2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2014 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
41 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
42 * as int64 which is wrong. MSVC doesn't define it at all, so just
46 #define MDB_THR_T DWORD
47 #include <sys/types.h>
50 # include <sys/param.h>
52 # define LITTLE_ENDIAN 1234
53 # define BIG_ENDIAN 4321
54 # define BYTE_ORDER LITTLE_ENDIAN
56 # define SSIZE_MAX INT_MAX
60 #include <sys/types.h>
62 #define MDB_PID_T pid_t
63 #define MDB_THR_T pthread_t
64 #include <sys/param.h>
67 #ifdef HAVE_SYS_FILE_H
83 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
84 #include <netinet/in.h>
85 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
88 #if defined(__APPLE__) || defined (BSD)
89 # define MDB_USE_POSIX_SEM 1
90 # define MDB_FDATASYNC fsync
91 #elif defined(ANDROID)
92 # define MDB_FDATASYNC fsync
97 #ifdef MDB_USE_POSIX_SEM
98 # define MDB_USE_HASH 1
99 #include <semaphore.h>
104 #include <valgrind/memcheck.h>
105 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
106 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
107 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
108 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
109 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
111 #define VGMEMP_CREATE(h,r,z)
112 #define VGMEMP_ALLOC(h,a,s)
113 #define VGMEMP_FREE(h,a)
114 #define VGMEMP_DESTROY(h)
115 #define VGMEMP_DEFINED(a,s)
119 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
120 /* Solaris just defines one or the other */
121 # define LITTLE_ENDIAN 1234
122 # define BIG_ENDIAN 4321
123 # ifdef _LITTLE_ENDIAN
124 # define BYTE_ORDER LITTLE_ENDIAN
126 # define BYTE_ORDER BIG_ENDIAN
129 # define BYTE_ORDER __BYTE_ORDER
133 #ifndef LITTLE_ENDIAN
134 #define LITTLE_ENDIAN __LITTLE_ENDIAN
137 #define BIG_ENDIAN __BIG_ENDIAN
140 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
141 #define MISALIGNED_OK 1
147 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
148 # error "Unknown or unsupported endianness (BYTE_ORDER)"
149 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
150 # error "Two's complement, reasonably sized integer types, please"
154 /** Put infrequently used env functions in separate section */
155 #define ESECT __attribute__ ((section("text_env")))
160 /** @defgroup internal LMDB Internals
163 /** @defgroup compat Compatibility Macros
164 * A bunch of macros to minimize the amount of platform-specific ifdefs
165 * needed throughout the rest of the code. When the features this library
166 * needs are similar enough to POSIX to be hidden in a one-or-two line
167 * replacement, this macro approach is used.
171 /** Wrapper around __func__, which is a C99 feature */
172 #if __STDC_VERSION__ >= 199901L
173 # define mdb_func_ __func__
174 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
175 # define mdb_func_ __FUNCTION__
177 /* If a debug message says <mdb_unknown>(), update the #if statements above */
178 # define mdb_func_ "<mdb_unknown>"
182 #define MDB_USE_HASH 1
183 #define MDB_PIDLOCK 0
184 #define THREAD_RET DWORD
185 #define pthread_t HANDLE
186 #define pthread_mutex_t HANDLE
187 #define pthread_cond_t HANDLE
188 #define pthread_key_t DWORD
189 #define pthread_self() GetCurrentThreadId()
190 #define pthread_key_create(x,y) \
191 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
192 #define pthread_key_delete(x) TlsFree(x)
193 #define pthread_getspecific(x) TlsGetValue(x)
194 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
195 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
196 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
197 #define pthread_cond_signal(x) SetEvent(*x)
198 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
199 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
200 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
201 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
202 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
203 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
204 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
205 #define getpid() GetCurrentProcessId()
206 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
207 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
208 #define ErrCode() GetLastError()
209 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
210 #define close(fd) (CloseHandle(fd) ? 0 : -1)
211 #define munmap(ptr,len) UnmapViewOfFile(ptr)
212 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
213 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
215 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
219 #define THREAD_RET void *
220 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
221 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
222 #define Z "z" /**< printf format modifier for size_t */
224 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
225 #define MDB_PIDLOCK 1
227 #ifdef MDB_USE_POSIX_SEM
229 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
230 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
231 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
232 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
235 mdb_sem_wait(sem_t *sem)
238 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
243 /** Lock the reader mutex.
245 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
246 /** Unlock the reader mutex.
248 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
250 /** Lock the writer mutex.
251 * Only a single write transaction is allowed at a time. Other writers
252 * will block waiting for this mutex.
254 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
255 /** Unlock the writer mutex.
257 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
258 #endif /* MDB_USE_POSIX_SEM */
260 /** Get the error code for the last failed system function.
262 #define ErrCode() errno
264 /** An abstraction for a file handle.
265 * On POSIX systems file handles are small integers. On Windows
266 * they're opaque pointers.
270 /** A value for an invalid file handle.
271 * Mainly used to initialize file variables and signify that they are
274 #define INVALID_HANDLE_VALUE (-1)
276 /** Get the size of a memory page for the system.
277 * This is the basic size that the platform's memory manager uses, and is
278 * fundamental to the use of memory-mapped files.
280 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
283 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
286 #define MNAME_LEN (sizeof(pthread_mutex_t))
292 /** A flag for opening a file and requesting synchronous data writes.
293 * This is only used when writing a meta page. It's not strictly needed;
294 * we could just do a normal write and then immediately perform a flush.
295 * But if this flag is available it saves us an extra system call.
297 * @note If O_DSYNC is undefined but exists in /usr/include,
298 * preferably set some compiler flag to get the definition.
299 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
302 # define MDB_DSYNC O_DSYNC
306 /** Function for flushing the data of a file. Define this to fsync
307 * if fdatasync() is not supported.
309 #ifndef MDB_FDATASYNC
310 # define MDB_FDATASYNC fdatasync
314 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
325 /** A page number in the database.
326 * Note that 64 bit page numbers are overkill, since pages themselves
327 * already represent 12-13 bits of addressable memory, and the OS will
328 * always limit applications to a maximum of 63 bits of address space.
330 * @note In the #MDB_node structure, we only store 48 bits of this value,
331 * which thus limits us to only 60 bits of addressable data.
333 typedef MDB_ID pgno_t;
335 /** A transaction ID.
336 * See struct MDB_txn.mt_txnid for details.
338 typedef MDB_ID txnid_t;
340 /** @defgroup debug Debug Macros
344 /** Enable debug output. Needs variable argument macros (a C99 feature).
