */
3146 mdb_cursors_close(txn, 1);
3148 /* Update parent's DB table. */
3149 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3150 parent->mt_numdbs = txn->mt_numdbs;
3151 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3152 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3153 for (i=2; i<txn->mt_numdbs; i++) {
3154 /* preserve parent's DB_NEW status */
3155 x = parent->mt_dbflags[i] & DB_NEW;
3156 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3159 dst = parent->mt_u.dirty_list;
3160 src = txn->mt_u.dirty_list;
3161 /* Remove anything in our dirty list from parent's spill list */
3162 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3164 pspill[0] = (pgno_t)-1;
3165 /* Mark our dirty pages as deleted in parent spill list */
3166 for (i=0, len=src[0].mid; ++i <= len; ) {
3167 MDB_ID pn = src[i].mid << 1;
3168 while (pn > pspill[x])
3170 if (pn == pspill[x]) {
3175 /* Squash deleted pagenums if we deleted any */
3176 for (x=y; ++x <= ps_len; )
3177 if (!(pspill[x] & 1))
3178 pspill[++y] = pspill[x];
3182 /* Find len = length of merging our dirty list with parent's */
3184 dst[0].mid = 0; /* simplify loops */
3185 if (parent->mt_parent) {
3186 len = x + src[0].mid;
3187 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3188 for (i = x; y && i; y--) {
3189 pgno_t yp = src[y].mid;
3190 while (yp < dst[i].mid)
3192 if (yp == dst[i].mid) {
3197 } else { /* Simplify the above for single-ancestor case */
3198 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3200 /* Merge our dirty list with parent's */
3202 for (i = len; y; dst[i--] = src[y--]) {
3203 pgno_t yp = src[y].mid;
3204 while (yp < dst[x].mid)
3205 dst[i--] = dst[x--];
3206 if (yp == dst[x].mid)
3207 free(dst[x--].mptr);
3209 mdb_tassert(txn, i == x);
3211 free(txn->mt_u.dirty_list);
3212 parent->mt_dirty_room = txn->mt_dirty_room;
3213 if (txn->mt_spill_pgs) {
3214 if (parent->mt_spill_pgs) {
3215 /* TODO: Prevent failure here, so parent does not fail */
3216 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3218 parent->mt_flags |= MDB_TXN_ERROR;
3219 mdb_midl_free(txn->mt_spill_pgs);
3220 mdb_midl_sort(parent->mt_spill_pgs);
3222 parent->mt_spill_pgs = txn->mt_spill_pgs;
3226 /* Append our loose page list to parent's */
3227 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3229 *lp = txn->mt_loose_pgs;
3231 parent->mt_child = NULL;
3232 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3237 if (txn != env->me_txn) {
3238 DPUTS("attempt to commit unknown transaction");
3243 mdb_cursors_close(txn, 0);
3245 if (!txn->mt_u.dirty_list[0].mid &&
3246 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS))) {
3247 if ((env->me_flags & MDB_RESIZING)
3248 && (rc = mdb_env_write_meta(txn))) {
3254 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3255 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3257 /* Update DB root pointers */
3258 if (txn->mt_numdbs > 2) {
3262 data.mv_size = sizeof(MDB_db);
3264 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3265 for (i = 2; i < txn->mt_numdbs; i++) {
3266 if (txn->mt_dbflags[i] & DB_DIRTY) {
3267 if (TXN_DBI_CHANGED(txn, i)) {
3271 data.mv_data = &txn->mt_dbs[i];
3272 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3279 rc = mdb_freelist_save(txn);
3283 mdb_midl_free(env->me_pghead);
3284 env->me_pghead = NULL;
3285 if (mdb_midl_shrink(&txn->mt_free_pgs))
3286 env->me_free_pgs = txn->mt_free_pgs;
3292 if ((rc = mdb_page_flush(txn, 0)) ||
3293 (rc = mdb_env_sync(env, 0)) ||
3294 (rc = mdb_env_write_meta(txn)))
3300 mdb_dbis_update(txn, 1);
3303 UNLOCK_MUTEX_W(env);
3313 /** Read the environment parameters of a DB environment before
3314 * mapping it into memory.
3315 * @param[in] env the environment handle
3316 * @param[out] meta address of where to store the meta information
3317 * @return 0 on success, non-zero on failure.
3320 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3326 enum { Size = sizeof(pbuf) };
3328 /* We don't know the page size yet, so use a minimum value.
3329 * Read both meta pages so we can use the latest one.
3332 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3336 memset(&ov, 0, sizeof(ov));
3338 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3339 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3342 rc = pread(env->me_fd, &pbuf, Size, off);
3345 if (rc == 0 && off == 0)
3347 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3348 DPRINTF(("read: %s", mdb_strerror(rc)));
3352 p = (MDB_page *)&pbuf;
3354 if (!F_ISSET(p->mp_flags, P_META)) {
3355 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3360 if (m->mm_magic != MDB_MAGIC) {
3361 DPUTS("meta has invalid magic");
3365 if (m->mm_version != MDB_DATA_VERSION) {
3366 DPRINTF(("database is version %u, expected version %u",
3367 m->mm_version, MDB_DATA_VERSION));
3368 return MDB_VERSION_MISMATCH;
3371 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3378 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3380 meta->mm_magic = MDB_MAGIC;
3381 meta->mm_version = MDB_DATA_VERSION;
3382 meta->mm_mapsize = env->me_mapsize;
3383 meta->mm_psize = env->me_psize;
3384 meta->mm_last_pg = 1;
3385 meta->mm_flags = env->me_flags & 0xffff;
3386 meta->mm_flags |= MDB_INTEGERKEY;
3387 meta->mm_dbs[0].md_root = P_INVALID;
3388 meta->mm_dbs[1].md_root = P_INVALID;
3391 /** Write the environment parameters of a freshly created DB environment.
3392 * @param[in] env the environment handle
3393 * @param[out] meta address of where to store the meta information
3394 * @return 0 on success, non-zero on failure.
3397 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3405 memset(&ov, 0, sizeof(ov));
3406 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3408 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3411 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3412 len = pwrite(fd, ptr, size, pos); \
3413 rc = (len >= 0); } while(0)
3416 DPUTS("writing new meta page");
3418 psize = env->me_psize;
3420 mdb_env_init_meta0(env, meta);
3422 p = calloc(2, psize);
3424 p->mp_flags = P_META;
3425 *(MDB_meta *)METADATA(p) = *meta;
3427 q = (MDB_page *)((char *)p + psize);
3429 q->mp_flags = P_META;
3430 *(MDB_meta *)METADATA(q) = *meta;
3432 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3435 else if ((unsigned) len == psize * 2)
3443 /** Update the environment info to commit a transaction.
3444 * @param[in] txn the transaction that's being committed
3445 * @return 0 on success, non-zero on failure.
3448 mdb_env_write_meta(MDB_txn *txn)
3451 MDB_meta meta, metab, *mp;
3453 int rc, len, toggle;
3462 toggle = txn->mt_txnid & 1;
3463 DPRINTF(("writing meta page %d for root page %"Z"u",
3464 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3467 mp = env->me_metas[toggle];
3469 if (env->me_flags & MDB_WRITEMAP) {
3470 /* Persist any changes of mapsize config */
3471 if (env->me_flags & MDB_RESIZING) {
3472 mp->mm_mapsize = env->me_mapsize;
3473 env->me_flags ^= MDB_RESIZING;
3475 mp->mm_dbs[0] = txn->mt_dbs[0];
3476 mp->mm_dbs[1] = txn->mt_dbs[1];
3477 mp->mm_last_pg = txn->mt_next_pgno - 1;
3478 mp->mm_txnid = txn->mt_txnid;
3479 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3480 unsigned meta_size = env->me_psize;
3481 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3484 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3485 if (meta_size < env->me_os_psize)
3486 meta_size += meta_size;
3491 if (MDB_MSYNC(ptr, meta_size, rc)) {
3498 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3499 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3501 ptr = (char *)&meta;
3502 if (env->me_flags & MDB_RESIZING) {
3503 /* Persist any changes of mapsize config */
3504 meta.mm_mapsize = env->me_mapsize;
3505 off = offsetof(MDB_meta, mm_mapsize);
3506 env->me_flags ^= MDB_RESIZING;
3508 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3510 len = sizeof(MDB_meta) - off;
3513 meta.mm_dbs[0] = txn->mt_dbs[0];
3514 meta.mm_dbs[1] = txn->mt_dbs[1];
3515 meta.mm_last_pg = txn->mt_next_pgno - 1;
3516 meta.mm_txnid = txn->mt_txnid;
3519 off += env->me_psize;
3522 /* Write to the SYNC fd */
3523 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3524 env->me_fd : env->me_mfd;
3527 memset(&ov, 0, sizeof(ov));
3529 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3533 rc = pwrite(mfd, ptr, len, off);
3536 rc = rc < 0 ? ErrCode() : EIO;
3537 DPUTS("write failed, disk error?");
3538 /* On a failure, the pagecache still contains the new data.
3539 * Write some old data back, to prevent it from being used.
3540 * Use the non-SYNC fd; we know it will fail anyway.
3542 meta.mm_last_pg = metab.mm_last_pg;
3543 meta.mm_txnid = metab.mm_txnid;
3545 memset(&ov, 0, sizeof(ov));
3547 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3549 r2 = pwrite(env->me_fd, ptr, len, off);
3550 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3553 env->me_flags |= MDB_FATAL_ERROR;
3557 /* Memory ordering issues are irrelevant; since the entire writer
3558 * is wrapped by wmutex, all of these changes will become visible
3559 * after the wmutex is unlocked. Since the DB is multi-version,
3560 * readers will get consistent data regardless of how fresh or
3561 * how stale their view of these values is.
3564 env->me_txns->mti_txnid = txn->mt_txnid;
3569 /** Check both meta pages to see which one is newer.
3570 * @param[in] env the environment handle
3571 * @return meta toggle (0 or 1).
3574 mdb_env_pick_meta(const MDB_env *env)
3576 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3580 mdb_env_create(MDB_env **env)
3584 e = calloc(1, sizeof(MDB_env));
3588 e->me_maxreaders = DEFAULT_READERS;
3589 e->me_maxdbs = e->me_numdbs = 2;
3590 e->me_fd = INVALID_HANDLE_VALUE;
3591 e->me_lfd = INVALID_HANDLE_VALUE;
3592 e->me_mfd = INVALID_HANDLE_VALUE;
3593 #ifdef MDB_USE_POSIX_SEM
3594 e->me_rmutex = SEM_FAILED;
3595 e->me_wmutex = SEM_FAILED;
3597 e->me_pid = getpid();
3598 GET_PAGESIZE(e->me_os_psize);
3599 VGMEMP_CREATE(e,0,0);
3605 mdb_env_map(MDB_env *env, void *addr, int newsize)
3608 unsigned int flags = env->me_flags;
3612 LONG sizelo, sizehi;
3615 if (flags & MDB_RDONLY) {
3620 msize = env->me_mapsize;
3621 sizelo = msize & 0xffffffff;
3622 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3625 /* Windows won't create mappings for zero length files.
3626 * Just allocate the maxsize right now.
3629 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3630 || !SetEndOfFile(env->me_fd)
3631 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3634 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3635 PAGE_READWRITE : PAGE_READONLY,
3636 sizehi, sizelo, NULL);
3639 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3640 FILE_MAP_WRITE : FILE_MAP_READ,
3642 rc = env->me_map ? 0 : ErrCode();
3647 int prot = PROT_READ;
3648 if (flags & MDB_WRITEMAP) {
3650 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3653 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3655 if (env->me_map == MAP_FAILED) {
3660 if (flags & MDB_NORDAHEAD) {
3661 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3663 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3665 #ifdef POSIX_MADV_RANDOM
3666 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3667 #endif /* POSIX_MADV_RANDOM */
3668 #endif /* MADV_RANDOM */
3672 /* Can happen because the address argument to mmap() is just a
3673 * hint. mmap() can pick another, e.g. if the range is in use.
3674 * The MAP_FIXED flag would prevent that, but then mmap could
3675 * instead unmap existing pages to make room for the new map.
3677 if (addr && env->me_map != addr)
3678 return EBUSY; /* TODO: Make a new MDB_* error code? */
3680 p = (MDB_page *)env->me_map;
3681 env->me_metas[0] = METADATA(p);
3682 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3688 mdb_env_set_mapsize(MDB_env *env, size_t size)
3690 /* If env is already open, caller is responsible for making
3691 * sure there are no active txns.
3699 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3701 if (size < env->me_mapsize) {
3702 /* If the configured size is smaller, make sure it's
3703 * still big enough. Silently round up to minimum if not.
3705 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3709 /* nothing actually changed */
3710 if (size == env->me_mapsize)
3714 munmap(env->me_map, env->me_mapsize);
3715 env->me_mapsize = size;
3716 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3717 rc = mdb_env_map(env, old, 1);
3721 env->me_flags |= MDB_RESIZING;
3723 env->me_mapsize = size;
3725 env->me_maxpg = env->me_mapsize / env->me_psize;
3730 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3734 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3739 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3741 if (env->me_map || readers < 1)
3743 env->me_maxreaders = readers;
3748 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3750 if (!env || !readers)
3752 *readers = env->me_maxreaders;
3756 /** Further setup required for opening an LMDB environment
3759 mdb_env_open2(MDB_env *env)
3761 unsigned int flags = env->me_flags;
3762 int i, newenv = 0, rc;
3766 /* See if we should use QueryLimited */
3768 if ((rc & 0xff) > 5)
3769 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3771 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3774 memset(&meta, 0, sizeof(meta));
3776 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3779 DPUTS("new mdbenv");
3781 env->me_psize = env->me_os_psize;
3782 if (env->me_psize > MAX_PAGESIZE)
3783 env->me_psize = MAX_PAGESIZE;
3785 env->me_psize = meta.mm_psize;
3788 /* Was a mapsize configured? */
3789 if (!env->me_mapsize) {
3790 /* If this is a new environment, take the default,
3791 * else use the size recorded in the existing env.
3793 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3795 if (env->me_mapsize < meta.mm_mapsize) {
3796 /* If the configured size is smaller, make sure it's
3797 * still big enough. Silently round up to minimum if not.
3799 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3800 if (env->me_mapsize < minsize)
3801 env->me_mapsize = minsize;
3803 if (env->me_mapsize != meta.mm_mapsize)
3804 env->me_flags |= MDB_RESIZING;
3807 rc = mdb_env_map(env, meta.mm_address, newenv || env->me_mapsize != meta.mm_mapsize);
3812 if (flags & MDB_FIXEDMAP)
3813 meta.mm_address = env->me_map;
3814 i = mdb_env_init_meta(env, &meta);
3815 if (i != MDB_SUCCESS) {
3820 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3821 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3823 #if !(MDB_MAXKEYSIZE)
3824 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3826 env->me_maxpg = env->me_mapsize / env->me_psize;
3830 int toggle = mdb_env_pick_meta(env);
3831 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3833 DPRINTF(("opened database version %u, pagesize %u",
3834 env->me_metas[0]->mm_version, env->me_psize));
3835 DPRINTF(("using meta page %d", toggle));
3836 DPRINTF(("depth: %u", db->md_depth));
3837 DPRINTF(("entries: %"Z"u", db->md_entries));
3838 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3839 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3840 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3841 DPRINTF(("root: %"Z"u", db->md_root));
3849 /** Release a reader thread's slot in the reader lock table.
3850 * This function is called automatically when a thread exits.
3851 * @param[in] ptr This points to the slot in the reader lock table.
3854 mdb_env_reader_dest(void *ptr)
3856 MDB_reader *reader = ptr;
3862 /** Junk for arranging thread-specific callbacks on Windows. This is
3863 * necessarily platform and compiler-specific. Windows supports up
3864 * to 1088 keys. Let's assume nobody opens more than 64 environments
3865 * in a single process, for now. They can override this if needed.