345 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
346 * read from and written to the database (used for free space management).
352 static int mdb_debug;
353 static txnid_t mdb_debug_start;
355 /** Print a debug message with printf formatting.
356 * Requires double parenthesis around 2 or more args.
358 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
359 # define DPRINTF0(fmt, ...) \
360 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
362 # define DPRINTF(args) ((void) 0)
364 /** Print a debug string.
365 * The string is printed literally, with no format processing.
367 #define DPUTS(arg) DPRINTF(("%s", arg))
368 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
370 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
373 /** @brief The maximum size of a database page.
375 * This is 32k, since it must fit in #MDB_page.%mp_upper.
377 * LMDB will use database pages < OS pages if needed.
378 * That causes more I/O in write transactions: The OS must
379 * know (read) the whole page before writing a partial page.
381 * Note that we don't currently support Huge pages. On Linux,
382 * regular data files cannot use Huge pages, and in general
383 * Huge pages aren't actually pageable. We rely on the OS
384 * demand-pager to read our data and page it out when memory
385 * pressure from other processes is high. So until OSs have
386 * actual paging support for Huge pages, they're not viable.
388 #define MAX_PAGESIZE 0x8000
390 /** The minimum number of keys required in a database page.
391 * Setting this to a larger value will place a smaller bound on the
392 * maximum size of a data item. Data items larger than this size will
393 * be pushed into overflow pages instead of being stored directly in
394 * the B-tree node. This value used to default to 4. With a page size
395 * of 4096 bytes that meant that any item larger than 1024 bytes would
396 * go into an overflow page. That also meant that on average 2-3KB of
397 * each overflow page was wasted space. The value cannot be lower than
398 * 2 because then there would no longer be a tree structure. With this
399 * value, items larger than 2KB will go into overflow pages, and on
400 * average only 1KB will be wasted.
402 #define MDB_MINKEYS 2
404 /** A stamp that identifies a file as an LMDB file.
405 * There's nothing special about this value other than that it is easily
406 * recognizable, and it will reflect any byte order mismatches.
408 #define MDB_MAGIC 0xBEEFC0DE
410 /** The version number for a database's datafile format. */
411 #define MDB_DATA_VERSION 1
412 /** The version number for a database's lockfile format. */
413 #define MDB_LOCK_VERSION 1
415 /** @brief The max size of a key we can write, or 0 for dynamic max.
417 * Define this as 0 to compute the max from the page size. 511
418 * is default for backwards compat: liblmdb <= 0.9.10 can break
419 * when modifying a DB with keys/dupsort data bigger than its max.
421 * Data items in an #MDB_DUPSORT database are also limited to
422 * this size, since they're actually keys of a sub-DB. Keys and
423 * #MDB_DUPSORT data items must fit on a node in a regular page.
425 #ifndef MDB_MAXKEYSIZE
426 #define MDB_MAXKEYSIZE 511
429 /** The maximum size of a key we can write to the environment. */
431 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
433 #define ENV_MAXKEY(env) ((env)->me_maxkey)
436 /** @brief The maximum size of a data item.
438 * We only store a 32 bit value for node sizes.
440 #define MAXDATASIZE 0xffffffffUL
443 /** Key size which fits in a #DKBUF.
446 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
449 * This is used for printing a hex dump of a key's contents.
451 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
452 /** Display a key in hex.
454 * Invoke a function to display a key in hex.
456 #define DKEY(x) mdb_dkey(x, kbuf)
462 /** An invalid page number.
463 * Mainly used to denote an empty tree.
465 #define P_INVALID (~(pgno_t)0)
467 /** Test if the flags \b f are set in a flag word \b w. */
468 #define F_ISSET(w, f) (((w) & (f)) == (f))
470 /** Round \b n up to an even number. */
471 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
473 /** Used for offsets within a single page.
474 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
477 typedef uint16_t indx_t;
479 /** Default size of memory map.
480 * This is certainly too small for any actual applications. Apps should always set
481 * the size explicitly using #mdb_env_set_mapsize().
483 #define DEFAULT_MAPSIZE 1048576
485 /** @defgroup readers Reader Lock Table
486 * Readers don't acquire any locks for their data access. Instead, they
487 * simply record their transaction ID in the reader table. The reader
488 * mutex is needed just to find an empty slot in the reader table. The
489 * slot's address is saved in thread-specific data so that subsequent read
490 * transactions started by the same thread need no further locking to proceed.
492 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
494 * No reader table is used if the database is on a read-only filesystem, or
495 * if #MDB_NOLOCK is set.
497 * Since the database uses multi-version concurrency control, readers don't
498 * actually need any locking. This table is used to keep track of which
499 * readers are using data from which old transactions, so that we'll know
500 * when a particular old transaction is no longer in use. Old transactions
501 * that have discarded any data pages can then have those pages reclaimed
502 * for use by a later write transaction.
504 * The lock table is constructed such that reader slots are aligned with the
505 * processor's cache line size. Any slot is only ever used by one thread.
506 * This alignment guarantees that there will be no contention or cache
507 * thrashing as threads update their own slot info, and also eliminates
508 * any need for locking when accessing a slot.
510 * A writer thread will scan every slot in the table to determine the oldest
511 * outstanding reader transaction. Any freed pages older than this will be
512 * reclaimed by the writer. The writer doesn't use any locks when scanning
513 * this table. This means that there's no guarantee that the writer will
514 * see the most up-to-date reader info, but that's not required for correct
515 * operation - all we need is to know the upper bound on the oldest reader,
516 * we don't care at all about the newest reader. So the only consequence of
517 * reading stale information here is that old pages might hang around a
518 * while longer before being reclaimed. That's actually good anyway, because
519 * the longer we delay reclaiming old pages, the more likely it is that a
520 * string of contiguous pages can be found after coalescing old pages from
521 * many old transactions together.
524 /** Number of slots in the reader table.
525 * This value was chosen somewhat arbitrarily. 126 readers plus a
526 * couple mutexes fit exactly into 8KB on my development machine.
527 * Applications should set the table size using #mdb_env_set_maxreaders().
529 #define DEFAULT_READERS 126
531 /** The size of a CPU cache line in bytes. We want our lock structures
532 * aligned to this size to avoid false cache line sharing in the
534 * This value works for most CPUs. For Itanium this should be 128.
540 /** The information we store in a single slot of the reader table.