3867 #ifndef MAX_TLS_KEYS
3868 #define MAX_TLS_KEYS 64
3870 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3871 static int mdb_tls_nkeys;
3873 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3877 case DLL_PROCESS_ATTACH: break;
3878 case DLL_THREAD_ATTACH: break;
3879 case DLL_THREAD_DETACH:
3880 for (i=0; i<mdb_tls_nkeys; i++) {
3881 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3883 mdb_env_reader_dest(r);
3887 case DLL_PROCESS_DETACH: break;
3892 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3894 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3898 /* Force some symbol references.
3899 * _tls_used forces the linker to create the TLS directory if not already done
3900 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3902 #pragma comment(linker, "/INCLUDE:_tls_used")
3903 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3904 #pragma const_seg(".CRT$XLB")
3905 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3906 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3909 #pragma comment(linker, "/INCLUDE:__tls_used")
3910 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3911 #pragma data_seg(".CRT$XLB")
3912 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3914 #endif /* WIN 32/64 */
3915 #endif /* !__GNUC__ */
3918 /** Downgrade the exclusive lock on the region back to shared */
3920 mdb_env_share_locks(MDB_env *env, int *excl)
3922 int rc = 0, toggle = mdb_env_pick_meta(env);
3924 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3929 /* First acquire a shared lock. The Unlock will
3930 * then release the existing exclusive lock.
3932 memset(&ov, 0, sizeof(ov));
3933 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3936 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3942 struct flock lock_info;
3943 /* The shared lock replaces the existing lock */
3944 memset((void *)&lock_info, 0, sizeof(lock_info));
3945 lock_info.l_type = F_RDLCK;
3946 lock_info.l_whence = SEEK_SET;
3947 lock_info.l_start = 0;
3948 lock_info.l_len = 1;
3949 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3950 (rc = ErrCode()) == EINTR) ;
3951 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3958 /** Try to get exlusive lock, otherwise shared.
3959 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3962 mdb_env_excl_lock(MDB_env *env, int *excl)
3966 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3970 memset(&ov, 0, sizeof(ov));
3971 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3978 struct flock lock_info;
3979 memset((void *)&lock_info, 0, sizeof(lock_info));
3980 lock_info.l_type = F_WRLCK;
3981 lock_info.l_whence = SEEK_SET;
3982 lock_info.l_start = 0;
3983 lock_info.l_len = 1;
3984 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3985 (rc = ErrCode()) == EINTR) ;
3989 # ifdef MDB_USE_POSIX_SEM
3990 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3993 lock_info.l_type = F_RDLCK;
3994 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3995 (rc = ErrCode()) == EINTR) ;
4005 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4007 * @(#) $Revision: 5.1 $
4008 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4009 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4011 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4015 * Please do not copyright this code. This code is in the public domain.
4017 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4018 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4019 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4020 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4021 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4022 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4023 * PERFORMANCE OF THIS SOFTWARE.
4026 * chongo <Landon Curt Noll> /\oo/\
4027 * http://www.isthe.com/chongo/
4029 * Share and Enjoy! :-)
4032 typedef unsigned long long mdb_hash_t;
4033 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4035 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4036 * @param[in] val value to hash
4037 * @param[in] hval initial value for hash
4038 * @return 64 bit hash
4040 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4041 * hval arg on the first call.
4044 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4046 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4047 unsigned char *end = s + val->mv_size;
4049 * FNV-1a hash each octet of the string
4052 /* xor the bottom with the current octet */
4053 hval ^= (mdb_hash_t)*s++;
4055 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4056 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4057 (hval << 7) + (hval << 8) + (hval << 40);
4059 /* return our new hash value */
4063 /** Hash the string and output the encoded hash.
4064 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4065 * very short name limits. We don't care about the encoding being reversible,
4066 * we just want to preserve as many bits of the input as possible in a
4067 * small printable string.
4068 * @param[in] str string to hash
4069 * @param[out] encbuf an array of 11 chars to hold the hash
4071 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4074 mdb_pack85(unsigned long l, char *out)
4078 for (i=0; i<5; i++) {
4079 *out++ = mdb_a85[l % 85];
4085 mdb_hash_enc(MDB_val *val, char *encbuf)
4087 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4089 mdb_pack85(h, encbuf);
4090 mdb_pack85(h>>32, encbuf+5);
4095 /** Open and/or initialize the lock region for the environment.
4096 * @param[in] env The LMDB environment.
4097 * @param[in] lpath The pathname of the file used for the lock region.
4098 * @param[in] mode The Unix permissions for the file, if we create it.
4099 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4100 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4101 * @return 0 on success, non-zero on failure.
4104 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4107 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4109 # define MDB_ERRCODE_ROFS EROFS
4110 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4111 # define MDB_CLOEXEC O_CLOEXEC
4114 # define MDB_CLOEXEC 0
4121 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4122 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4123 FILE_ATTRIBUTE_NORMAL, NULL);
4125 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4127 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4129 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4134 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4135 /* Lose record locks when exec*() */
4136 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4137 fcntl(env->me_lfd, F_SETFD, fdflags);
4140 if (!(env->me_flags & MDB_NOTLS)) {
4141 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4144 env->me_flags |= MDB_ENV_TXKEY;
4146 /* Windows TLS callbacks need help finding their TLS info. */
4147 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4151 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4155 /* Try to get exclusive lock. If we succeed, then
4156 * nobody is using the lock region and we should initialize it.
4158 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4161 size = GetFileSize(env->me_lfd, NULL);
4163 size = lseek(env->me_lfd, 0, SEEK_END);
4164 if (size == -1) goto fail_errno;
4166 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4167 if (size < rsize && *excl > 0) {
4169 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4170 || !SetEndOfFile(env->me_lfd))
4173 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4177 size = rsize - sizeof(MDB_txninfo);
4178 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4183 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4185 if (!mh) goto fail_errno;
4186 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4188 if (!env->me_txns) goto fail_errno;
4190 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4192 if (m == MAP_FAILED) goto fail_errno;
4198 BY_HANDLE_FILE_INFORMATION stbuf;
4207 if (!mdb_sec_inited) {
4208 InitializeSecurityDescriptor(&mdb_null_sd,
4209 SECURITY_DESCRIPTOR_REVISION);
4210 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4211 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4212 mdb_all_sa.bInheritHandle = FALSE;
4213 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4216 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4217 idbuf.volume = stbuf.dwVolumeSerialNumber;
4218 idbuf.nhigh = stbuf.nFileIndexHigh;
4219 idbuf.nlow = stbuf.nFileIndexLow;
4220 val.mv_data = &idbuf;
4221 val.mv_size = sizeof(idbuf);
4222 mdb_hash_enc(&val, encbuf);
4223 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4224 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4225 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4226 if (!env->me_rmutex) goto fail_errno;
4227 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4228 if (!env->me_wmutex) goto fail_errno;
4229 #elif defined(MDB_USE_POSIX_SEM)
4238 #if defined(__NetBSD__)
4239 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4241 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4242 idbuf.dev = stbuf.st_dev;
4243 idbuf.ino = stbuf.st_ino;
4244 val.mv_data = &idbuf;
4245 val.mv_size = sizeof(idbuf);
4246 mdb_hash_enc(&val, encbuf);
4247 #ifdef MDB_SHORT_SEMNAMES
4248 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4250 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4251 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4252 /* Clean up after a previous run, if needed: Try to
4253 * remove both semaphores before doing anything else.
4255 sem_unlink(env->me_txns->mti_rmname);
4256 sem_unlink(env->me_txns->mti_wmname);
4257 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4258 O_CREAT|O_EXCL, mode, 1);
4259 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4260 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4261 O_CREAT|O_EXCL, mode, 1);
4262 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4263 #else /* MDB_USE_POSIX_SEM */
4264 pthread_mutexattr_t mattr;
4266 if ((rc = pthread_mutexattr_init(&mattr))
4267 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4268 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4269 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4271 pthread_mutexattr_destroy(&mattr);
4272 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4274 env->me_txns->mti_magic = MDB_MAGIC;
4275 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4276 env->me_txns->mti_txnid = 0;
4277 env->me_txns->mti_numreaders = 0;
4280 if (env->me_txns->mti_magic != MDB_MAGIC) {
4281 DPUTS("lock region has invalid magic");
4285 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4286 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4287 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4288 rc = MDB_VERSION_MISMATCH;
4292 if (rc && rc != EACCES && rc != EAGAIN) {
4296 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4297 if (!env->me_rmutex) goto fail_errno;
4298 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4299 if (!env->me_wmutex) goto fail_errno;
4300 #elif defined(MDB_USE_POSIX_SEM)
4301 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4302 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4303 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4304 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4315 /** The name of the lock file in the DB environment */
4316 #define LOCKNAME "/lock.mdb"
4317 /** The name of the data file in the DB environment */
4318 #define DATANAME "/data.mdb"
4319 /** The suffix of the lock file when no subdir is used */
4320 #define LOCKSUFF "-lock"
4321 /** Only a subset of the @ref mdb_env flags can be changed
4322 * at runtime. Changing other flags requires closing the
4323 * environment and re-opening it with the new flags.
4325 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4326 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4327 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4329 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4330 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4334 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4336 int oflags, rc, len, excl = -1;
4337 char *lpath, *dpath;
4339 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4343 if (flags & MDB_NOSUBDIR) {
4344 rc = len + sizeof(LOCKSUFF) + len + 1;
4346 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4351 if (flags & MDB_NOSUBDIR) {
4352 dpath = lpath + len + sizeof(LOCKSUFF);
4353 sprintf(lpath, "%s" LOCKSUFF, path);
4354 strcpy(dpath, path);
4356 dpath = lpath + len + sizeof(LOCKNAME);
4357 sprintf(lpath, "%s" LOCKNAME, path);
4358 sprintf(dpath, "%s" DATANAME, path);
4362 flags |= env->me_flags;
4363 if (flags & MDB_RDONLY) {
4364 /* silently ignore WRITEMAP when we're only getting read access */
4365 flags &= ~MDB_WRITEMAP;
4367 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4368 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4371 env->me_flags = flags |= MDB_ENV_ACTIVE;
4375 env->me_path = strdup(path);
4376 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4377 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4378 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4379 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4384 /* For RDONLY, get lockfile after we know datafile exists */
4385 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4386 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4392 if (F_ISSET(flags, MDB_RDONLY)) {
4393 oflags = GENERIC_READ;
4394 len = OPEN_EXISTING;
4396 oflags = GENERIC_READ|GENERIC_WRITE;
4399 mode = FILE_ATTRIBUTE_NORMAL;
4400 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4401 NULL, len, mode, NULL);
4403 if (F_ISSET(flags, MDB_RDONLY))
4406 oflags = O_RDWR | O_CREAT;
4408 env->me_fd = open(dpath, oflags, mode);
4410 if (env->me_fd == INVALID_HANDLE_VALUE) {
4415 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4416 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4421 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4422 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4423 env->me_mfd = env->me_fd;
4425 /* Synchronous fd for meta writes. Needed even with
4426 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4429 len = OPEN_EXISTING;
4430 env->me_mfd = CreateFile(dpath, oflags,
4431 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4432 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4435 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4437 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4442 DPRINTF(("opened dbenv %p", (void *) env));
4444 rc = mdb_env_share_locks(env, &excl);
4448 if (!((flags & MDB_RDONLY) ||
4449 (env->me_pbuf = calloc(1, env->me_psize))))
4455 mdb_env_close0(env, excl);
4461 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4463 mdb_env_close0(MDB_env *env, int excl)
4467 if (!(env->me_flags & MDB_ENV_ACTIVE))
4470 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4471 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4472 free(env->me_dbxs[i].md_name.mv_data);
4475 free(env->me_dbiseqs);
4476 free(env->me_dbflags);
4479 free(env->me_dirty_list);
4480 mdb_midl_free(env->me_free_pgs);
4482 if (env->me_flags & MDB_ENV_TXKEY) {
4483 pthread_key_delete(env->me_txkey);
4485 /* Delete our key from the global list */
4486 for (i=0; i<mdb_tls_nkeys; i++)
4487 if (mdb_tls_keys[i] == env->me_txkey) {
4488 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4496 munmap(env->me_map, env->me_mapsize);
4498 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4499 (void) close(env->me_mfd);
4500 if (env->me_fd != INVALID_HANDLE_VALUE)
4501 (void) close(env->me_fd);
4503 MDB_PID_T pid = env->me_pid;
4504 /* Clearing readers is done in this function because
4505 * me_txkey with its destructor must be disabled first.
4507 for (i = env->me_numreaders; --i >= 0; )
4508 if (env->me_txns->mti_readers[i].mr_pid == pid)
4509 env->me_txns->mti_readers[i].mr_pid = 0;
4511 if (env->me_rmutex) {
4512 CloseHandle(env->me_rmutex);
4513 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4515 /* Windows automatically destroys the mutexes when
4516 * the last handle closes.
4518 #elif defined(MDB_USE_POSIX_SEM)
4519 if (env->me_rmutex != SEM_FAILED) {
4520 sem_close(env->me_rmutex);
4521 if (env->me_wmutex != SEM_FAILED)
4522 sem_close(env->me_wmutex);
4523 /* If we have the filelock: If we are the
4524 * only remaining user, clean up semaphores.
4527 mdb_env_excl_lock(env, &excl);
4529 sem_unlink(env->me_txns->mti_rmname);
4530 sem_unlink(env->me_txns->mti_wmname);
4534 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4536 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4539 /* Unlock the lockfile. Windows would have unlocked it
4540 * after closing anyway, but not necessarily at once.
4542 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4545 (void) close(env->me_lfd);
4548 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4553 mdb_env_close(MDB_env *env)
4560 VGMEMP_DESTROY(env);
4561 while ((dp = env->me_dpages) != NULL) {
4562 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4563 env->me_dpages = dp->mp_next;
4567 mdb_env_close0(env, 0);
4571 /** Compare two items pointing at aligned size_t's */
4573 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4575 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4576 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4579 /** Compare two items pointing at aligned unsigned int's */
4581 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4583 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4584 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4587 /** Compare two items pointing at unsigned ints of unknown alignment.
4588 * Nodes and keys are guaranteed to be 2-byte aligned.
4591 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4593 #if BYTE_ORDER == LITTLE_ENDIAN
4594 unsigned short *u, *c;
4597 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4598 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4601 } while(!x && u > (unsigned short *)a->mv_data);
4604 unsigned short *u, *c, *end;
4607 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4608 u = (unsigned short *)a->mv_data;
4609 c = (unsigned short *)b->mv_data;
4612 } while(!x && u < end);
4617 /** Compare two items pointing at size_t's of unknown alignment. */
4618 #ifdef MISALIGNED_OK
4619 # define mdb_cmp_clong mdb_cmp_long
4621 # define mdb_cmp_clong mdb_cmp_cint
4624 /** Compare two items lexically */
4626 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4633 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4639 diff = memcmp(a->mv_data, b->mv_data, len);
4640 return diff ? diff : len_diff<0 ? -1 : len_diff;
4643 /** Compare two items in reverse byte order */
4645 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4647 const unsigned char *p1, *p2, *p1_lim;
4651 p1_lim = (const unsigned char *)a->mv_data;
4652 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4653 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4655 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4661 while (p1 > p1_lim) {
4662 diff = *--p1 - *--p2;
4666 return len_diff<0 ? -1 : len_diff;
4669 /** Search for key within a page, using binary search.
4670 * Returns the smallest entry larger or equal to the key.
4671 * If exactp is non-null, stores whether the found entry was an exact match
4672 * in *exactp (1 or 0).
4673 * Updates the cursor index with the index of the found entry.
4674 * If no entry larger or equal to the key is found, returns NULL.
4677 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4679 unsigned int i = 0, nkeys;
4682 MDB_page *mp = mc->mc_pg[mc->mc_top];
4683 MDB_node *node = NULL;
4688 nkeys = NUMKEYS(mp);
4690 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4691 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4694 low = IS_LEAF(mp) ? 0 : 1;
4696 cmp = mc->mc_dbx->md_cmp;
4698 /* Branch pages have no data, so if using integer keys,
4699 * alignment is guaranteed. Use faster mdb_cmp_int.
4701 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4702 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4709 nodekey.mv_size = mc->mc_db->md_pad;
4710 node = NODEPTR(mp, 0); /* fake */
4711 while (low <= high) {
4712 i = (low + high) >> 1;
4713 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4714 rc = cmp(key, &nodekey);
4715 DPRINTF(("found leaf index %u [%s], rc = %i",
4716 i, DKEY(&nodekey), rc));
4725 while (low <= high) {
4726 i = (low + high) >> 1;
4728 node = NODEPTR(mp, i);
4729 nodekey.mv_size = NODEKSZ(node);
4730 nodekey.mv_data = NODEKEY(node);
4732 rc = cmp(key, &nodekey);
4735 DPRINTF(("found leaf index %u [%s], rc = %i",
4736 i, DKEY(&nodekey), rc));
4738 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4739 i, DKEY(&nodekey), NODEPGNO(node), rc));
4750 if (rc > 0) { /* Found entry is less than the key. */
4751 i++; /* Skip to get the smallest entry larger than key. */
4753 node = NODEPTR(mp, i);
4756 *exactp = (rc == 0 && nkeys > 0);
4757 /* store the key index */
4758 mc->mc_ki[mc->mc_top] = i;
4760 /* There is no entry larger or equal to the key. */
4763 /* nodeptr is fake for LEAF2 */
4769 mdb_cursor_adjust(MDB_cursor *mc, func)
4773 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4774 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4781 /** Pop a page off the top of the cursor's stack. */
4783 mdb_cursor_pop(MDB_cursor *mc)
4787 MDB_page *top = mc->mc_pg[mc->mc_top];
4793 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4794 DDBI(mc), (void *) mc));
4798 /** Push a page onto the top of the cursor's stack. */
4800 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4802 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4803 DDBI(mc), (void *) mc));
4805 if (mc->mc_snum >= CURSOR_STACK) {
4806 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4807 return MDB_CURSOR_FULL;
4810 mc->mc_top = mc->mc_snum++;
4811 mc->mc_pg[mc->mc_top] = mp;
4812 mc->mc_ki[mc->mc_top] = 0;
4817 /** Find the address of the page corresponding to a given page number.
4818 * @param[in] txn the transaction for this access.
4819 * @param[in] pgno the page number for the page to retrieve.
4820 * @param[out] ret address of a pointer where the page's address will be stored.
4821 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4822 * @return 0 on success, non-zero on failure.
4825 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4827 MDB_env *env = txn->mt_env;
4831 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4835 MDB_ID2L dl = tx2->mt_u.dirty_list;
4837 /* Spilled pages were dirtied in this txn and flushed
4838 * because the dirty list got full. Bring this page
4839 * back in from the map (but don't unspill it here,
4840 * leave that unless page_touch happens again).
4842 if (tx2->mt_spill_pgs) {
4843 MDB_ID pn = pgno << 1;
4844 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4845 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4846 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4851 unsigned x = mdb_mid2l_search(dl, pgno);
4852 if (x <= dl[0].mid && dl[x].mid == pgno) {
4858 } while ((tx2 = tx2->mt_parent) != NULL);
4861 if (pgno < txn->mt_next_pgno) {
4863 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4865 DPRINTF(("page %"Z"u not found", pgno));
4866 txn->mt_flags |= MDB_TXN_ERROR;
4867 return MDB_PAGE_NOTFOUND;
4877 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4878 * The cursor is at the root page, set up the rest of it.
4881 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4883 MDB_page *mp = mc->mc_pg[mc->mc_top];
4887 while (IS_BRANCH(mp)) {
4891 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4892 mdb_cassert(mc, NUMKEYS(mp) > 1);
4893 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4895 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4897 if (flags & MDB_PS_LAST)
4898 i = NUMKEYS(mp) - 1;
4901 node = mdb_node_search(mc, key, &exact);
4903 i = NUMKEYS(mp) - 1;
4905 i = mc->mc_ki[mc->mc_top];
4907 mdb_cassert(mc, i > 0);
4911 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4914 mdb_cassert(mc, i < NUMKEYS(mp));
4915 node = NODEPTR(mp, i);
4917 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4920 mc->mc_ki[mc->mc_top] = i;
4921 if ((rc = mdb_cursor_push(mc, mp)))
4924 if (flags & MDB_PS_MODIFY) {
4925 if ((rc = mdb_page_touch(mc)) != 0)
4927 mp = mc->mc_pg[mc->mc_top];
4932 DPRINTF(("internal error, index points to a %02X page!?",
4934 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4935 return MDB_CORRUPTED;
4938 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4939 key ? DKEY(key) : "null"));
4940 mc->mc_flags |= C_INITIALIZED;
4941 mc->mc_flags &= ~C_EOF;
4946 /** Search for the lowest key under the current branch page.
4947 * This just bypasses a NUMKEYS check in the current page
4948 * before calling mdb_page_search_root(), because the callers
4949 * are all in situations where the current page is known to
4953 mdb_page_search_lowest(MDB_cursor *mc)
4955 MDB_page *mp = mc->mc_pg[mc->mc_top];
4956 MDB_node *node = NODEPTR(mp, 0);
4959 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4962 mc->mc_ki[mc->mc_top] = 0;
4963 if ((rc = mdb_cursor_push(mc, mp)))
4965 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4968 /** Search for the page a given key should be in.
4969 * Push it and its parent pages on the cursor stack.
4970 * @param[in,out] mc the cursor for this operation.
4971 * @param[in] key the key to search for, or NULL for first/last page.
4972 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4973 * are touched (updated with new page numbers).
4974 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4975 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4976 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4977 * @return 0 on success, non-zero on failure.
4980 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4985 /* Make sure the txn is still viable, then find the root from
4986 * the txn's db table and set it as the root of the cursor's stack.
4988 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4989 DPUTS("transaction has failed, must abort");
4992 /* Make sure we're using an up-to-date root */
4993 if (*mc->mc_dbflag & DB_STALE) {
4995 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
4997 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4998 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5005 MDB_node *leaf = mdb_node_search(&mc2,
5006 &mc->mc_dbx->md_name, &exact);
5008 return MDB_NOTFOUND;
5009 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5012 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5014 /* The txn may not know this DBI, or another process may
5015 * have dropped and recreated the DB with other flags.
5017 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5018 return MDB_INCOMPATIBLE;
5019 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5021 *mc->mc_dbflag &= ~DB_STALE;
5023 root = mc->mc_db->md_root;
5025 if (root == P_INVALID) { /* Tree is empty. */
5026 DPUTS("tree is empty");
5027 return MDB_NOTFOUND;
5031 mdb_cassert(mc, root > 1);
5032 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5033 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5039 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5040 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5042 if (flags & MDB_PS_MODIFY) {
5043 if ((rc = mdb_page_touch(mc)))
5047 if (flags & MDB_PS_ROOTONLY)
5050 return mdb_page_search_root(mc, key, flags);
5054 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5056 MDB_txn *txn = mc->mc_txn;
5057 pgno_t pg = mp->mp_pgno;
5058 unsigned x = 0, ovpages = mp->mp_pages;
5059 MDB_env *env = txn->mt_env;
5060 MDB_IDL sl = txn->mt_spill_pgs;
5061 MDB_ID pn = pg << 1;
5064 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5065 /* If the page is dirty or on the spill list we just acquired it,
5066 * so we should give it back to our current free list, if any.
5067 * Otherwise put it onto the list of pages we freed in this txn.
5069 * Won't create me_pghead: me_pglast must be inited along with it.
5070 * Unsupported in nested txns: They would need to hide the page
5071 * range in ancestor txns' dirty and spilled lists.
5073 if (env->me_pghead &&
5075 ((mp->mp_flags & P_DIRTY) ||
5076 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5080 MDB_ID2 *dl, ix, iy;
5081 rc = mdb_midl_need(&env->me_pghead, ovpages);
5084 if (!(mp->mp_flags & P_DIRTY)) {
5085 /* This page is no longer spilled */
5092 /* Remove from dirty list */
5093 dl = txn->mt_u.dirty_list;
5095 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5101 mdb_cassert(mc, x > 1);
5103 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5104 txn->mt_flags |= MDB_TXN_ERROR;
5105 return MDB_CORRUPTED;
5108 if (!(env->me_flags & MDB_WRITEMAP))
5109 mdb_dpage_free(env, mp);
5111 /* Insert in me_pghead */
5112 mop = env->me_pghead;
5113 j = mop[0] + ovpages;
5114 for (i = mop[0]; i && mop[i] < pg; i--)
5120 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5124 mc->mc_db->md_overflow_pages -= ovpages;
5128 /** Return the data associated with a given node.
5129 * @param[in] txn The transaction for this operation.
5130 * @param[in] leaf The node being read.
5131 * @param[out] data Updated to point to the node's data.
5132 * @return 0 on success, non-zero on failure.
5135 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5137 MDB_page *omp; /* overflow page */
5141 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5142 data->mv_size = NODEDSZ(leaf);
5143 data->mv_data = NODEDATA(leaf);
5147 /* Read overflow data.
5149 data->mv_size = NODEDSZ(leaf);
5150 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5151 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5152 DPRINTF(("read overflow page %"Z"u failed", pgno));
5155 data->mv_data = METADATA(omp);
5161 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5162 MDB_val *key, MDB_val *data)
5169 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5171 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5174 if (txn->mt_flags & MDB_TXN_ERROR)
5177 mdb_cursor_init(&mc, txn, dbi, &mx);
5178 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5181 /** Find a sibling for a page.
5182 * Replaces the page at the top of the cursor's stack with the
5183 * specified sibling, if one exists.
5184 * @param[in] mc The cursor for this operation.
5185 * @param[in] move_right Non-zero if the right sibling is requested,
5186 * otherwise the left sibling.
5187 * @return 0 on success, non-zero on failure.
5190 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5196 if (mc->mc_snum < 2) {
5197 return MDB_NOTFOUND; /* root has no siblings */
5201 DPRINTF(("parent page is page %"Z"u, index %u",
5202 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5204 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5205 : (mc->mc_ki[mc->mc_top] == 0)) {
5206 DPRINTF(("no more keys left, moving to %s sibling",
5207 move_right ? "right" : "left"));
5208 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5209 /* undo cursor_pop before returning */
5216 mc->mc_ki[mc->mc_top]++;
5218 mc->mc_ki[mc->mc_top]--;
5219 DPRINTF(("just moving to %s index key %u",
5220 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5222 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5224 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5225 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5226 /* mc will be inconsistent if caller does mc_snum++ as above */
5227 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5231 mdb_cursor_push(mc, mp);
5233 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5238 /** Move the cursor to the next data item. */
5240 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5246 if (mc->mc_flags & C_EOF) {
5247 return MDB_NOTFOUND;
5250 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5252 mp = mc->mc_pg[mc->mc_top];
5254 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5255 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5256 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5257 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5258 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5259 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5260 if (rc == MDB_SUCCESS)
5261 MDB_GET_KEY(leaf, key);
5266 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5267 if (op == MDB_NEXT_DUP)
5268 return MDB_NOTFOUND;
5272 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5273 mdb_dbg_pgno(mp), (void *) mc));
5274 if (mc->mc_flags & C_DEL)
5277 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5278 DPUTS("=====> move to next sibling page");
5279 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5280 mc->mc_flags |= C_EOF;
5283 mp = mc->mc_pg[mc->mc_top];
5284 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5286 mc->mc_ki[mc->mc_top]++;
5289 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5290 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5293 key->mv_size = mc->mc_db->md_pad;
5294 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5298 mdb_cassert(mc, IS_LEAF(mp));
5299 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5301 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5302 mdb_xcursor_init1(mc, leaf);
5305 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5308 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5309 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5310 if (rc != MDB_SUCCESS)
5315 MDB_GET_KEY(leaf, key);
5319 /** Move the cursor to the previous data item. */
5321 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5327 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5329 mp = mc->mc_pg[mc->mc_top];
5331 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5332 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5333 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5334 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5335 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5336 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5337 if (rc == MDB_SUCCESS)
5338 MDB_GET_KEY(leaf, key);
5342 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5343 if (op == MDB_PREV_DUP)
5344 return MDB_NOTFOUND;
5349 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5350 mdb_dbg_pgno(mp), (void *) mc));
5352 if (mc->mc_ki[mc->mc_top] == 0) {
5353 DPUTS("=====> move to prev sibling page");
5354 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5357 mp = mc->mc_pg[mc->mc_top];
5358 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5359 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5361 mc->mc_ki[mc->mc_top]--;
5363 mc->mc_flags &= ~C_EOF;
5365 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5366 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5369 key->mv_size = mc->mc_db->md_pad;
5370 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5374 mdb_cassert(mc, IS_LEAF(mp));
5375 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5377 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5378 mdb_xcursor_init1(mc, leaf);
5381 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5384 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5385 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5386 if (rc != MDB_SUCCESS)
5391 MDB_GET_KEY(leaf, key);
5395 /** Set the cursor on a specific data item. */
5397 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5398 MDB_cursor_op op, int *exactp)
5402 MDB_node *leaf = NULL;
5405 if (key->mv_size == 0)
5406 return MDB_BAD_VALSIZE;
5409 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5411 /* See if we're already on the right page */
5412 if (mc->mc_flags & C_INITIALIZED) {
5415 mp = mc->mc_pg[mc->mc_top];
5417 mc->mc_ki[mc->mc_top] = 0;
5418 return MDB_NOTFOUND;
5420 if (mp->mp_flags & P_LEAF2) {
5421 nodekey.mv_size = mc->mc_db->md_pad;
5422 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5424 leaf = NODEPTR(mp, 0);
5425 MDB_GET_KEY2(leaf, nodekey);
5427 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5429 /* Probably happens rarely, but first node on the page
5430 * was the one we wanted.
5432 mc->mc_ki[mc->mc_top] = 0;
5439 unsigned int nkeys = NUMKEYS(mp);
5441 if (mp->mp_flags & P_LEAF2) {
5442 nodekey.mv_data = LEAF2KEY(mp,
5443 nkeys-1, nodekey.mv_size);
5445 leaf = NODEPTR(mp, nkeys-1);
5446 MDB_GET_KEY2(leaf, nodekey);
5448 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5450 /* last node was the one we wanted */
5451 mc->mc_ki[mc->mc_top] = nkeys-1;
5457 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5458 /* This is definitely the right page, skip search_page */
5459 if (mp->mp_flags & P_LEAF2) {
5460 nodekey.mv_data = LEAF2KEY(mp,
5461 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5463 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5464 MDB_GET_KEY2(leaf, nodekey);
5466 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5468 /* current node was the one we wanted */
5478 /* If any parents have right-sibs, search.
5479 * Otherwise, there's nothing further.