541 * In addition to a transaction ID, we also record the process and
542 * thread ID that owns a slot, so that we can detect stale information,
543 * e.g. threads or processes that went away without cleaning up.
544 * @note We currently don't check for stale records. We simply re-init
545 * the table when we know that we're the only process opening the
548 typedef struct MDB_rxbody {
549 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
550 * Multiple readers that start at the same time will probably have the
551 * same ID here. Again, it's not important to exclude them from
552 * anything; all we need to know is which version of the DB they
553 * started from so we can avoid overwriting any data used in that
554 * particular version.
557 /** The process ID of the process owning this reader txn. */
559 /** The thread ID of the thread owning this txn. */
563 /** The actual reader record, with cacheline padding. */
564 typedef struct MDB_reader {
567 /** shorthand for mrb_txnid */
568 #define mr_txnid mru.mrx.mrb_txnid
569 #define mr_pid mru.mrx.mrb_pid
570 #define mr_tid mru.mrx.mrb_tid
571 /** cache line alignment */
572 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
576 /** The header for the reader table.
577 * The table resides in a memory-mapped file. (This is a different file
578 * than is used for the main database.)
580 * For POSIX the actual mutexes reside in the shared memory of this
581 * mapped file. On Windows, mutexes are named objects allocated by the
582 * kernel; we store the mutex names in this mapped file so that other
583 * processes can grab them. This same approach is also used on
584 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
585 * process-shared POSIX mutexes. For these cases where a named object
586 * is used, the object name is derived from a 64 bit FNV hash of the
587 * environment pathname. As such, naming collisions are extremely
588 * unlikely. If a collision occurs, the results are unpredictable.
590 typedef struct MDB_txbody {
591 /** Stamp identifying this as an LMDB file. It must be set
594 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
596 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
597 char mtb_rmname[MNAME_LEN];
599 /** Mutex protecting access to this table.
600 * This is the reader lock that #LOCK_MUTEX_R acquires.
602 pthread_mutex_t mtb_mutex;
604 /** The ID of the last transaction committed to the database.
605 * This is recorded here only for convenience; the value can always
606 * be determined by reading the main database meta pages.
609 /** The number of slots that have been used in the reader table.
610 * This always records the maximum count, it is not decremented
611 * when readers release their slots.
613 unsigned mtb_numreaders;
616 /** The actual reader table definition. */
617 typedef struct MDB_txninfo {
620 #define mti_magic mt1.mtb.mtb_magic
621 #define mti_format mt1.mtb.mtb_format
622 #define mti_mutex mt1.mtb.mtb_mutex
623 #define mti_rmname mt1.mtb.mtb_rmname
624 #define mti_txnid mt1.mtb.mtb_txnid
625 #define mti_numreaders mt1.mtb.mtb_numreaders
626 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
629 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
630 char mt2_wmname[MNAME_LEN];
631 #define mti_wmname mt2.mt2_wmname
633 pthread_mutex_t mt2_wmutex;
634 #define mti_wmutex mt2.mt2_wmutex
636 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
638 MDB_reader mti_readers[1];
641 /** Lockfile format signature: version, features and field layout */
642 #define MDB_LOCK_FORMAT \
644 ((MDB_LOCK_VERSION) \
645 /* Flags which describe functionality */ \
646 + (((MDB_PIDLOCK) != 0) << 16)))
649 /** Common header for all page types.
650 * Overflow records occupy a number of contiguous pages with no
651 * headers on any page after the first.
653 typedef struct MDB_page {
654 #define mp_pgno mp_p.p_pgno
655 #define mp_next mp_p.p_next
657 pgno_t p_pgno; /**< page number */
658 void * p_next; /**< for in-memory list of freed structs */
661 /** @defgroup mdb_page Page Flags
663 * Flags for the page headers.
666 #define P_BRANCH 0x01 /**< branch page */
667 #define P_LEAF 0x02 /**< leaf page */
668 #define P_OVERFLOW 0x04 /**< overflow page */
669 #define P_META 0x08 /**< meta page */
670 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
671 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
672 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
673 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
674 #define P_KEEP 0x8000 /**< leave this page alone during spill */
676 uint16_t mp_flags; /**< @ref mdb_page */
677 #define mp_lower mp_pb.pb.pb_lower
678 #define mp_upper mp_pb.pb.pb_upper
679 #define mp_pages mp_pb.pb_pages
682 indx_t pb_lower; /**< lower bound of free space */
683 indx_t pb_upper; /**< upper bound of free space */
685 uint32_t pb_pages; /**< number of overflow pages */
687 indx_t mp_ptrs[1]; /**< dynamic size */
690 /** Size of the page header, excluding dynamic data at the end */
691 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
693 /** Address of first usable data byte in a page, after the header */
694 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
696 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
699 #define MP_HIBASE PAGEHDRSZ
701 #define MP_LOBASE PAGEHDRSZ
705 /** Number of nodes on a page */
706 #define NUMKEYS(p) (((p)->mp_lower - MP_LOBASE) >> 1)
708 /** The amount of space remaining in the page */
709 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
711 /** The percentage of space used in the page, in tenths of a percent. */
712 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
713 ((env)->me_psize - PAGEHDRSZ))
714 /** The minimum page fill factor, in tenths of a percent.
715 * Pages emptier than this are candidates for merging.
717 #define FILL_THRESHOLD 250
719 /** Test if a page is a leaf page */
720 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
721 /** Test if a page is a LEAF2 page */
722 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
723 /** Test if a page is a branch page */
724 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
725 /** Test if a page is an overflow page */
726 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
727 /** Test if a page is a sub page */
728 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
730 /** The number of overflow pages needed to store the given size. */
731 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
733 /** Link in #MDB_txn.%mt_loose_pages list */
734 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)METADATA(p))
736 /** Header for a single key/data pair within a page.
737 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
738 * We guarantee 2-byte alignment for 'MDB_node's.
740 typedef struct MDB_node {
741 /** lo and hi are used for data size on leaf nodes and for
742 * child pgno on branch nodes. On 64 bit platforms, flags
743 * is also used for pgno. (Branch nodes have no flags).
744 * They are in host byte order in case that lets some
745 * accesses be optimized into a 32-bit word access.
747 #if BYTE_ORDER == LITTLE_ENDIAN
748 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
750 unsigned short mn_hi, mn_lo;
752 /** @defgroup mdb_node Node Flags
754 * Flags for node headers.
757 #define F_BIGDATA 0x01 /**< data put on overflow page */
758 #define F_SUBDATA 0x02 /**< data is a sub-database */
759 #define F_DUPDATA 0x04 /**< data has duplicates */
761 /** valid flags for #mdb_node_add() */
762 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
765 unsigned short mn_flags; /**< @ref mdb_node */
766 unsigned short mn_ksize; /**< key size */
767 char mn_data[1]; /**< key and data are appended here */
770 /** Size of the node header, excluding dynamic data at the end */
771 #define NODESIZE offsetof(MDB_node, mn_data)
773 /** Bit position of top word in page number, for shifting mn_flags */
774 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
776 /** Size of a node in a branch page with a given key.
777 * This is just the node header plus the key, there is no data.
779 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
781 /** Size of a node in a leaf page with a given key and data.
782 * This is node header plus key plus data size.
784 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
786 /** Address of node \b i in page \b p */
787 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + MP_HIBASE))
789 /** Address of the key for the node */
790 #define NODEKEY(node) (void *)((node)->mn_data)
792 /** Address of the data for a node */
793 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
795 /** Get the page number pointed to by a branch node */
796 #define NODEPGNO(node) \
797 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
798 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
799 /** Set the page number in a branch node */
800 #define SETPGNO(node,pgno) do { \
801 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
802 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
804 /** Get the size of the data in a leaf node */
805 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
806 /** Set the size of the data for a leaf node */
807 #define SETDSZ(node,size) do { \
808 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
809 /** The size of a key in a node */
810 #define NODEKSZ(node) ((node)->mn_ksize)
812 /** Copy a page number from src to dst */
814 #define COPY_PGNO(dst,src) dst = src
816 #if SIZE_MAX > 4294967295UL
817 #define COPY_PGNO(dst,src) do { \
818 unsigned short *s, *d; \
819 s = (unsigned short *)&(src); \
820 d = (unsigned short *)&(dst); \
827 #define COPY_PGNO(dst,src) do { \
828 unsigned short *s, *d; \
829 s = (unsigned short *)&(src); \
830 d = (unsigned short *)&(dst); \
836 /** The address of a key in a LEAF2 page.
837 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
838 * There are no node headers, keys are stored contiguously.
840 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
842 /** Set the \b node's key into \b keyptr, if requested. */
843 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
844 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
846 /** Set the \b node's key into \b key. */
847 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
849 /** Information about a single database in the environment. */
850 typedef struct MDB_db {
851 uint32_t md_pad; /**< also ksize for LEAF2 pages */
852 uint16_t md_flags; /**< @ref mdb_dbi_open */
853 uint16_t md_depth; /**< depth of this tree */
854 pgno_t md_branch_pages; /**< number of internal pages */
855 pgno_t md_leaf_pages; /**< number of leaf pages */
856 pgno_t md_overflow_pages; /**< number of overflow pages */
857 size_t md_entries; /**< number of data items */
858 pgno_t md_root; /**< the root page of this tree */
861 /** mdb_dbi_open flags */
862 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
863 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
864 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
865 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
867 /** Handle for the DB used to track free pages. */
869 /** Handle for the default DB. */
872 /** Meta page content.
873 * A meta page is the start point for accessing a database snapshot.
874 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
876 typedef struct MDB_meta {
877 /** Stamp identifying this as an LMDB file. It must be set
880 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
882 void *mm_address; /**< address for fixed mapping */
883 size_t mm_mapsize; /**< size of mmap region */
884 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
885 /** The size of pages used in this DB */
886 #define mm_psize mm_dbs[0].md_pad
887 /** Any persistent environment flags. @ref mdb_env */
888 #define mm_flags mm_dbs[0].md_flags
889 pgno_t mm_last_pg; /**< last used page in file */
890 txnid_t mm_txnid; /**< txnid that committed this page */
893 /** Buffer for a stack-allocated meta page.
894 * The members define size and alignment, and silence type
895 * aliasing warnings. They are not used directly; that could
896 * mean incorrectly using several union members in parallel.
898 typedef union MDB_metabuf {
901 char mm_pad[PAGEHDRSZ];
906 /** Auxiliary DB info.
907 * The information here is mostly static/read-only. There is
908 * only a single copy of this record in the environment.
910 typedef struct MDB_dbx {
911 MDB_val md_name; /**< name of the database */
912 MDB_cmp_func *md_cmp; /**< function for comparing keys */
913 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
914 MDB_rel_func *md_rel; /**< user relocate function */
915 void *md_relctx; /**< user-provided context for md_rel */
918 /** A database transaction.
919 * Every operation requires a transaction handle.
922 MDB_txn *mt_parent; /**< parent of a nested txn */
923 MDB_txn *mt_child; /**< nested txn under this txn */
924 pgno_t mt_next_pgno; /**< next unallocated page */
925 /** The ID of this transaction. IDs are integers incrementing from 1.
926 * Only committed write transactions increment the ID. If a transaction
927 * aborts, the ID may be re-used by the next writer.
930 MDB_env *mt_env; /**< the DB environment */
931 /** The list of pages that became unused during this transaction.
934 /** The list of loose pages that became unused and may be reused
935 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
937 MDB_page *mt_loose_pgs;
938 /** The sorted list of dirty pages we temporarily wrote to disk
939 * because the dirty list was full. page numbers in here are
940 * shifted left by 1, deleted slots have the LSB set.
942 MDB_IDL mt_spill_pgs;
944 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
946 /** For read txns: This thread/txn's reader table slot, or NULL. */
949 /** Array of records for each DB known in the environment. */
951 /** Array of MDB_db records for each known DB */
953 /** @defgroup mt_dbflag Transaction DB Flags
957 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
958 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
959 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
960 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
962 /** In write txns, array of cursors for each DB */
963 MDB_cursor **mt_cursors;
964 /** Array of flags for each DB */
965 unsigned char *mt_dbflags;
966 /** Number of DB records in use. This number only ever increments;
967 * we don't decrement it when individual DB handles are closed.
971 /** @defgroup mdb_txn Transaction Flags
975 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
976 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
977 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
978 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
980 unsigned int mt_flags; /**< @ref mdb_txn */
981 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
982 * Includes ancestor txns' dirty pages not hidden by other txns'
983 * dirty/spilled pages. Thus commit(nested txn) has room to merge
984 * dirty_list into mt_parent after freeing hidden mt_parent pages.
986 unsigned int mt_dirty_room;
989 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
990 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
991 * raise this on a 64 bit machine.
993 #define CURSOR_STACK 32
997 /** Cursors are used for all DB operations.