5481 for (i=0; i<mc->mc_top; i++)
5483 NUMKEYS(mc->mc_pg[i])-1)
5485 if (i == mc->mc_top) {
5486 /* There are no other pages */
5487 mc->mc_ki[mc->mc_top] = nkeys;
5488 return MDB_NOTFOUND;
5492 /* There are no other pages */
5493 mc->mc_ki[mc->mc_top] = 0;
5494 if (op == MDB_SET_RANGE && !exactp) {
5498 return MDB_NOTFOUND;
5502 rc = mdb_page_search(mc, key, 0);
5503 if (rc != MDB_SUCCESS)
5506 mp = mc->mc_pg[mc->mc_top];
5507 mdb_cassert(mc, IS_LEAF(mp));
5510 leaf = mdb_node_search(mc, key, exactp);
5511 if (exactp != NULL && !*exactp) {
5512 /* MDB_SET specified and not an exact match. */
5513 return MDB_NOTFOUND;
5517 DPUTS("===> inexact leaf not found, goto sibling");
5518 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5519 return rc; /* no entries matched */
5520 mp = mc->mc_pg[mc->mc_top];
5521 mdb_cassert(mc, IS_LEAF(mp));
5522 leaf = NODEPTR(mp, 0);
5526 mc->mc_flags |= C_INITIALIZED;
5527 mc->mc_flags &= ~C_EOF;
5530 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5531 key->mv_size = mc->mc_db->md_pad;
5532 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5537 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5538 mdb_xcursor_init1(mc, leaf);
5541 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5542 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5543 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5546 if (op == MDB_GET_BOTH) {
5552 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5553 if (rc != MDB_SUCCESS)
5556 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5558 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5560 rc = mc->mc_dbx->md_dcmp(data, &d2);
5562 if (op == MDB_GET_BOTH || rc > 0)
5563 return MDB_NOTFOUND;
5570 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5571 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5576 /* The key already matches in all other cases */
5577 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5578 MDB_GET_KEY(leaf, key);
5579 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5584 /** Move the cursor to the first item in the database. */
5586 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5592 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5594 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5595 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5596 if (rc != MDB_SUCCESS)
5599 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5601 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5602 mc->mc_flags |= C_INITIALIZED;
5603 mc->mc_flags &= ~C_EOF;
5605 mc->mc_ki[mc->mc_top] = 0;
5607 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5608 key->mv_size = mc->mc_db->md_pad;
5609 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5614 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5615 mdb_xcursor_init1(mc, leaf);
5616 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5620 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5624 MDB_GET_KEY(leaf, key);
5628 /** Move the cursor to the last item in the database. */
5630 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5636 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5638 if (!(mc->mc_flags & C_EOF)) {
5640 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5641 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5642 if (rc != MDB_SUCCESS)
5645 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5648 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5649 mc->mc_flags |= C_INITIALIZED|C_EOF;
5650 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5652 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5653 key->mv_size = mc->mc_db->md_pad;
5654 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5659 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5660 mdb_xcursor_init1(mc, leaf);
5661 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5665 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5670 MDB_GET_KEY(leaf, key);
5675 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5680 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5685 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5689 case MDB_GET_CURRENT:
5690 if (!(mc->mc_flags & C_INITIALIZED)) {
5693 MDB_page *mp = mc->mc_pg[mc->mc_top];
5694 int nkeys = NUMKEYS(mp);
5695 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5696 mc->mc_ki[mc->mc_top] = nkeys;
5702 key->mv_size = mc->mc_db->md_pad;
5703 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5705 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5706 MDB_GET_KEY(leaf, key);
5708 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5709 if (mc->mc_flags & C_DEL)
5710 mdb_xcursor_init1(mc, leaf);
5711 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5713 rc = mdb_node_read(mc->mc_txn, leaf, data);
5720 case MDB_GET_BOTH_RANGE:
5725 if (mc->mc_xcursor == NULL) {
5726 rc = MDB_INCOMPATIBLE;
5736 rc = mdb_cursor_set(mc, key, data, op,
5737 op == MDB_SET_RANGE ? NULL : &exact);
5740 case MDB_GET_MULTIPLE:
5741 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5745 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5746 rc = MDB_INCOMPATIBLE;
5750 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5751 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5754 case MDB_NEXT_MULTIPLE:
5759 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5760 rc = MDB_INCOMPATIBLE;
5763 if (!(mc->mc_flags & C_INITIALIZED))
5764 rc = mdb_cursor_first(mc, key, data);
5766 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5767 if (rc == MDB_SUCCESS) {
5768 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5771 mx = &mc->mc_xcursor->mx_cursor;
5772 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5774 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5775 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5783 case MDB_NEXT_NODUP:
5784 if (!(mc->mc_flags & C_INITIALIZED))
5785 rc = mdb_cursor_first(mc, key, data);
5787 rc = mdb_cursor_next(mc, key, data, op);
5791 case MDB_PREV_NODUP:
5792 if (!(mc->mc_flags & C_INITIALIZED)) {
5793 rc = mdb_cursor_last(mc, key, data);
5796 mc->mc_flags |= C_INITIALIZED;
5797 mc->mc_ki[mc->mc_top]++;
5799 rc = mdb_cursor_prev(mc, key, data, op);
5802 rc = mdb_cursor_first(mc, key, data);
5805 mfunc = mdb_cursor_first;
5807 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5811 if (mc->mc_xcursor == NULL) {
5812 rc = MDB_INCOMPATIBLE;
5815 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5819 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5822 rc = mdb_cursor_last(mc, key, data);
5825 mfunc = mdb_cursor_last;
5828 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5833 if (mc->mc_flags & C_DEL)
5834 mc->mc_flags ^= C_DEL;
5839 /** Touch all the pages in the cursor stack. Set mc_top.
5840 * Makes sure all the pages are writable, before attempting a write operation.
5841 * @param[in] mc The cursor to operate on.
5844 mdb_cursor_touch(MDB_cursor *mc)
5846 int rc = MDB_SUCCESS;
5848 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5851 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5853 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5854 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5857 *mc->mc_dbflag |= DB_DIRTY;
5862 rc = mdb_page_touch(mc);
5863 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5864 mc->mc_top = mc->mc_snum-1;
5869 /** Do not spill pages to disk if txn is getting full, may fail instead */
5870 #define MDB_NOSPILL 0x8000
5873 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5876 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5878 MDB_node *leaf = NULL;
5881 MDB_val xdata, *rdata, dkey, olddata;
5883 int do_sub = 0, insert_key, insert_data;
5884 unsigned int mcount = 0, dcount = 0, nospill;
5887 unsigned int nflags;
5890 if (mc == NULL || key == NULL)
5893 env = mc->mc_txn->mt_env;
5895 /* Check this first so counter will always be zero on any
5898 if (flags & MDB_MULTIPLE) {
5899 dcount = data[1].mv_size;
5900 data[1].mv_size = 0;
5901 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5902 return MDB_INCOMPATIBLE;
5905 nospill = flags & MDB_NOSPILL;
5906 flags &= ~MDB_NOSPILL;
5908 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5909 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5911 if (key->mv_size-1 >= ENV_MAXKEY(env))
5912 return MDB_BAD_VALSIZE;
5914 #if SIZE_MAX > MAXDATASIZE
5915 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5916 return MDB_BAD_VALSIZE;
5918 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5919 return MDB_BAD_VALSIZE;
5922 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5923 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5927 if (flags == MDB_CURRENT) {
5928 if (!(mc->mc_flags & C_INITIALIZED))
5931 } else if (mc->mc_db->md_root == P_INVALID) {
5932 /* new database, cursor has nothing to point to */
5935 mc->mc_flags &= ~C_INITIALIZED;
5940 if (flags & MDB_APPEND) {
5942 rc = mdb_cursor_last(mc, &k2, &d2);
5944 rc = mc->mc_dbx->md_cmp(key, &k2);
5947 mc->mc_ki[mc->mc_top]++;
5949 /* new key is <= last key */
5954 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5956 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5957 DPRINTF(("duplicate key [%s]", DKEY(key)));
5959 return MDB_KEYEXIST;
5961 if (rc && rc != MDB_NOTFOUND)
5965 if (mc->mc_flags & C_DEL)
5966 mc->mc_flags ^= C_DEL;
5968 /* Cursor is positioned, check for room in the dirty list */
5970 if (flags & MDB_MULTIPLE) {
5972 xdata.mv_size = data->mv_size * dcount;
5976 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5980 if (rc == MDB_NO_ROOT) {
5982 /* new database, write a root leaf page */
5983 DPUTS("allocating new root leaf page");
5984 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5987 mdb_cursor_push(mc, np);
5988 mc->mc_db->md_root = np->mp_pgno;
5989 mc->mc_db->md_depth++;
5990 *mc->mc_dbflag |= DB_DIRTY;
5991 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5993 np->mp_flags |= P_LEAF2;
5994 mc->mc_flags |= C_INITIALIZED;
5996 /* make sure all cursor pages are writable */
5997 rc2 = mdb_cursor_touch(mc);
6002 insert_key = insert_data = rc;
6004 /* The key does not exist */
6005 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6006 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6007 LEAFSIZE(key, data) > env->me_nodemax)
6009 /* Too big for a node, insert in sub-DB. Set up an empty
6010 * "old sub-page" for prep_subDB to expand to a full page.
6012 fp_flags = P_LEAF|P_DIRTY;
6014 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6015 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6016 olddata.mv_size = PAGEHDRSZ;
6020 /* there's only a key anyway, so this is a no-op */
6021 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6023 unsigned int ksize = mc->mc_db->md_pad;
6024 if (key->mv_size != ksize)
6025 return MDB_BAD_VALSIZE;
6026 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6027 memcpy(ptr, key->mv_data, ksize);
6032 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6033 olddata.mv_size = NODEDSZ(leaf);
6034 olddata.mv_data = NODEDATA(leaf);
6037 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6038 /* Prepare (sub-)page/sub-DB to accept the new item,
6039 * if needed. fp: old sub-page or a header faking
6040 * it. mp: new (sub-)page. offset: growth in page
6041 * size. xdata: node data with new page or DB.
6043 unsigned i, offset = 0;
6044 mp = fp = xdata.mv_data = env->me_pbuf;
6045 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6047 /* Was a single item before, must convert now */
6048 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6049 /* Just overwrite the current item */
6050 if (flags == MDB_CURRENT)
6053 #if UINT_MAX < SIZE_MAX
6054 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6055 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6057 /* does data match? */
6058 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6059 if (flags & MDB_NODUPDATA)
6060 return MDB_KEYEXIST;
6065 /* Back up original data item */
6066 dkey.mv_size = olddata.mv_size;
6067 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6069 /* Make sub-page header for the dup items, with dummy body */
6070 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6071 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6072 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6073 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6074 fp->mp_flags |= P_LEAF2;
6075 fp->mp_pad = data->mv_size;
6076 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6078 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6079 (dkey.mv_size & 1) + (data->mv_size & 1);
6081 fp->mp_upper = xdata.mv_size - PAGEBASE;
6082 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6083 } else if (leaf->mn_flags & F_SUBDATA) {
6084 /* Data is on sub-DB, just store it */
6085 flags |= F_DUPDATA|F_SUBDATA;
6088 /* Data is on sub-page */
6089 fp = olddata.mv_data;
6092 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6093 offset = EVEN(NODESIZE + sizeof(indx_t) +
6097 offset = fp->mp_pad;
6098 if (SIZELEFT(fp) < offset) {
6099 offset *= 4; /* space for 4 more */
6102 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6104 fp->mp_flags |= P_DIRTY;
6105 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6106 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6110 xdata.mv_size = olddata.mv_size + offset;
6113 fp_flags = fp->mp_flags;
6114 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6115 /* Too big for a sub-page, convert to sub-DB */
6116 fp_flags &= ~P_SUBP;
6118 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6119 fp_flags |= P_LEAF2;
6120 dummy.md_pad = fp->mp_pad;
6121 dummy.md_flags = MDB_DUPFIXED;
6122 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6123 dummy.md_flags |= MDB_INTEGERKEY;
6129 dummy.md_branch_pages = 0;
6130 dummy.md_leaf_pages = 1;
6131 dummy.md_overflow_pages = 0;
6132 dummy.md_entries = NUMKEYS(fp);
6133 xdata.mv_size = sizeof(MDB_db);
6134 xdata.mv_data = &dummy;
6135 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6137 offset = env->me_psize - olddata.mv_size;
6138 flags |= F_DUPDATA|F_SUBDATA;
6139 dummy.md_root = mp->mp_pgno;
6142 mp->mp_flags = fp_flags | P_DIRTY;
6143 mp->mp_pad = fp->mp_pad;
6144 mp->mp_lower = fp->mp_lower;
6145 mp->mp_upper = fp->mp_upper + offset;
6146 if (fp_flags & P_LEAF2) {
6147 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6149 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6150 olddata.mv_size - fp->mp_upper - PAGEBASE);
6151 for (i=0; i<NUMKEYS(fp); i++)
6152 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6160 mdb_node_del(mc, 0);
6164 /* overflow page overwrites need special handling */
6165 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6168 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6170 memcpy(&pg, olddata.mv_data, sizeof(pg));
6171 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6173 ovpages = omp->mp_pages;
6175 /* Is the ov page large enough? */
6176 if (ovpages >= dpages) {
6177 if (!(omp->mp_flags & P_DIRTY) &&
6178 (level || (env->me_flags & MDB_WRITEMAP)))
6180 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6183 level = 0; /* dirty in this txn or clean */
6186 if (omp->mp_flags & P_DIRTY) {
6187 /* yes, overwrite it. Note in this case we don't
6188 * bother to try shrinking the page if the new data
6189 * is smaller than the overflow threshold.
6192 /* It is writable only in a parent txn */
6193 size_t sz = (size_t) env->me_psize * ovpages, off;
6194 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6200 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6201 mdb_cassert(mc, rc2 == 0);
6202 if (!(flags & MDB_RESERVE)) {
6203 /* Copy end of page, adjusting alignment so
6204 * compiler may copy words instead of bytes.
6206 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6207 memcpy((size_t *)((char *)np + off),
6208 (size_t *)((char *)omp + off), sz - off);
6211 memcpy(np, omp, sz); /* Copy beginning of page */
6214 SETDSZ(leaf, data->mv_size);
6215 if (F_ISSET(flags, MDB_RESERVE))
6216 data->mv_data = METADATA(omp);
6218 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6222 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6224 } else if (data->mv_size == olddata.mv_size) {
6225 /* same size, just replace it. Note that we could
6226 * also reuse this node if the new data is smaller,
6227 * but instead we opt to shrink the node in that case.
6229 if (F_ISSET(flags, MDB_RESERVE))
6230 data->mv_data = olddata.mv_data;
6231 else if (!(mc->mc_flags & C_SUB))
6232 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6234 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6237 mdb_node_del(mc, 0);
6243 nflags = flags & NODE_ADD_FLAGS;
6244 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6245 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6246 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6247 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6249 nflags |= MDB_SPLIT_REPLACE;
6250 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6252 /* There is room already in this leaf page. */
6253 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6254 if (rc == 0 && insert_key) {
6255 /* Adjust other cursors pointing to mp */
6256 MDB_cursor *m2, *m3;
6257 MDB_dbi dbi = mc->mc_dbi;
6258 unsigned i = mc->mc_top;
6259 MDB_page *mp = mc->mc_pg[i];
6261 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6262 if (mc->mc_flags & C_SUB)
6263 m3 = &m2->mc_xcursor->mx_cursor;
6266 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6267 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6274 if (rc == MDB_SUCCESS) {
6275 /* Now store the actual data in the child DB. Note that we're
6276 * storing the user data in the keys field, so there are strict
6277 * size limits on dupdata. The actual data fields of the child
6278 * DB are all zero size.
6286 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6287 if (flags & MDB_CURRENT) {
6288 xflags = MDB_CURRENT|MDB_NOSPILL;
6290 mdb_xcursor_init1(mc, leaf);
6291 xflags = (flags & MDB_NODUPDATA) ?