998 * A cursor holds a path of (page pointer, key index) from the DB
999 * root to a position in the DB, plus other state. #MDB_DUPSORT
1000 * cursors include an xcursor to the current data item. Write txns
1001 * track their cursors and keep them up to date when data moves.
1002 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1003 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1006 /** Next cursor on this DB in this txn */
1007 MDB_cursor *mc_next;
1008 /** Backup of the original cursor if this cursor is a shadow */
1009 MDB_cursor *mc_backup;
1010 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1011 struct MDB_xcursor *mc_xcursor;
1012 /** The transaction that owns this cursor */
1014 /** The database handle this cursor operates on */
1016 /** The database record for this cursor */
1018 /** The database auxiliary record for this cursor */
1020 /** The @ref mt_dbflag for this database */
1021 unsigned char *mc_dbflag;
1022 unsigned short mc_snum; /**< number of pushed pages */
1023 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1024 /** @defgroup mdb_cursor Cursor Flags
1026 * Cursor state flags.
1029 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1030 #define C_EOF 0x02 /**< No more data */
1031 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1032 #define C_DEL 0x08 /**< last op was a cursor_del */
1033 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1034 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1036 unsigned int mc_flags; /**< @ref mdb_cursor */
1037 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1038 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1041 /** Context for sorted-dup records.
1042 * We could have gone to a fully recursive design, with arbitrarily
1043 * deep nesting of sub-databases. But for now we only handle these
1044 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1046 typedef struct MDB_xcursor {
1047 /** A sub-cursor for traversing the Dup DB */
1048 MDB_cursor mx_cursor;
1049 /** The database record for this Dup DB */
1051 /** The auxiliary DB record for this Dup DB */
1053 /** The @ref mt_dbflag for this Dup DB */
1054 unsigned char mx_dbflag;
1057 /** State of FreeDB old pages, stored in the MDB_env */
1058 typedef struct MDB_pgstate {
1059 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1060 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1063 /** The database environment. */
1065 HANDLE me_fd; /**< The main data file */
1066 HANDLE me_lfd; /**< The lock file */
1067 HANDLE me_mfd; /**< just for writing the meta pages */
1068 /** Failed to update the meta page. Probably an I/O error. */
1069 #define MDB_FATAL_ERROR 0x80000000U
1070 /** Some fields are initialized. */
1071 #define MDB_ENV_ACTIVE 0x20000000U
1072 /** me_txkey is set */
1073 #define MDB_ENV_TXKEY 0x10000000U
1074 uint32_t me_flags; /**< @ref mdb_env */
1075 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1076 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1077 unsigned int me_maxreaders; /**< size of the reader table */
1078 unsigned int me_numreaders; /**< max numreaders set by this env */
1079 MDB_dbi me_numdbs; /**< number of DBs opened */
1080 MDB_dbi me_maxdbs; /**< size of the DB table */
1081 MDB_PID_T me_pid; /**< process ID of this env */
1082 char *me_path; /**< path to the DB files */
1083 char *me_map; /**< the memory map of the data file */
1084 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1085 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1086 void *me_pbuf; /**< scratch area for DUPSORT put() */
1087 MDB_txn *me_txn; /**< current write transaction */
1088 size_t me_mapsize; /**< size of the data memory map */
1089 off_t me_size; /**< current file size */
1090 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1091 MDB_dbx *me_dbxs; /**< array of static DB info */
1092 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1093 pthread_key_t me_txkey; /**< thread-key for readers */
1094 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1095 # define me_pglast me_pgstate.mf_pglast
1096 # define me_pghead me_pgstate.mf_pghead
1097 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1098 /** IDL of pages that became unused in a write txn */
1099 MDB_IDL me_free_pgs;
1100 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1101 MDB_ID2L me_dirty_list;
1102 /** Max number of freelist items that can fit in a single overflow page */
1104 /** Max size of a node on a page */
1105 unsigned int me_nodemax;
1106 #if !(MDB_MAXKEYSIZE)
1107 unsigned int me_maxkey; /**< max size of a key */
1109 int me_live_reader; /**< have liveness lock in reader table */
1111 int me_pidquery; /**< Used in OpenProcess */
1112 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1114 #elif defined(MDB_USE_POSIX_SEM)
1115 sem_t *me_rmutex; /* Shared mutexes are not supported */
1118 void *me_userctx; /**< User-settable context */
1119 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1122 /** Nested transaction */
1123 typedef struct MDB_ntxn {
1124 MDB_txn mnt_txn; /**< the transaction */
1125 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1128 /** max number of pages to commit in one writev() call */
1129 #define MDB_COMMIT_PAGES 64
1130 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1131 #undef MDB_COMMIT_PAGES
1132 #define MDB_COMMIT_PAGES IOV_MAX
1135 /** max bytes to write in one call */
1136 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1138 /** Check \b txn and \b dbi arguments to a function */
1139 #define TXN_DBI_EXIST(txn, dbi) \
1140 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1142 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1143 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1144 static int mdb_page_touch(MDB_cursor *mc);
1146 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1147 static int mdb_page_search_root(MDB_cursor *mc,
1148 MDB_val *key, int modify);
1149 #define MDB_PS_MODIFY 1
1150 #define MDB_PS_ROOTONLY 2
1151 #define MDB_PS_FIRST 4
1152 #define MDB_PS_LAST 8
1153 static int mdb_page_search(MDB_cursor *mc,
1154 MDB_val *key, int flags);
1155 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1157 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1158 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1159 pgno_t newpgno, unsigned int nflags);
1161 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1162 static int mdb_env_pick_meta(const MDB_env *env);
1163 static int mdb_env_write_meta(MDB_txn *txn);
1164 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1165 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1167 static void mdb_env_close0(MDB_env *env, int excl);
1169 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1170 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1171 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1172 static void mdb_node_del(MDB_cursor *mc, int ksize);
1173 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1174 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1175 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1176 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1177 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1179 static int mdb_rebalance(MDB_cursor *mc);
1180 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1182 static void mdb_cursor_pop(MDB_cursor *mc);
1183 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1185 static int mdb_cursor_del0(MDB_cursor *mc);
1186 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1187 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1188 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1189 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1190 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1192 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1193 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1195 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1196 static void mdb_xcursor_init0(MDB_cursor *mc);