6292 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6294 /* converted, write the original data first */
6296 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6300 /* Adjust other cursors pointing to mp */
6302 unsigned i = mc->mc_top;
6303 MDB_page *mp = mc->mc_pg[i];
6305 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6306 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6307 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6308 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6309 mdb_xcursor_init1(m2, leaf);
6313 /* we've done our job */
6316 ecount = mc->mc_xcursor->mx_db.md_entries;
6317 if (flags & MDB_APPENDDUP)
6318 xflags |= MDB_APPEND;
6319 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6320 if (flags & F_SUBDATA) {
6321 void *db = NODEDATA(leaf);
6322 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6324 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6326 /* Increment count unless we just replaced an existing item. */
6328 mc->mc_db->md_entries++;
6330 /* Invalidate txn if we created an empty sub-DB */
6333 /* If we succeeded and the key didn't exist before,
6334 * make sure the cursor is marked valid.
6336 mc->mc_flags |= C_INITIALIZED;
6338 if (flags & MDB_MULTIPLE) {
6341 /* let caller know how many succeeded, if any */
6342 data[1].mv_size = mcount;
6343 if (mcount < dcount) {
6344 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6345 insert_key = insert_data = 0;
6352 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6355 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6360 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6366 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6367 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6369 if (!(mc->mc_flags & C_INITIALIZED))
6372 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6373 return MDB_NOTFOUND;
6375 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6378 rc = mdb_cursor_touch(mc);
6382 mp = mc->mc_pg[mc->mc_top];
6385 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6387 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6388 if (flags & MDB_NODUPDATA) {
6389 /* mdb_cursor_del0() will subtract the final entry */
6390 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6392 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6393 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6395 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6398 /* If sub-DB still has entries, we're done */
6399 if (mc->mc_xcursor->mx_db.md_entries) {
6400 if (leaf->mn_flags & F_SUBDATA) {
6401 /* update subDB info */
6402 void *db = NODEDATA(leaf);
6403 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6406 /* shrink fake page */
6407 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6408 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6409 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6410 /* fix other sub-DB cursors pointed at this fake page */
6411 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6412 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6413 if (m2->mc_pg[mc->mc_top] == mp &&
6414 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6415 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6418 mc->mc_db->md_entries--;
6419 mc->mc_flags |= C_DEL;
6422 /* otherwise fall thru and delete the sub-DB */
6425 if (leaf->mn_flags & F_SUBDATA) {
6426 /* add all the child DB's pages to the free list */
6427 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6433 /* add overflow pages to free list */
6434 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6438 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6439 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6440 (rc = mdb_ovpage_free(mc, omp)))
6445 return mdb_cursor_del0(mc);
6448 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6452 /** Allocate and initialize new pages for a database.
6453 * @param[in] mc a cursor on the database being added to.
6454 * @param[in] flags flags defining what type of page is being allocated.
6455 * @param[in] num the number of pages to allocate. This is usually 1,
6456 * unless allocating overflow pages for a large record.
6457 * @param[out] mp Address of a page, or NULL on failure.
6458 * @return 0 on success, non-zero on failure.
6461 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6466 if ((rc = mdb_page_alloc(mc, num, &np)))
6468 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6469 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6470 np->mp_flags = flags | P_DIRTY;
6471 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6472 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6475 mc->mc_db->md_branch_pages++;
6476 else if (IS_LEAF(np))
6477 mc->mc_db->md_leaf_pages++;
6478 else if (IS_OVERFLOW(np)) {
6479 mc->mc_db->md_overflow_pages += num;
6487 /** Calculate the size of a leaf node.
6488 * The size depends on the environment's page size; if a data item
6489 * is too large it will be put onto an overflow page and the node
6490 * size will only include the key and not the data. Sizes are always
6491 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6492 * of the #MDB_node headers.
6493 * @param[in] env The environment handle.
6494 * @param[in] key The key for the node.
6495 * @param[in] data The data for the node.
6496 * @return The number of bytes needed to store the node.
6499 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6503 sz = LEAFSIZE(key, data);
6504 if (sz > env->me_nodemax) {
6505 /* put on overflow page */
6506 sz -= data->mv_size - sizeof(pgno_t);
6509 return EVEN(sz + sizeof(indx_t));
6512 /** Calculate the size of a branch node.
6513 * The size should depend on the environment's page size but since
6514 * we currently don't support spilling large keys onto overflow
6515 * pages, it's simply the size of the #MDB_node header plus the
6516 * size of the key. Sizes are always rounded up to an even number
6517 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6518 * @param[in] env The environment handle.
6519 * @param[in] key The key for the node.
6520 * @return The number of bytes needed to store the node.
6523 mdb_branch_size(MDB_env *env, MDB_val *key)
6528 if (sz > env->me_nodemax) {
6529 /* put on overflow page */
6530 /* not implemented */
6531 /* sz -= key->size - sizeof(pgno_t); */
6534 return sz + sizeof(indx_t);
6537 /** Add a node to the page pointed to by the cursor.
6538 * @param[in] mc The cursor for this operation.
6539 * @param[in] indx The index on the page where the new node should be added.
6540 * @param[in] key The key for the new node.
6541 * @param[in] data The data for the new node, if any.
6542 * @param[in] pgno The page number, if adding a branch node.
6543 * @param[in] flags Flags for the node.
6544 * @return 0 on success, non-zero on failure. Possible errors are:
6546 * <li>ENOMEM - failed to allocate overflow pages for the node.
6547 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6548 * should never happen since all callers already calculate the
6549 * page's free space before calling this function.
6553 mdb_node_add(MDB_cursor *mc, indx_t indx,
6554 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6557 size_t node_size = NODESIZE;
6561 MDB_page *mp = mc->mc_pg[mc->mc_top];
6562 MDB_page *ofp = NULL; /* overflow page */
6565 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6567 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6568 IS_LEAF(mp) ? "leaf" : "branch",
6569 IS_SUBP(mp) ? "sub-" : "",
6570 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6571 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6574 /* Move higher keys up one slot. */
6575 int ksize = mc->mc_db->md_pad, dif;
6576 char *ptr = LEAF2KEY(mp, indx, ksize);
6577 dif = NUMKEYS(mp) - indx;
6579 memmove(ptr+ksize, ptr, dif*ksize);
6580 /* insert new key */
6581 memcpy(ptr, key->mv_data, ksize);
6583 /* Just using these for counting */
6584 mp->mp_lower += sizeof(indx_t);
6585 mp->mp_upper -= ksize - sizeof(indx_t);
6589 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6591 node_size += key->mv_size;
6593 mdb_cassert(mc, data);
6594 if (F_ISSET(flags, F_BIGDATA)) {
6595 /* Data already on overflow page. */
6596 node_size += sizeof(pgno_t);
6597 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6598 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6600 /* Put data on overflow page. */
6601 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6602 data->mv_size, node_size+data->mv_size));
6603 node_size = EVEN(node_size + sizeof(pgno_t));
6604 if ((ssize_t)node_size > room)
6606 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6608 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6612 node_size += data->mv_size;
6615 node_size = EVEN(node_size);
6616 if ((ssize_t)node_size > room)
6620 /* Move higher pointers up one slot. */
6621 for (i = NUMKEYS(mp); i > indx; i--)
6622 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6624 /* Adjust free space offsets. */
6625 ofs = mp->mp_upper - node_size;
6626 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6627 mp->mp_ptrs[indx] = ofs;
6629 mp->mp_lower += sizeof(indx_t);
6631 /* Write the node data. */
6632 node = NODEPTR(mp, indx);
6633 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6634 node->mn_flags = flags;
6636 SETDSZ(node,data->mv_size);
6641 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6644 mdb_cassert(mc, key);
6646 if (F_ISSET(flags, F_BIGDATA))
6647 memcpy(node->mn_data + key->mv_size, data->mv_data,
6649 else if (F_ISSET(flags, MDB_RESERVE))
6650 data->mv_data = node->mn_data + key->mv_size;
6652 memcpy(node->mn_data + key->mv_size, data->mv_data,
6655 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6657 if (F_ISSET(flags, MDB_RESERVE))
6658 data->mv_data = METADATA(ofp);
6660 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6667 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6668 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6669 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6670 DPRINTF(("node size = %"Z"u", node_size));
6671 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6672 return MDB_PAGE_FULL;
6675 /** Delete the specified node from a page.
6676 * @param[in] mc Cursor pointing to the node to delete.
6677 * @param[in] ksize The size of a node. Only used if the page is
6678 * part of a #MDB_DUPFIXED database.
6681 mdb_node_del(MDB_cursor *mc, int ksize)
6683 MDB_page *mp = mc->mc_pg[mc->mc_top];
6684 indx_t indx = mc->mc_ki[mc->mc_top];
6686 indx_t i, j, numkeys, ptr;
6690 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6691 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6692 numkeys = NUMKEYS(mp);
6693 mdb_cassert(mc, indx < numkeys);
6696 int x = numkeys - 1 - indx;
6697 base = LEAF2KEY(mp, indx, ksize);
6699 memmove(base, base + ksize, x * ksize);
6700 mp->mp_lower -= sizeof(indx_t);
6701 mp->mp_upper += ksize - sizeof(indx_t);
6705 node = NODEPTR(mp, indx);
6706 sz = NODESIZE + node->mn_ksize;
6708 if (F_ISSET(node->mn_flags, F_BIGDATA))
6709 sz += sizeof(pgno_t);
6711 sz += NODEDSZ(node);
6715 ptr = mp->mp_ptrs[indx];
6716 for (i = j = 0; i < numkeys; i++) {
6718 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6719 if (mp->mp_ptrs[i] < ptr)
6720 mp->mp_ptrs[j] += sz;
6725 base = (char *)mp + mp->mp_upper + PAGEBASE;
6726 memmove(base + sz, base, ptr - mp->mp_upper);
6728 mp->mp_lower -= sizeof(indx_t);
6732 /** Compact the main page after deleting a node on a subpage.
6733 * @param[in] mp The main page to operate on.
6734 * @param[in] indx The index of the subpage on the main page.
6737 mdb_node_shrink(MDB_page *mp, indx_t indx)
6743 indx_t i, numkeys, ptr;
6745 node = NODEPTR(mp, indx);
6746 sp = (MDB_page *)NODEDATA(node);
6747 delta = SIZELEFT(sp);
6748 xp = (MDB_page *)((char *)sp + delta);
6750 /* shift subpage upward */
6752 nsize = NUMKEYS(sp) * sp->mp_pad;
6754 return; /* do not make the node uneven-sized */
6755 memmove(METADATA(xp), METADATA(sp), nsize);
6758 numkeys = NUMKEYS(sp);
6759 for (i=numkeys-1; i>=0; i--)
6760 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6762 xp->mp_upper = sp->mp_lower;
6763 xp->mp_lower = sp->mp_lower;
6764 xp->mp_flags = sp->mp_flags;
6765 xp->mp_pad = sp->mp_pad;
6766 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6768 nsize = NODEDSZ(node) - delta;
6769 SETDSZ(node, nsize);
6771 /* shift lower nodes upward */
6772 ptr = mp->mp_ptrs[indx];
6773 numkeys = NUMKEYS(mp);
6774 for (i = 0; i < numkeys; i++) {
6775 if (mp->mp_ptrs[i] <= ptr)
6776 mp->mp_ptrs[i] += delta;
6779 base = (char *)mp + mp->mp_upper + PAGEBASE;
6780 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6781 mp->mp_upper += delta;
6784 /** Initial setup of a sorted-dups cursor.
6785 * Sorted duplicates are implemented as a sub-database for the given key.
6786 * The duplicate data items are actually keys of the sub-database.
6787 * Operations on the duplicate data items are performed using a sub-cursor
6788 * initialized when the sub-database is first accessed. This function does
6789 * the preliminary setup of the sub-cursor, filling in the fields that
6790 * depend only on the parent DB.
6791 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6794 mdb_xcursor_init0(MDB_cursor *mc)
6796 MDB_xcursor *mx = mc->mc_xcursor;
6798 mx->mx_cursor.mc_xcursor = NULL;
6799 mx->mx_cursor.mc_txn = mc->mc_txn;
6800 mx->mx_cursor.mc_db = &mx->mx_db;
6801 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6802 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6803 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6804 mx->mx_cursor.mc_snum = 0;
6805 mx->mx_cursor.mc_top = 0;
6806 mx->mx_cursor.mc_flags = C_SUB;
6807 mx->mx_dbx.md_name.mv_size = 0;
6808 mx->mx_dbx.md_name.mv_data = NULL;
6809 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6810 mx->mx_dbx.md_dcmp = NULL;
6811 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6814 /** Final setup of a sorted-dups cursor.
6815 * Sets up the fields that depend on the data from the main cursor.
6816 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6817 * @param[in] node The data containing the #MDB_db record for the
6818 * sorted-dup database.
6821 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6823 MDB_xcursor *mx = mc->mc_xcursor;
6825 if (node->mn_flags & F_SUBDATA) {
6826 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6827 mx->mx_cursor.mc_pg[0] = 0;
6828 mx->mx_cursor.mc_snum = 0;
6829 mx->mx_cursor.mc_top = 0;
6830 mx->mx_cursor.mc_flags = C_SUB;
6832 MDB_page *fp = NODEDATA(node);
6833 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6834 mx->mx_db.md_flags = 0;
6835 mx->mx_db.md_depth = 1;
6836 mx->mx_db.md_branch_pages = 0;
6837 mx->mx_db.md_leaf_pages = 1;
6838 mx->mx_db.md_overflow_pages = 0;
6839 mx->mx_db.md_entries = NUMKEYS(fp);
6840 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6841 mx->mx_cursor.mc_snum = 1;
6842 mx->mx_cursor.mc_top = 0;
6843 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6844 mx->mx_cursor.mc_pg[0] = fp;
6845 mx->mx_cursor.mc_ki[0] = 0;
6846 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6847 mx->mx_db.md_flags = MDB_DUPFIXED;
6848 mx->mx_db.md_pad = fp->mp_pad;
6849 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6850 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6853 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6854 mx->mx_db.md_root));
6855 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6856 #if UINT_MAX < SIZE_MAX
6857 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6858 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6862 /** Initialize a cursor for a given transaction and database. */
6864 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6867 mc->mc_backup = NULL;
6870 mc->mc_db = &txn->mt_dbs[dbi];
6871 mc->mc_dbx = &txn->mt_dbxs[dbi];
6872 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6877 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6878 mdb_tassert(txn, mx != NULL);
6879 mc->mc_xcursor = mx;
6880 mdb_xcursor_init0(mc);
6882 mc->mc_xcursor = NULL;
6884 if (*mc->mc_dbflag & DB_STALE) {
6885 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6890 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6893 size_t size = sizeof(MDB_cursor);
6895 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6898 if (txn->mt_flags & MDB_TXN_ERROR)
6901 /* Allow read access to the freelist */
6902 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6905 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6906 size += sizeof(MDB_xcursor);
6908 if ((mc = malloc(size)) != NULL) {
6909 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6910 if (txn->mt_cursors) {
6911 mc->mc_next = txn->mt_cursors[dbi];
6912 txn->mt_cursors[dbi] = mc;
6913 mc->mc_flags |= C_UNTRACK;
6925 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6927 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
6930 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6933 if (txn->mt_flags & MDB_TXN_ERROR)
6936 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6940 /* Return the count of duplicate data items for the current key */
6942 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6946 if (mc == NULL || countp == NULL)
6949 if (mc->mc_xcursor == NULL)
6950 return MDB_INCOMPATIBLE;
6952 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6955 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6956 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6959 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6962 *countp = mc->mc_xcursor->mx_db.md_entries;
6968 mdb_cursor_close(MDB_cursor *mc)
6970 if (mc && !mc->mc_backup) {
6971 /* remove from txn, if tracked */
6972 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6973 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6974 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6976 *prev = mc->mc_next;
6983 mdb_cursor_txn(MDB_cursor *mc)
6985 if (!mc) return NULL;
6990 mdb_cursor_dbi(MDB_cursor *mc)
6995 /** Replace the key for a branch node with a new key.
6996 * @param[in] mc Cursor pointing to the node to operate on.
6997 * @param[in] key The new key to use.
6998 * @return 0 on success, non-zero on failure.