1197 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1199 static int mdb_drop0(MDB_cursor *mc, int subs);
1200 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1203 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1207 static SECURITY_DESCRIPTOR mdb_null_sd;
1208 static SECURITY_ATTRIBUTES mdb_all_sa;
1209 static int mdb_sec_inited;
1212 /** Return the library version info. */
1214 mdb_version(int *major, int *minor, int *patch)
1216 if (major) *major = MDB_VERSION_MAJOR;
1217 if (minor) *minor = MDB_VERSION_MINOR;
1218 if (patch) *patch = MDB_VERSION_PATCH;
1219 return MDB_VERSION_STRING;
1222 /** Table of descriptions for LMDB @ref errors */
1223 static char *const mdb_errstr[] = {
1224 "MDB_KEYEXIST: Key/data pair already exists",
1225 "MDB_NOTFOUND: No matching key/data pair found",
1226 "MDB_PAGE_NOTFOUND: Requested page not found",
1227 "MDB_CORRUPTED: Located page was wrong type",
1228 "MDB_PANIC: Update of meta page failed",
1229 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1230 "MDB_INVALID: File is not an LMDB file",
1231 "MDB_MAP_FULL: Environment mapsize limit reached",
1232 "MDB_DBS_FULL: Environment maxdbs limit reached",
1233 "MDB_READERS_FULL: Environment maxreaders limit reached",
1234 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1235 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1236 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1237 "MDB_PAGE_FULL: Internal error - page has no more space",
1238 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1239 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1240 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1241 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1242 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1246 mdb_strerror(int err)
1250 return ("Successful return: 0");
1252 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1253 i = err - MDB_KEYEXIST;
1254 return mdb_errstr[i];
1257 return strerror(err);
1260 /** assert(3) variant in cursor context */
1261 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1262 /** assert(3) variant in transaction context */
1263 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1264 /** assert(3) variant in environment context */
1265 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1268 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1269 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1272 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1273 const char *func, const char *file, int line)
1276 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1277 file, line, expr_txt, func);
1278 if (env->me_assert_func)
1279 env->me_assert_func(env, buf);
1280 fprintf(stderr, "%s\n", buf);
1284 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1288 /** Return the page number of \b mp which may be sub-page, for debug output */
1290 mdb_dbg_pgno(MDB_page *mp)
1293 COPY_PGNO(ret, mp->mp_pgno);
1297 /** Display a key in hexadecimal and return the address of the result.
1298 * @param[in] key the key to display
1299 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1300 * @return The key in hexadecimal form.
1303 mdb_dkey(MDB_val *key, char *buf)
1306 unsigned char *c = key->mv_data;
1312 if (key->mv_size > DKBUF_MAXKEYSIZE)
1313 return "MDB_MAXKEYSIZE";
1314 /* may want to make this a dynamic check: if the key is mostly
1315 * printable characters, print it as-is instead of converting to hex.
1319 for (i=0; i<key->mv_size; i++)
1320 ptr += sprintf(ptr, "%02x", *c++);
1322 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1328 mdb_leafnode_type(MDB_node *n)
1330 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1331 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1332 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1335 /** Display all the keys in the page. */
1337 mdb_page_list(MDB_page *mp)
1339 pgno_t pgno = mdb_dbg_pgno(mp);
1340 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1342 unsigned int i, nkeys, nsize, total = 0;
1346 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1347 case P_BRANCH: type = "Branch page"; break;
1348 case P_LEAF: type = "Leaf page"; break;
1349 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1350 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1351 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1353 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1354 pgno, mp->mp_pages, state);
1357 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1358 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1361 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1365 nkeys = NUMKEYS(mp);
1366 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1368 for (i=0; i<nkeys; i++) {
1369 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1370 key.mv_size = nsize = mp->mp_pad;
1371 key.mv_data = LEAF2KEY(mp, i, nsize);
1373 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1376 node = NODEPTR(mp, i);
1377 key.mv_size = node->mn_ksize;
1378 key.mv_data = node->mn_data;
1379 nsize = NODESIZE + key.mv_size;
1380 if (IS_BRANCH(mp)) {
1381 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1385 if (F_ISSET(node->mn_flags, F_BIGDATA))
1386 nsize += sizeof(pgno_t);
1388 nsize += NODEDSZ(node);
1390 nsize += sizeof(indx_t);
1391 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1392 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1394 total = EVEN(total);
1396 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1397 IS_LEAF2(mp) ? PAGEHDRSZ : MP_HIBASE + mp->mp_lower, total, SIZELEFT(mp));
1401 mdb_cursor_chk(MDB_cursor *mc)
1407 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1408 for (i=0; i<mc->mc_top; i++) {
1410 node = NODEPTR(mp, mc->mc_ki[i]);
1411 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1414 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1420 /** Count all the pages in each DB and in the freelist
1421 * and make sure it matches the actual number of pages
1423 * All named DBs must be open for a correct count.
1425 static void mdb_audit(MDB_txn *txn)
1429 MDB_ID freecount, count;
1434 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1435 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1436 freecount += *(MDB_ID *)data.mv_data;
1437 mdb_tassert(txn, rc == MDB_NOTFOUND);
1440 for (i = 0; i<txn->mt_numdbs; i++) {
1442 if (!(txn->mt_dbflags[i] & DB_VALID))
1444 mdb_cursor_init(&mc, txn, i, &mx);
1445 if (txn->mt_dbs[i].md_root == P_INVALID)
1447 count += txn->mt_dbs[i].md_branch_pages +
1448 txn->mt_dbs[i].md_leaf_pages +
1449 txn->mt_dbs[i].md_overflow_pages;
1450 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1451 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1452 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1455 mp = mc.mc_pg[mc.mc_top];
1456 for (j=0; j<NUMKEYS(mp); j++) {
1457 MDB_node *leaf = NODEPTR(mp, j);
1458 if (leaf->mn_flags & F_SUBDATA) {
1460 memcpy(&db, NODEDATA(leaf), sizeof(db));
1461 count += db.md_branch_pages + db.md_leaf_pages +
1462 db.md_overflow_pages;
1466 mdb_tassert(txn, rc == MDB_NOTFOUND);
1469 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1470 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1471 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1477 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1479 return txn->mt_dbxs[dbi].md_cmp(a, b);
1483 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1485 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1488 /** Allocate memory for a page.
1489 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1492 mdb_page_malloc(MDB_txn *txn, unsigned num)
1494 MDB_env *env = txn->mt_env;
1495 MDB_page *ret = env->me_dpages;
1496 size_t psize = env->me_psize, sz = psize, off;
1497 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1498 * For a single page alloc, we init everything after the page header.
1499 * For multi-page, we init the final page; if the caller needed that
1500 * many pages they will be filling in at least up to the last page.