7001 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7007 int delta, ksize, oksize;
7008 indx_t ptr, i, numkeys, indx;
7011 indx = mc->mc_ki[mc->mc_top];
7012 mp = mc->mc_pg[mc->mc_top];
7013 node = NODEPTR(mp, indx);
7014 ptr = mp->mp_ptrs[indx];
7018 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7019 k2.mv_data = NODEKEY(node);
7020 k2.mv_size = node->mn_ksize;
7021 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7023 mdb_dkey(&k2, kbuf2),
7029 /* Sizes must be 2-byte aligned. */
7030 ksize = EVEN(key->mv_size);
7031 oksize = EVEN(node->mn_ksize);
7032 delta = ksize - oksize;
7034 /* Shift node contents if EVEN(key length) changed. */
7036 if (delta > 0 && SIZELEFT(mp) < delta) {
7038 /* not enough space left, do a delete and split */
7039 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7040 pgno = NODEPGNO(node);
7041 mdb_node_del(mc, 0);
7042 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7045 numkeys = NUMKEYS(mp);
7046 for (i = 0; i < numkeys; i++) {
7047 if (mp->mp_ptrs[i] <= ptr)
7048 mp->mp_ptrs[i] -= delta;
7051 base = (char *)mp + mp->mp_upper + PAGEBASE;
7052 len = ptr - mp->mp_upper + NODESIZE;
7053 memmove(base - delta, base, len);
7054 mp->mp_upper -= delta;
7056 node = NODEPTR(mp, indx);
7059 /* But even if no shift was needed, update ksize */
7060 if (node->mn_ksize != key->mv_size)
7061 node->mn_ksize = key->mv_size;
7064 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7070 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7072 /** Move a node from csrc to cdst.
7075 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7082 unsigned short flags;
7086 /* Mark src and dst as dirty. */
7087 if ((rc = mdb_page_touch(csrc)) ||
7088 (rc = mdb_page_touch(cdst)))
7091 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7092 key.mv_size = csrc->mc_db->md_pad;
7093 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7095 data.mv_data = NULL;
7099 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7100 mdb_cassert(csrc, !((size_t)srcnode & 1));
7101 srcpg = NODEPGNO(srcnode);
7102 flags = srcnode->mn_flags;
7103 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7104 unsigned int snum = csrc->mc_snum;
7106 /* must find the lowest key below src */
7107 rc = mdb_page_search_lowest(csrc);
7110 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7111 key.mv_size = csrc->mc_db->md_pad;
7112 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7114 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7115 key.mv_size = NODEKSZ(s2);
7116 key.mv_data = NODEKEY(s2);
7118 csrc->mc_snum = snum--;
7119 csrc->mc_top = snum;
7121 key.mv_size = NODEKSZ(srcnode);
7122 key.mv_data = NODEKEY(srcnode);
7124 data.mv_size = NODEDSZ(srcnode);
7125 data.mv_data = NODEDATA(srcnode);
7127 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7128 unsigned int snum = cdst->mc_snum;
7131 /* must find the lowest key below dst */
7132 mdb_cursor_copy(cdst, &mn);
7133 rc = mdb_page_search_lowest(&mn);
7136 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7137 bkey.mv_size = mn.mc_db->md_pad;
7138 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7140 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7141 bkey.mv_size = NODEKSZ(s2);
7142 bkey.mv_data = NODEKEY(s2);
7144 mn.mc_snum = snum--;
7147 rc = mdb_update_key(&mn, &bkey);
7152 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7153 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7154 csrc->mc_ki[csrc->mc_top],
7156 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7157 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7159 /* Add the node to the destination page.
7161 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7162 if (rc != MDB_SUCCESS)
7165 /* Delete the node from the source page.
7167 mdb_node_del(csrc, key.mv_size);
7170 /* Adjust other cursors pointing to mp */
7171 MDB_cursor *m2, *m3;
7172 MDB_dbi dbi = csrc->mc_dbi;
7173 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7175 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7176 if (csrc->mc_flags & C_SUB)
7177 m3 = &m2->mc_xcursor->mx_cursor;
7180 if (m3 == csrc) continue;
7181 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7182 csrc->mc_ki[csrc->mc_top]) {
7183 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7184 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7189 /* Update the parent separators.
7191 if (csrc->mc_ki[csrc->mc_top] == 0) {
7192 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7193 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7194 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7196 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7197 key.mv_size = NODEKSZ(srcnode);
7198 key.mv_data = NODEKEY(srcnode);
7200 DPRINTF(("update separator for source page %"Z"u to [%s]",
7201 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7202 mdb_cursor_copy(csrc, &mn);
7205 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7208 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7210 indx_t ix = csrc->mc_ki[csrc->mc_top];
7211 nullkey.mv_size = 0;
7212 csrc->mc_ki[csrc->mc_top] = 0;
7213 rc = mdb_update_key(csrc, &nullkey);
7214 csrc->mc_ki[csrc->mc_top] = ix;
7215 mdb_cassert(csrc, rc == MDB_SUCCESS);
7219 if (cdst->mc_ki[cdst->mc_top] == 0) {
7220 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7221 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7222 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7224 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7225 key.mv_size = NODEKSZ(srcnode);
7226 key.mv_data = NODEKEY(srcnode);
7228 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7229 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7230 mdb_cursor_copy(cdst, &mn);
7233 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7236 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7238 indx_t ix = cdst->mc_ki[cdst->mc_top];
7239 nullkey.mv_size = 0;
7240 cdst->mc_ki[cdst->mc_top] = 0;
7241 rc = mdb_update_key(cdst, &nullkey);
7242 cdst->mc_ki[cdst->mc_top] = ix;
7243 mdb_cassert(csrc, rc == MDB_SUCCESS);
7250 /** Merge one page into another.
7251 * The nodes from the page pointed to by \b csrc will
7252 * be copied to the page pointed to by \b cdst and then
7253 * the \b csrc page will be freed.
7254 * @param[in] csrc Cursor pointing to the source page.
7255 * @param[in] cdst Cursor pointing to the destination page.
7256 * @return 0 on success, non-zero on failure.
7259 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7261 MDB_page *psrc, *pdst;
7268 psrc = csrc->mc_pg[csrc->mc_top];
7269 pdst = cdst->mc_pg[cdst->mc_top];
7271 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7273 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7274 mdb_cassert(csrc, cdst->mc_snum > 1);
7276 /* Mark dst as dirty. */
7277 if ((rc = mdb_page_touch(cdst)))
7280 /* Move all nodes from src to dst.
7282 j = nkeys = NUMKEYS(pdst);
7283 if (IS_LEAF2(psrc)) {
7284 key.mv_size = csrc->mc_db->md_pad;
7285 key.mv_data = METADATA(psrc);
7286 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7287 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7288 if (rc != MDB_SUCCESS)
7290 key.mv_data = (char *)key.mv_data + key.mv_size;
7293 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7294 srcnode = NODEPTR(psrc, i);
7295 if (i == 0 && IS_BRANCH(psrc)) {
7298 mdb_cursor_copy(csrc, &mn);
7299 /* must find the lowest key below src */
7300 rc = mdb_page_search_lowest(&mn);
7303 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7304 key.mv_size = mn.mc_db->md_pad;
7305 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7307 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7308 key.mv_size = NODEKSZ(s2);
7309 key.mv_data = NODEKEY(s2);
7312 key.mv_size = srcnode->mn_ksize;
7313 key.mv_data = NODEKEY(srcnode);
7316 data.mv_size = NODEDSZ(srcnode);
7317 data.mv_data = NODEDATA(srcnode);
7318 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7319 if (rc != MDB_SUCCESS)
7324 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7325 pdst->mp_pgno, NUMKEYS(pdst),
7326 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7328 /* Unlink the src page from parent and add to free list.
7331 mdb_node_del(csrc, 0);
7332 if (csrc->mc_ki[csrc->mc_top] == 0) {
7334 rc = mdb_update_key(csrc, &key);
7342 psrc = csrc->mc_pg[csrc->mc_top];
7343 /* If not operating on FreeDB, allow this page to be reused
7344 * in this txn. Otherwise just add to free list.
7346 rc = mdb_page_loose(csrc, psrc);
7350 csrc->mc_db->md_leaf_pages--;
7352 csrc->mc_db->md_branch_pages--;
7354 /* Adjust other cursors pointing to mp */
7355 MDB_cursor *m2, *m3;
7356 MDB_dbi dbi = csrc->mc_dbi;
7358 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7359 if (csrc->mc_flags & C_SUB)
7360 m3 = &m2->mc_xcursor->mx_cursor;
7363 if (m3 == csrc) continue;
7364 if (m3->mc_snum < csrc->mc_snum) continue;
7365 if (m3->mc_pg[csrc->mc_top] == psrc) {
7366 m3->mc_pg[csrc->mc_top] = pdst;
7367 m3->mc_ki[csrc->mc_top] += nkeys;
7372 unsigned int snum = cdst->mc_snum;
7373 uint16_t depth = cdst->mc_db->md_depth;
7374 mdb_cursor_pop(cdst);
7375 rc = mdb_rebalance(cdst);
7376 /* Did the tree shrink? */
7377 if (depth > cdst->mc_db->md_depth)
7379 cdst->mc_snum = snum;
7380 cdst->mc_top = snum-1;
7385 /** Copy the contents of a cursor.
7386 * @param[in] csrc The cursor to copy from.
7387 * @param[out] cdst The cursor to copy to.
7390 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7394 cdst->mc_txn = csrc->mc_txn;
7395 cdst->mc_dbi = csrc->mc_dbi;
7396 cdst->mc_db = csrc->mc_db;
7397 cdst->mc_dbx = csrc->mc_dbx;
7398 cdst->mc_snum = csrc->mc_snum;
7399 cdst->mc_top = csrc->mc_top;
7400 cdst->mc_flags = csrc->mc_flags;
7402 for (i=0; i<csrc->mc_snum; i++) {
7403 cdst->mc_pg[i] = csrc->mc_pg[i];
7404 cdst->mc_ki[i] = csrc->mc_ki[i];
7408 /** Rebalance the tree after a delete operation.
7409 * @param[in] mc Cursor pointing to the page where rebalancing
7411 * @return 0 on success, non-zero on failure.
7414 mdb_rebalance(MDB_cursor *mc)
7418 unsigned int ptop, minkeys;
7422 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7423 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7424 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7425 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7426 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7428 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7429 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7430 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7431 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7435 if (mc->mc_snum < 2) {
7436 MDB_page *mp = mc->mc_pg[0];
7438 DPUTS("Can't rebalance a subpage, ignoring");
7441 if (NUMKEYS(mp) == 0) {
7442 DPUTS("tree is completely empty");
7443 mc->mc_db->md_root = P_INVALID;
7444 mc->mc_db->md_depth = 0;
7445 mc->mc_db->md_leaf_pages = 0;
7446 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7449 /* Adjust cursors pointing to mp */
7452 mc->mc_flags &= ~C_INITIALIZED;
7454 MDB_cursor *m2, *m3;
7455 MDB_dbi dbi = mc->mc_dbi;
7457 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7458 if (mc->mc_flags & C_SUB)
7459 m3 = &m2->mc_xcursor->mx_cursor;
7462 if (m3->mc_snum < mc->mc_snum) continue;
7463 if (m3->mc_pg[0] == mp) {
7466 m3->mc_flags &= ~C_INITIALIZED;
7470 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7472 DPUTS("collapsing root page!");
7473 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7476 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7477 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7480 mc->mc_db->md_depth--;
7481 mc->mc_db->md_branch_pages--;
7482 mc->mc_ki[0] = mc->mc_ki[1];
7483 for (i = 1; i<mc->mc_db->md_depth; i++) {
7484 mc->mc_pg[i] = mc->mc_pg[i+1];
7485 mc->mc_ki[i] = mc->mc_ki[i+1];
7488 /* Adjust other cursors pointing to mp */
7489 MDB_cursor *m2, *m3;
7490 MDB_dbi dbi = mc->mc_dbi;
7492 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7493 if (mc->mc_flags & C_SUB)
7494 m3 = &m2->mc_xcursor->mx_cursor;
7497 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7498 if (m3->mc_pg[0] == mp) {
7501 for (i=0; i<m3->mc_snum; i++) {
7502 m3->mc_pg[i] = m3->mc_pg[i+1];
7503 m3->mc_ki[i] = m3->mc_ki[i+1];
7509 DPUTS("root page doesn't need rebalancing");
7513 /* The parent (branch page) must have at least 2 pointers,
7514 * otherwise the tree is invalid.
7516 ptop = mc->mc_top-1;
7517 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7519 /* Leaf page fill factor is below the threshold.
7520 * Try to move keys from left or right neighbor, or
7521 * merge with a neighbor page.
7526 mdb_cursor_copy(mc, &mn);
7527 mn.mc_xcursor = NULL;
7529 oldki = mc->mc_ki[mc->mc_top];
7530 if (mc->mc_ki[ptop] == 0) {
7531 /* We're the leftmost leaf in our parent.
7533 DPUTS("reading right neighbor");
7535 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7536 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7539 mn.mc_ki[mn.mc_top] = 0;
7540 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7542 /* There is at least one neighbor to the left.
7544 DPUTS("reading left neighbor");
7546 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7547 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7550 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7551 mc->mc_ki[mc->mc_top] = 0;
7554 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7555 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7556 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7558 /* If the neighbor page is above threshold and has enough keys,
7559 * move one key from it. Otherwise we should try to merge them.
7560 * (A branch page must never have less than 2 keys.)
7562 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7563 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7564 rc = mdb_node_move(&mn, mc);
7565 if (mc->mc_ki[ptop]) {
7569 if (mc->mc_ki[ptop] == 0) {
7570 rc = mdb_page_merge(&mn, mc);
7572 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7573 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7574 rc = mdb_page_merge(mc, &mn);
7575 mdb_cursor_copy(&mn, mc);
7577 mc->mc_flags &= ~C_EOF;
7579 mc->mc_ki[mc->mc_top] = oldki;
7583 /** Complete a delete operation started by #mdb_cursor_del(). */
7585 mdb_cursor_del0(MDB_cursor *mc)
7592 ki = mc->mc_ki[mc->mc_top];
7593 mdb_node_del(mc, mc->mc_db->md_pad);
7594 mc->mc_db->md_entries--;
7595 rc = mdb_rebalance(mc);
7597 if (rc == MDB_SUCCESS) {
7598 MDB_cursor *m2, *m3;
7599 MDB_dbi dbi = mc->mc_dbi;
7601 mp = mc->mc_pg[mc->mc_top];
7602 nkeys = NUMKEYS(mp);
7604 /* if mc points past last node in page, find next sibling */
7605 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7606 rc = mdb_cursor_sibling(mc, 1);
7607 if (rc == MDB_NOTFOUND) {
7608 mc->mc_flags |= C_EOF;
7613 /* Adjust other cursors pointing to mp */
7614 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7615 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7616 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7618 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7620 if (m3->mc_pg[mc->mc_top] == mp) {
7621 if (m3->mc_ki[mc->mc_top] >= ki) {
7622 m3->mc_flags |= C_DEL;
7623 if (m3->mc_ki[mc->mc_top] > ki)
7624 m3->mc_ki[mc->mc_top]--;
7626 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7627 rc = mdb_cursor_sibling(m3, 1);
7628 if (rc == MDB_NOTFOUND) {
7629 m3->mc_flags |= C_EOF;
7635 mc->mc_flags |= C_DEL;
7639 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7644 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7645 MDB_val *key, MDB_val *data)
7647 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7650 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7651 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7653 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7654 /* must ignore any data */
7658 return mdb_del0(txn, dbi, key, data, 0);
7662 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7663 MDB_val *key, MDB_val *data, unsigned flags)
7668 MDB_val rdata, *xdata;
7672 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7674 mdb_cursor_init(&mc, txn, dbi, &mx);
7683 flags |= MDB_NODUPDATA;
7685 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7687 /* let mdb_page_split know about this cursor if needed:
7688 * delete will trigger a rebalance; if it needs to move
7689 * a node from one page to another, it will have to
7690 * update the parent's separator key(s). If the new sepkey
7691 * is larger than the current one, the parent page may
7692 * run out of space, triggering a split. We need this
7693 * cursor to be consistent until the end of the rebalance.
7695 mc.mc_flags |= C_UNTRACK;
7696 mc.mc_next = txn->mt_cursors[dbi];
7697 txn->mt_cursors[dbi] = &mc;
7698 rc = mdb_cursor_del(&mc, flags);
7699 txn->mt_cursors[dbi] = mc.mc_next;
7704 /** Split a page and insert a new node.
7705 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7706 * The cursor will be updated to point to the actual page and index where
7707 * the node got inserted after the split.
7708 * @param[in] newkey The key for the newly inserted node.
7709 * @param[in] newdata The data for the newly inserted node.
7710 * @param[in] newpgno The page number, if the new node is a branch node.
7711 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7712 * @return 0 on success, non-zero on failure.
7715 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7716 unsigned int nflags)
7719 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7722 int i, j, split_indx, nkeys, pmax;
7723 MDB_env *env = mc->mc_txn->mt_env;
7725 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7726 MDB_page *copy = NULL;
7727 MDB_page *mp, *rp, *pp;
7732 mp = mc->mc_pg[mc->mc_top];
7733 newindx = mc->mc_ki[mc->mc_top];
7734 nkeys = NUMKEYS(mp);
7736 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7737 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7738 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7740 /* Create a right sibling. */
7741 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7743 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7745 if (mc->mc_snum < 2) {
7746 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7748 /* shift current top to make room for new parent */
7749 mc->mc_pg[1] = mc->mc_pg[0];
7750 mc->mc_ki[1] = mc->mc_ki[0];
7753 mc->mc_db->md_root = pp->mp_pgno;
7754 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7755 mc->mc_db->md_depth++;
7758 /* Add left (implicit) pointer. */
7759 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7760 /* undo the pre-push */
7761 mc->mc_pg[0] = mc->mc_pg[1];
7762 mc->mc_ki[0] = mc->mc_ki[1];
7763 mc->mc_db->md_root = mp->mp_pgno;
7764 mc->mc_db->md_depth--;
7771 ptop = mc->mc_top-1;
7772 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7775 mc->mc_flags |= C_SPLITTING;
7776 mdb_cursor_copy(mc, &mn);
7777 mn.mc_pg[mn.mc_top] = rp;
7778 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7780 if (nflags & MDB_APPEND) {
7781 mn.mc_ki[mn.mc_top] = 0;
7783 split_indx = newindx;
7787 split_indx = (nkeys+1) / 2;
7792 unsigned int lsize, rsize, ksize;
7793 /* Move half of the keys to the right sibling */
7794 x = mc->mc_ki[mc->mc_top] - split_indx;
7795 ksize = mc->mc_db->md_pad;
7796 split = LEAF2KEY(mp, split_indx, ksize);
7797 rsize = (nkeys - split_indx) * ksize;
7798 lsize = (nkeys - split_indx) * sizeof(indx_t);
7799 mp->mp_lower -= lsize;
7800 rp->mp_lower += lsize;
7801 mp->mp_upper += rsize - lsize;
7802 rp->mp_upper -= rsize - lsize;
7803 sepkey.mv_size = ksize;
7804 if (newindx == split_indx) {
7805 sepkey.mv_data = newkey->mv_data;
7807 sepkey.mv_data = split;
7810 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7811 memcpy(rp->mp_ptrs, split, rsize);
7812 sepkey.mv_data = rp->mp_ptrs;
7813 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7814 memcpy(ins, newkey->mv_data, ksize);
7815 mp->mp_lower += sizeof(indx_t);
7816 mp->mp_upper -= ksize - sizeof(indx_t);
7819 memcpy(rp->mp_ptrs, split, x * ksize);
7820 ins = LEAF2KEY(rp, x, ksize);
7821 memcpy(ins, newkey->mv_data, ksize);
7822 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7823 rp->mp_lower += sizeof(indx_t);
7824 rp->mp_upper -= ksize - sizeof(indx_t);
7825 mc->mc_ki[mc->mc_top] = x;
7826 mc->mc_pg[mc->mc_top] = rp;
7829 int psize, nsize, k;
7830 /* Maximum free space in an empty page */
7831 pmax = env->me_psize - PAGEHDRSZ;
7833 nsize = mdb_leaf_size(env, newkey, newdata);
7835 nsize = mdb_branch_size(env, newkey);
7836 nsize = EVEN(nsize);
7838 /* grab a page to hold a temporary copy */
7839 copy = mdb_page_malloc(mc->mc_txn, 1);
7844 copy->mp_pgno = mp->mp_pgno;
7845 copy->mp_flags = mp->mp_flags;
7846 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
7847 copy->mp_upper = env->me_psize - PAGEBASE;
7849 /* prepare to insert */
7850 for (i=0, j=0; i<nkeys; i++) {
7852 copy->mp_ptrs[j++] = 0;
7854 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7857 /* When items are relatively large the split point needs
7858 * to be checked, because being off-by-one will make the
7859 * difference between success or failure in mdb_node_add.
7861 * It's also relevant if a page happens to be laid out
7862 * such that one half of its nodes are all "small" and
7863 * the other half of its nodes are "large." If the new
7864 * item is also "large" and falls on the half with
7865 * "large" nodes, it also may not fit.
7867 * As a final tweak, if the new item goes on the last
7868 * spot on the page (and thus, onto the new page), bias
7869 * the split so the new page is emptier than the old page.
7870 * This yields better packing during sequential inserts.
7872 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7873 /* Find split point */
7875 if (newindx <= split_indx || newindx >= nkeys) {
7877 k = newindx >= nkeys ? nkeys : split_indx+2;
7882 for (; i!=k; i+=j) {
7887 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
7888 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7890 if (F_ISSET(node->mn_flags, F_BIGDATA))
7891 psize += sizeof(pgno_t);
7893 psize += NODEDSZ(node);
7895 psize = EVEN(psize);
7897 if (psize > pmax || i == k-j) {
7898 split_indx = i + (j<0);
7903 if (split_indx == newindx) {
7904 sepkey.mv_size = newkey->mv_size;
7905 sepkey.mv_data = newkey->mv_data;
7907 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
7908 sepkey.mv_size = node->mn_ksize;
7909 sepkey.mv_data = NODEKEY(node);
7914 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7916 /* Copy separator key to the parent.
7918 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7922 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7927 if (mn.mc_snum == mc->mc_snum) {
7928 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7929 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7930 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7931 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7936 /* Right page might now have changed parent.
7937 * Check if left page also changed parent.
7939 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7940 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7941 for (i=0; i<ptop; i++) {
7942 mc->mc_pg[i] = mn.mc_pg[i];
7943 mc->mc_ki[i] = mn.mc_ki[i];
7945 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7946 if (mn.mc_ki[ptop]) {
7947 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7949 /* find right page's left sibling */
7950 mc->mc_ki[ptop] = mn.mc_ki[ptop];
7951 mdb_cursor_sibling(mc, 0);
7956 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7959 mc->mc_flags ^= C_SPLITTING;
7960 if (rc != MDB_SUCCESS) {
7963 if (nflags & MDB_APPEND) {
7964 mc->mc_pg[mc->mc_top] = rp;
7965 mc->mc_ki[mc->mc_top] = 0;
7966 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7969 for (i=0; i<mc->mc_top; i++)
7970 mc->mc_ki[i] = mn.mc_ki[i];
7971 } else if (!IS_LEAF2(mp)) {
7973 mc->mc_pg[mc->mc_top] = rp;
7978 rkey.mv_data = newkey->mv_data;
7979 rkey.mv_size = newkey->mv_size;
7985 /* Update index for the new key. */
7986 mc->mc_ki[mc->mc_top] = j;
7988 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
7989 rkey.mv_data = NODEKEY(node);
7990 rkey.mv_size = node->mn_ksize;
7992 xdata.mv_data = NODEDATA(node);
7993 xdata.mv_size = NODEDSZ(node);
7996 pgno = NODEPGNO(node);
7997 flags = node->mn_flags;
8000 if (!IS_LEAF(mp) && j == 0) {
8001 /* First branch index doesn't need key data. */
8005 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8011 mc->mc_pg[mc->mc_top] = copy;
8016 } while (i != split_indx);
8018 nkeys = NUMKEYS(copy);
8019 for (i=0; i<nkeys; i++)
8020 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8021 mp->mp_lower = copy->mp_lower;
8022 mp->mp_upper = copy->mp_upper;
8023 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8024 env->me_psize - copy->mp_upper - PAGEBASE);
8026 /* reset back to original page */
8027 if (newindx < split_indx) {
8028 mc->mc_pg[mc->mc_top] = mp;
8029 if (nflags & MDB_RESERVE) {
8030 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8031 if (!(node->mn_flags & F_BIGDATA))
8032 newdata->mv_data = NODEDATA(node);
8035 mc->mc_pg[mc->mc_top] = rp;
8037 /* Make sure mc_ki is still valid.
8039 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8040 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8041 for (i=0; i<=ptop; i++) {
8042 mc->mc_pg[i] = mn.mc_pg[i];
8043 mc->mc_ki[i] = mn.mc_ki[i];
8050 /* Adjust other cursors pointing to mp */
8051 MDB_cursor *m2, *m3;
8052 MDB_dbi dbi = mc->mc_dbi;
8053 int fixup = NUMKEYS(mp);
8055 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8056 if (mc->mc_flags & C_SUB)
8057 m3 = &m2->mc_xcursor->mx_cursor;
8062 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8064 if (m3->mc_flags & C_SPLITTING)
8069 for (k=m3->mc_top; k>=0; k--) {
8070 m3->mc_ki[k+1] = m3->mc_ki[k];
8071 m3->mc_pg[k+1] = m3->mc_pg[k];
8073 if (m3->mc_ki[0] >= split_indx) {
8078 m3->mc_pg[0] = mc->mc_pg[0];
8082 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8083 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8084 m3->mc_ki[mc->mc_top]++;
8085 if (m3->mc_ki[mc->mc_top] >= fixup) {
8086 m3->mc_pg[mc->mc_top] = rp;
8087 m3->mc_ki[mc->mc_top] -= fixup;
8088 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8090 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8091 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8096 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8099 if (copy) /* tmp page */
8100 mdb_page_free(env, copy);
8102 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8107 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8108 MDB_val *key, MDB_val *data, unsigned int flags)
8113 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8116 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8119 mdb_cursor_init(&mc, txn, dbi, &mx);
8120 return mdb_cursor_put(&mc, key, data, flags);
8124 #define MDB_WBUF (1024*1024)
8127 /** State needed for a compacting copy. */
8128 typedef struct mdb_copy {
8129 pthread_mutex_t mc_mutex;
8130 pthread_cond_t mc_cond;
8137 pgno_t mc_next_pgno;
8140 volatile int mc_new;
8145 /** Dedicated writer thread for compacting copy. */
8146 static THREAD_RET ESECT
8147 mdb_env_copythr(void *arg)
8151 int toggle = 0, wsize, rc;
8154 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8157 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8160 pthread_mutex_lock(&my->mc_mutex);
8162 pthread_cond_signal(&my->mc_cond);
8165 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8166 if (my->mc_new < 0) {
8171 wsize = my->mc_wlen[toggle];
8172 ptr = my->mc_wbuf[toggle];
8175 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8179 } else if (len > 0) {
8193 /* If there's an overflow page tail, write it too */
8194 if (my->mc_olen[toggle]) {
8195 wsize = my->mc_olen[toggle];
8196 ptr = my->mc_over[toggle];
8197 my->mc_olen[toggle] = 0;
8200 my->mc_wlen[toggle] = 0;
8202 pthread_cond_signal(&my->mc_cond);
8204 pthread_cond_signal(&my->mc_cond);
8205 pthread_mutex_unlock(&my->mc_mutex);
8206 return (THREAD_RET)0;
8210 /** Tell the writer thread there's a buffer ready to write */
8212 mdb_env_cthr_toggle(mdb_copy *my, int st)
8214 int toggle = my->mc_toggle ^ 1;
8215 pthread_mutex_lock(&my->mc_mutex);
8216 if (my->mc_status) {
8217 pthread_mutex_unlock(&my->mc_mutex);
8218 return my->mc_status;
8220 while (my->mc_new == 1)
8221 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8223 my->mc_toggle = toggle;
8224 pthread_cond_signal(&my->mc_cond);
8225 pthread_mutex_unlock(&my->mc_mutex);
8229 /** Depth-first tree traversal for compacting copy. */
8231 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8234 MDB_txn *txn = my->mc_txn;
8236 MDB_page *mo, *mp, *leaf;
8241 /* Empty DB, nothing to do */
8242 if (*pg == P_INVALID)
8249 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8252 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8256 /* Make cursor pages writable */
8257 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8261 for (i=0; i<mc.mc_top; i++) {
8262 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8263 mc.mc_pg[i] = (MDB_page *)ptr;
8264 ptr += my->mc_env->me_psize;
8267 /* This is writable space for a leaf page. Usually not needed. */
8268 leaf = (MDB_page *)ptr;
8270 toggle = my->mc_toggle;
8271 while (mc.mc_snum > 0) {
8273 mp = mc.mc_pg[mc.mc_top];
8277 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8278 for (i=0; i<n; i++) {
8279 ni = NODEPTR(mp, i);
8280 if (ni->mn_flags & F_BIGDATA) {
8284 /* Need writable leaf */
8286 mc.mc_pg[mc.mc_top] = leaf;
8287 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8289 ni = NODEPTR(mp, i);
8292 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8293 rc = mdb_page_get(txn, pg, &omp, NULL);
8296 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8297 rc = mdb_env_cthr_toggle(my, 1);
8300 toggle = my->mc_toggle;
8302 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8303 memcpy(mo, omp, my->mc_env->me_psize);
8304 mo->mp_pgno = my->mc_next_pgno;
8305 my->mc_next_pgno += omp->mp_pages;
8306 my->mc_wlen[toggle] += my->mc_env->me_psize;
8307 if (omp->mp_pages > 1) {
8308 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8309 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8310 rc = mdb_env_cthr_toggle(my, 1);
8313 toggle = my->mc_toggle;
8315 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8316 } else if (ni->mn_flags & F_SUBDATA) {
8319 /* Need writable leaf */
8321 mc.mc_pg[mc.mc_top] = leaf;
8322 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8324 ni = NODEPTR(mp, i);
8327 memcpy(&db, NODEDATA(ni), sizeof(db));
8328 my->mc_toggle = toggle;
8329 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8332 toggle = my->mc_toggle;
8333 memcpy(NODEDATA(ni), &db, sizeof(db));
8338 mc.mc_ki[mc.mc_top]++;
8339 if (mc.mc_ki[mc.mc_top] < n) {
8342 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8344 rc = mdb_page_get(txn, pg, &mp, NULL);
8349 mc.mc_ki[mc.mc_top] = 0;
8350 if (IS_BRANCH(mp)) {
8351 /* Whenever we advance to a sibling branch page,
8352 * we must proceed all the way down to its first leaf.