1504 VGMEMP_ALLOC(env, ret, sz);
1505 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1506 env->me_dpages = ret->mp_next;
1509 psize -= off = PAGEHDRSZ;
1514 if ((ret = malloc(sz)) != NULL) {
1515 VGMEMP_ALLOC(env, ret, sz);
1516 if (!(env->me_flags & MDB_NOMEMINIT)) {
1517 memset((char *)ret + off, 0, psize);
1521 txn->mt_flags |= MDB_TXN_ERROR;
1525 /** Free a single page.
1526 * Saves single pages to a list, for future reuse.
1527 * (This is not used for multi-page overflow pages.)
1530 mdb_page_free(MDB_env *env, MDB_page *mp)
1532 mp->mp_next = env->me_dpages;
1533 VGMEMP_FREE(env, mp);
1534 env->me_dpages = mp;
1537 /** Free a dirty page */
1539 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1541 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1542 mdb_page_free(env, dp);
1544 /* large pages just get freed directly */
1545 VGMEMP_FREE(env, dp);
1550 /** Return all dirty pages to dpage list */
1552 mdb_dlist_free(MDB_txn *txn)
1554 MDB_env *env = txn->mt_env;
1555 MDB_ID2L dl = txn->mt_u.dirty_list;
1556 unsigned i, n = dl[0].mid;
1558 for (i = 1; i <= n; i++) {
1559 mdb_dpage_free(env, dl[i].mptr);
1564 /** Loosen or free a single page.
1565 * Saves single pages to a list for future reuse
1566 * in this same txn. It has been pulled from the freeDB
1567 * and already resides on the dirty list, but has been
1568 * deleted. Use these pages first before pulling again
1571 * If the page wasn't dirtied in this txn, just add it
1572 * to this txn's free list.
1575 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1578 pgno_t pgno = mp->mp_pgno;
1580 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1581 if (mc->mc_txn->mt_parent) {
1582 MDB_ID2 *dl = mc->mc_txn->mt_u.dirty_list;
1583 /* If txn has a parent, make sure the page is in our
1587 unsigned x = mdb_mid2l_search(dl, pgno);
1588 if (x <= dl[0].mid && dl[x].mid == pgno) {
1589 if (mp != dl[x].mptr) { /* bad cursor? */
1590 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1591 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
1592 return MDB_CORRUPTED;
1599 /* no parent txn, so it's just ours */
1604 NEXT_LOOSE_PAGE(mp) = mc->mc_txn->mt_loose_pgs;
1605 mc->mc_txn->mt_loose_pgs = mp;
1606 mp->mp_flags |= P_LOOSE;
1608 int rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, pgno);
1616 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1617 * @param[in] mc A cursor handle for the current operation.
1618 * @param[in] pflags Flags of the pages to update:
1619 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1620 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1621 * @return 0 on success, non-zero on failure.
1624 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1626 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1627 MDB_txn *txn = mc->mc_txn;
1633 int rc = MDB_SUCCESS, level;
1635 /* Mark pages seen by cursors */
1636 if (mc->mc_flags & C_UNTRACK)
1637 mc = NULL; /* will find mc in mt_cursors */
1638 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1639 for (; mc; mc=mc->mc_next) {
1640 if (!(mc->mc_flags & C_INITIALIZED))
1642 for (m3 = mc;; m3 = &mx->mx_cursor) {
1644 for (j=0; j<m3->mc_snum; j++) {
1646 if ((mp->mp_flags & Mask) == pflags)
1647 mp->mp_flags ^= P_KEEP;
1649 mx = m3->mc_xcursor;
1650 /* Proceed to mx if it is at a sub-database */
1651 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1653 if (! (mp && (mp->mp_flags & P_LEAF)))
1655 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1656 if (!(leaf->mn_flags & F_SUBDATA))
1664 /* Loose pages shouldn't be spilled */
1665 for (dp = txn->mt_loose_pgs; dp; dp = NEXT_LOOSE_PAGE(dp)) {
1666 if ((dp->mp_flags & Mask) == pflags)
1667 dp->mp_flags ^= P_KEEP;
1671 /* Mark dirty root pages */
1672 for (i=0; i<txn->mt_numdbs; i++) {
1673 if (txn->mt_dbflags[i] & DB_DIRTY) {
1674 pgno_t pgno = txn->mt_dbs[i].md_root;
1675 if (pgno == P_INVALID)
1677 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1679 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1680 dp->mp_flags ^= P_KEEP;
1688 static int mdb_page_flush(MDB_txn *txn, int keep);
1690 /** Spill pages from the dirty list back to disk.
1691 * This is intended to prevent running into #MDB_TXN_FULL situations,
1692 * but note that they may still occur in a few cases:
1693 * 1) our estimate of the txn size could be too small. Currently this
1694 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1695 * 2) child txns may run out of space if their parents dirtied a
1696 * lot of pages and never spilled them. TODO: we probably should do
1697 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1698 * the parent's dirty_room is below a given threshold.
1700 * Otherwise, if not using nested txns, it is expected that apps will
1701 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1702 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1703 * If the txn never references them again, they can be left alone.
1704 * If the txn only reads them, they can be used without any fuss.
1705 * If the txn writes them again, they can be dirtied immediately without
1706 * going thru all of the work of #mdb_page_touch(). Such references are
1707 * handled by #mdb_page_unspill().
1709 * Also note, we never spill DB root pages, nor pages of active cursors,
1710 * because we'll need these back again soon anyway. And in nested txns,
1711 * we can't spill a page in a child txn if it was already spilled in a
1712 * parent txn. That would alter the parent txns' data even though
1713 * the child hasn't committed yet, and we'd have no way to undo it if
1714 * the child aborted.
1716 * @param[in] m0 cursor A cursor handle identifying the transaction and
1717 * database for which we are checking space.
1718 * @param[in] key For a put operation, the key being stored.
1719 * @param[in] data For a put operation, the data being stored.
1720 * @return 0 on success, non-zero on failure.