8354 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8357 mc.mc_pg[mc.mc_top] = mp;
8361 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8362 rc = mdb_env_cthr_toggle(my, 1);
8365 toggle = my->mc_toggle;
8367 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8368 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8369 mo->mp_pgno = my->mc_next_pgno++;
8370 my->mc_wlen[toggle] += my->mc_env->me_psize;
8372 /* Update parent if there is one */
8373 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8374 SETPGNO(ni, mo->mp_pgno);
8375 mdb_cursor_pop(&mc);
8377 /* Otherwise we're done */
8387 /** Copy environment with compaction. */
8389 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8394 MDB_txn *txn = NULL;
8399 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8400 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8401 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_psize);
8402 if (my.mc_wbuf[0] == NULL)
8405 pthread_mutex_init(&my.mc_mutex, NULL);
8406 pthread_cond_init(&my.mc_cond, NULL);
8407 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_psize, MDB_WBUF*2);
8411 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8416 my.mc_next_pgno = 2;
8422 THREAD_CREATE(thr, mdb_env_copythr, &my);
8424 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8428 mp = (MDB_page *)my.mc_wbuf[0];
8429 memset(mp, 0, 2*env->me_psize);
8431 mp->mp_flags = P_META;
8432 mm = (MDB_meta *)METADATA(mp);
8433 mdb_env_init_meta0(env, mm);
8434 mm->mm_address = env->me_metas[0]->mm_address;
8436 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8438 mp->mp_flags = P_META;
8439 *(MDB_meta *)METADATA(mp) = *mm;
8440 mm = (MDB_meta *)METADATA(mp);
8442 /* Count the number of free pages, subtract from lastpg to find
8443 * number of active pages
8446 MDB_ID freecount = 0;
8449 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8450 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8451 freecount += *(MDB_ID *)data.mv_data;
8452 freecount += txn->mt_dbs[0].md_branch_pages +
8453 txn->mt_dbs[0].md_leaf_pages +
8454 txn->mt_dbs[0].md_overflow_pages;
8456 /* Set metapage 1 */
8457 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8458 mm->mm_dbs[1] = txn->mt_dbs[1];
8459 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8462 my.mc_wlen[0] = env->me_psize * 2;
8464 pthread_mutex_lock(&my.mc_mutex);
8466 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8467 pthread_mutex_unlock(&my.mc_mutex);
8468 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8469 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8470 rc = mdb_env_cthr_toggle(&my, 1);
8471 mdb_env_cthr_toggle(&my, -1);
8472 pthread_mutex_lock(&my.mc_mutex);
8474 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8475 pthread_mutex_unlock(&my.mc_mutex);
8480 CloseHandle(my.mc_cond);
8481 CloseHandle(my.mc_mutex);
8482 _aligned_free(my.mc_wbuf[0]);
8484 pthread_cond_destroy(&my.mc_cond);
8485 pthread_mutex_destroy(&my.mc_mutex);
8486 free(my.mc_wbuf[0]);
8491 /** Copy environment as-is. */
8493 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8495 MDB_txn *txn = NULL;
8501 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8505 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8508 /* Do the lock/unlock of the reader mutex before starting the
8509 * write txn. Otherwise other read txns could block writers.
8511 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8516 /* We must start the actual read txn after blocking writers */
8517 mdb_txn_reset0(txn, "reset-stage1");
8519 /* Temporarily block writers until we snapshot the meta pages */
8522 rc = mdb_txn_renew0(txn);
8524 UNLOCK_MUTEX_W(env);
8529 wsize = env->me_psize * 2;
8533 DO_WRITE(rc, fd, ptr, w2, len);
8537 } else if (len > 0) {
8543 /* Non-blocking or async handles are not supported */
8549 UNLOCK_MUTEX_W(env);
8554 w2 = txn->mt_next_pgno * env->me_psize;
8557 LARGE_INTEGER fsize;
8558 GetFileSizeEx(env->me_fd, &fsize);
8559 if (w2 > fsize.QuadPart)
8560 w2 = fsize.QuadPart;
8565 fstat(env->me_fd, &st);
8566 if (w2 > (size_t)st.st_size)
8572 if (wsize > MAX_WRITE)
8576 DO_WRITE(rc, fd, ptr, w2, len);
8580 } else if (len > 0) {
8597 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8599 if (flags & MDB_CP_COMPACT)
8600 return mdb_env_copyfd1(env, fd);
8602 return mdb_env_copyfd0(env, fd);
8606 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8608 return mdb_env_copyfd2(env, fd, 0);
8612 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8616 HANDLE newfd = INVALID_HANDLE_VALUE;
8618 if (env->me_flags & MDB_NOSUBDIR) {
8619 lpath = (char *)path;
8622 len += sizeof(DATANAME);
8623 lpath = malloc(len);
8626 sprintf(lpath, "%s" DATANAME, path);
8629 /* The destination path must exist, but the destination file must not.
8630 * We don't want the OS to cache the writes, since the source data is
8631 * already in the OS cache.
8634 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8635 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8637 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8639 if (newfd == INVALID_HANDLE_VALUE) {
8645 /* Set O_DIRECT if the file system supports it */
8646 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8647 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8649 #ifdef F_NOCACHE /* __APPLE__ */
8650 rc = fcntl(newfd, F_NOCACHE, 1);
8657 rc = mdb_env_copyfd2(env, newfd, flags);
8660 if (!(env->me_flags & MDB_NOSUBDIR))
8662 if (newfd != INVALID_HANDLE_VALUE)
8663 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8670 mdb_env_copy(MDB_env *env, const char *path)
8672 return mdb_env_copy2(env, path, 0);
8676 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8678 if ((flag & CHANGEABLE) != flag)
8681 env->me_flags |= flag;
8683 env->me_flags &= ~flag;
8688 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8693 *arg = env->me_flags;
8698 mdb_env_set_userctx(MDB_env *env, void *ctx)
8702 env->me_userctx = ctx;
8707 mdb_env_get_userctx(MDB_env *env)
8709 return env ? env->me_userctx : NULL;
8713 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8718 env->me_assert_func = func;
8724 mdb_env_get_path(MDB_env *env, const char **arg)
8729 *arg = env->me_path;
8734 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8743 /** Common code for #mdb_stat() and #mdb_env_stat().
8744 * @param[in] env the environment to operate in.
8745 * @param[in] db the #MDB_db record containing the stats to return.
8746 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8747 * @return 0, this function always succeeds.
8750 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8752 arg->ms_psize = env->me_psize;
8753 arg->ms_depth = db->md_depth;
8754 arg->ms_branch_pages = db->md_branch_pages;
8755 arg->ms_leaf_pages = db->md_leaf_pages;
8756 arg->ms_overflow_pages = db->md_overflow_pages;
8757 arg->ms_entries = db->md_entries;
8763 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8767 if (env == NULL || arg == NULL)
8770 toggle = mdb_env_pick_meta(env);
8772 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8776 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8780 if (env == NULL || arg == NULL)
8783 toggle = mdb_env_pick_meta(env);
8784 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8785 arg->me_mapsize = env->me_mapsize;
8786 arg->me_maxreaders = env->me_maxreaders;
8788 /* me_numreaders may be zero if this process never used any readers. Use
8789 * the shared numreader count if it exists.
8791 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8793 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8794 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8798 /** Set the default comparison functions for a database.
8799 * Called immediately after a database is opened to set the defaults.
8800 * The user can then override them with #mdb_set_compare() or
8801 * #mdb_set_dupsort().
8802 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8803 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8806 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8808 uint16_t f = txn->mt_dbs[dbi].md_flags;
8810 txn->mt_dbxs[dbi].md_cmp =
8811 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8812 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8814 txn->mt_dbxs[dbi].md_dcmp =
8815 !(f & MDB_DUPSORT) ? 0 :
8816 ((f & MDB_INTEGERDUP)
8817 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8818 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8821 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8826 int rc, dbflag, exact;
8827 unsigned int unused = 0, seq;
8830 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8831 mdb_default_cmp(txn, FREE_DBI);
8834 if ((flags & VALID_FLAGS) != flags)
8836 if (txn->mt_flags & MDB_TXN_ERROR)
8842 if (flags & PERSISTENT_FLAGS) {
8843 uint16_t f2 = flags & PERSISTENT_FLAGS;
8844 /* make sure flag changes get committed */
8845 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8846 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8847 txn->mt_flags |= MDB_TXN_DIRTY;
8850 mdb_default_cmp(txn, MAIN_DBI);
8854 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8855 mdb_default_cmp(txn, MAIN_DBI);
8858 /* Is the DB already open? */
8860 for (i=2; i<txn->mt_numdbs; i++) {
8861 if (!txn->mt_dbxs[i].md_name.mv_size) {
8862 /* Remember this free slot */
8863 if (!unused) unused = i;
8866 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8867 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8873 /* If no free slot and max hit, fail */
8874 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8875 return MDB_DBS_FULL;
8877 /* Cannot mix named databases with some mainDB flags */
8878 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8879 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8881 /* Find the DB info */
8882 dbflag = DB_NEW|DB_VALID;
8885 key.mv_data = (void *)name;
8886 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8887 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8888 if (rc == MDB_SUCCESS) {
8889 /* make sure this is actually a DB */
8890 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8891 if (!(node->mn_flags & F_SUBDATA))
8892 return MDB_INCOMPATIBLE;
8893 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8894 /* Create if requested */
8896 data.mv_size = sizeof(MDB_db);
8897 data.mv_data = &dummy;
8898 memset(&dummy, 0, sizeof(dummy));
8899 dummy.md_root = P_INVALID;
8900 dummy.md_flags = flags & PERSISTENT_FLAGS;
8901 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8905 /* OK, got info, add to table */
8906 if (rc == MDB_SUCCESS) {
8907 unsigned int slot = unused ? unused : txn->mt_numdbs;
8908 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8909 txn->mt_dbxs[slot].md_name.mv_size = len;
8910 txn->mt_dbxs[slot].md_rel = NULL;
8911 txn->mt_dbflags[slot] = dbflag;
8912 /* txn-> and env-> are the same in read txns, use
8913 * tmp variable to avoid undefined assignment
8915 seq = ++txn->mt_env->me_dbiseqs[slot];
8916 txn->mt_dbiseqs[slot] = seq;
8918 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8920 mdb_default_cmp(txn, slot);
8929 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8931 if (!arg || !TXN_DBI_EXIST(txn, dbi))
8934 if (txn->mt_flags & MDB_TXN_ERROR)
8937 if (txn->mt_dbflags[dbi] & DB_STALE) {
8940 /* Stale, must read the DB's root. cursor_init does it for us. */
8941 mdb_cursor_init(&mc, txn, dbi, &mx);
8943 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8946 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8949 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8951 ptr = env->me_dbxs[dbi].md_name.mv_data;
8952 /* If there was no name, this was already closed */
8954 env->me_dbxs[dbi].md_name.mv_data = NULL;
8955 env->me_dbxs[dbi].md_name.mv_size = 0;
8956 env->me_dbflags[dbi] = 0;
8957 env->me_dbiseqs[dbi]++;
8962 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8964 /* We could return the flags for the FREE_DBI too but what's the point? */
8965 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8967 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8971 /** Add all the DB's pages to the free list.
8972 * @param[in] mc Cursor on the DB to free.
8973 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8974 * @return 0 on success, non-zero on failure.
8977 mdb_drop0(MDB_cursor *mc, int subs)
8981 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8982 if (rc == MDB_SUCCESS) {
8983 MDB_txn *txn = mc->mc_txn;
8988 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8989 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8992 mdb_cursor_copy(mc, &mx);
8993 while (mc->mc_snum > 0) {
8994 MDB_page *mp = mc->mc_pg[mc->mc_top];
8995 unsigned n = NUMKEYS(mp);
8997 for (i=0; i<n; i++) {
8998 ni = NODEPTR(mp, i);
8999 if (ni->mn_flags & F_BIGDATA) {
9002 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9003 rc = mdb_page_get(txn, pg, &omp, NULL);
9006 mdb_cassert(mc, IS_OVERFLOW(omp));
9007 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9011 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9012 mdb_xcursor_init1(mc, ni);
9013 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9019 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9021 for (i=0; i<n; i++) {
9023 ni = NODEPTR(mp, i);
9026 mdb_midl_xappend(txn->mt_free_pgs, pg);
9031 mc->mc_ki[mc->mc_top] = i;
9032 rc = mdb_cursor_sibling(mc, 1);
9034 if (rc != MDB_NOTFOUND)
9036 /* no more siblings, go back to beginning
9037 * of previous level.
9041 for (i=1; i<mc->mc_snum; i++) {
9043 mc->mc_pg[i] = mx.mc_pg[i];
9048 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9051 txn->mt_flags |= MDB_TXN_ERROR;
9052 } else if (rc == MDB_NOTFOUND) {
9058 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9060 MDB_cursor *mc, *m2;
9063 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9066 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9069 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9072 rc = mdb_cursor_open(txn, dbi, &mc);
9076 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9077 /* Invalidate the dropped DB's cursors */
9078 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9079 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9083 /* Can't delete the main DB */
9084 if (del && dbi > MAIN_DBI) {
9085 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9087 txn->mt_dbflags[dbi] = DB_STALE;
9088 mdb_dbi_close(txn->mt_env, dbi);
9090 txn->mt_flags |= MDB_TXN_ERROR;
9093 /* reset the DB record, mark it dirty */
9094 txn->mt_dbflags[dbi] |= DB_DIRTY;
9095 txn->mt_dbs[dbi].md_depth = 0;
9096 txn->mt_dbs[dbi].md_branch_pages = 0;
9097 txn->mt_dbs[dbi].md_leaf_pages = 0;
9098 txn->mt_dbs[dbi].md_overflow_pages = 0;
9099 txn->mt_dbs[dbi].md_entries = 0;
9100 txn->mt_dbs[dbi].md_root = P_INVALID;
9102 txn->mt_flags |= MDB_TXN_DIRTY;
9105 mdb_cursor_close(mc);
9109 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9111 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9114 txn->mt_dbxs[dbi].md_cmp = cmp;
9118 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9120 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9123 txn->mt_dbxs[dbi].md_dcmp = cmp;
9127 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9129 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9132 txn->mt_dbxs[dbi].md_rel = rel;
9136 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9138 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9141 txn->mt_dbxs[dbi].md_relctx = ctx;
9146 mdb_env_get_maxkeysize(MDB_env *env)
9148 return ENV_MAXKEY(env);
9152 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9154 unsigned int i, rdrs;
9157 int rc = 0, first = 1;
9161 if (!env->me_txns) {
9162 return func("(no reader locks)\n", ctx);
9164 rdrs = env->me_txns->mti_numreaders;
9165 mr = env->me_txns->mti_readers;
9166 for (i=0; i<rdrs; i++) {
9168 txnid_t txnid = mr[i].mr_txnid;
9169 sprintf(buf, txnid == (txnid_t)-1 ?
9170 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9171 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9174 rc = func(" pid thread txnid\n", ctx);
9178 rc = func(buf, ctx);
9184 rc = func("(no active readers)\n", ctx);
9189 /** Insert pid into list if not already present.
9190 * return -1 if already present.
9193 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9195 /* binary search of pid in list */
9197 unsigned cursor = 1;
9199 unsigned n = ids[0];
9202 unsigned pivot = n >> 1;
9203 cursor = base + pivot + 1;
9204 val = pid - ids[cursor];
9209 } else if ( val > 0 ) {
9214 /* found, so it's a duplicate */
9223 for (n = ids[0]; n > cursor; n--)
9230 mdb_reader_check(MDB_env *env, int *dead)
9232 unsigned int i, j, rdrs;
9234 MDB_PID_T *pids, pid;
9243 rdrs = env->me_txns->mti_numreaders;
9244 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9248 mr = env->me_txns->mti_readers;
9249 for (i=0; i<rdrs; i++) {
9250 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9252 if (mdb_pid_insert(pids, pid) == 0) {
9253 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9255 /* Recheck, a new process may have reused pid */
9256 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9257 for (j=i; j<rdrs; j++)
9258 if (mr[j].mr_pid == pid) {
9259 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9260 (unsigned) pid, mr[j].mr_txnid));
9265 UNLOCK_MUTEX_R(env);