1723 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1725 MDB_txn *txn = m0->mc_txn;
1727 MDB_ID2L dl = txn->mt_u.dirty_list;
1728 unsigned int i, j, need;
1731 if (m0->mc_flags & C_SUB)
1734 /* Estimate how much space this op will take */
1735 i = m0->mc_db->md_depth;
1736 /* Named DBs also dirty the main DB */
1737 if (m0->mc_dbi > MAIN_DBI)
1738 i += txn->mt_dbs[MAIN_DBI].md_depth;
1739 /* For puts, roughly factor in the key+data size */
1741 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1742 i += i; /* double it for good measure */
1745 if (txn->mt_dirty_room > i)
1748 if (!txn->mt_spill_pgs) {
1749 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1750 if (!txn->mt_spill_pgs)
1753 /* purge deleted slots */
1754 MDB_IDL sl = txn->mt_spill_pgs;
1755 unsigned int num = sl[0];
1757 for (i=1; i<=num; i++) {
1764 /* Preserve pages which may soon be dirtied again */
1765 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1768 /* Less aggressive spill - we originally spilled the entire dirty list,
1769 * with a few exceptions for cursor pages and DB root pages. But this
1770 * turns out to be a lot of wasted effort because in a large txn many
1771 * of those pages will need to be used again. So now we spill only 1/8th
1772 * of the dirty pages. Testing revealed this to be a good tradeoff,
1773 * better than 1/2, 1/4, or 1/10.
1775 if (need < MDB_IDL_UM_MAX / 8)
1776 need = MDB_IDL_UM_MAX / 8;
1778 /* Save the page IDs of all the pages we're flushing */
1779 /* flush from the tail forward, this saves a lot of shifting later on. */
1780 for (i=dl[0].mid; i && need; i--) {
1781 MDB_ID pn = dl[i].mid << 1;
1783 if (dp->mp_flags & P_KEEP)
1785 /* Can't spill twice, make sure it's not already in a parent's
1788 if (txn->mt_parent) {
1790 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1791 if (tx2->mt_spill_pgs) {
1792 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1793 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1794 dp->mp_flags |= P_KEEP;
1802 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1806 mdb_midl_sort(txn->mt_spill_pgs);
1808 /* Flush the spilled part of dirty list */
1809 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1812 /* Reset any dirty pages we kept that page_flush didn't see */
1813 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1816 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1820 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1822 mdb_find_oldest(MDB_txn *txn)
1825 txnid_t mr, oldest = txn->mt_txnid - 1;
1826 if (txn->mt_env->me_txns) {
1827 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1828 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1839 /** Add a page to the txn's dirty list */
1841 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1844 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1846 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1847 insert = mdb_mid2l_append;
1849 insert = mdb_mid2l_insert;
1851 mid.mid = mp->mp_pgno;
1853 rc = insert(txn->mt_u.dirty_list, &mid);
1854 mdb_tassert(txn, rc == 0);
1855 txn->mt_dirty_room--;
1858 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1859 * me_pghead and mt_next_pgno.
1861 * If there are free pages available from older transactions, they
1862 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1863 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1864 * and move me_pglast to say which records were consumed. Only this
1865 * function can create me_pghead and move me_pglast/mt_next_pgno.
1866 * @param[in] mc cursor A cursor handle identifying the transaction and
1867 * database for which we are allocating.
1868 * @param[in] num the number of pages to allocate.
1869 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1870 * will always be satisfied by a single contiguous chunk of memory.
1871 * @return 0 on success, non-zero on failure.
1874 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1876 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1877 /* Get at most <Max_retries> more freeDB records once me_pghead
1878 * has enough pages. If not enough, use new pages from the map.
1879 * If <Paranoid> and mc is updating the freeDB, only get new
1880 * records if me_pghead is empty. Then the freelist cannot play
1881 * catch-up with itself by growing while trying to save it.
1883 enum { Paranoid = 1, Max_retries = 500 };
1885 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1887 int rc, retry = num * 20;
1888 MDB_txn *txn = mc->mc_txn;
1889 MDB_env *env = txn->mt_env;
1890 pgno_t pgno, *mop = env->me_pghead;
1891 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1893 txnid_t oldest = 0, last;
1897 /* If there are any loose pages, just use them */
1898 if (num == 1 && txn->mt_loose_pgs) {
1899 np = txn->mt_loose_pgs;
1900 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1907 /* If our dirty list is already full, we can't do anything */
1908 if (txn->mt_dirty_room == 0) {
1913 for (op = MDB_FIRST;; op = MDB_NEXT) {
1916 pgno_t *idl, old_id, new_id;
1918 /* Seek a big enough contiguous page range. Prefer
1919 * pages at the tail, just truncating the list.
1925 if (mop[i-n2] == pgno+n2)
1932 if (op == MDB_FIRST) { /* 1st iteration */
1933 /* Prepare to fetch more and coalesce */
1934 oldest = mdb_find_oldest(txn);
1935 last = env->me_pglast;
1936 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1939 key.mv_data = &last; /* will look up last+1 */
1940 key.mv_size = sizeof(last);
1942 if (Paranoid && mc->mc_dbi == FREE_DBI)
1945 if (Paranoid && retry < 0 && mop_len)
1949 /* Do not fetch more if the record will be too recent */
1952 rc = mdb_cursor_get(&m2, &key, NULL, op);
1954 if (rc == MDB_NOTFOUND)
1958 last = *(txnid_t*)key.mv_data;
1961 np = m2.mc_pg[m2.mc_top];
1962 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1963 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1966 idl = (MDB_ID *) data.mv_data;
1969 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
1974 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1976 mop = env->me_pghead;
1978 env->me_pglast = last;
1980 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1981 last, txn->mt_dbs[FREE_DBI].md_root, i));
1983 DPRINTF(("IDL %"Z"u", idl[k]));
1985 /* Merge in descending sorted order */
1988 mop[0] = (pgno_t)-1;
1992 for (; old_id < new_id; old_id = mop[--j])
1999 /* Use new pages from the map when nothing suitable in the freeDB */
2001 pgno = txn->mt_next_pgno;
2002 if (pgno + num >= env->me_maxpg) {
2003 DPUTS("DB size maxed out");
2009 if (env->me_flags & MDB_WRITEMAP) {
2010 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2012 if (!(np = mdb_page_malloc(txn, num))) {
2018 mop[0] = mop_len -= num;
2019 /* Move any stragglers down */
2020 for (j = i-num; j < mop_len; )
2021 mop[++j] = mop[++i];
2023 txn->mt_next_pgno = pgno + num;
2026 mdb_page_dirty(txn, np);
2032 txn->mt_flags |= MDB_TXN_ERROR;
2036 /** Copy the used portions of a non-overflow page.
2037 * @param[in] dst page to copy into
2038 * @param[in] src page to copy from
2039 * @param[in] psize size of a page
2042 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2044 enum { Align = sizeof(pgno_t) };
2045 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2047 /* If page isn't full, just copy the used portion. Adjust
2048 * alignment so memcpy may copy words instead of bytes.
2050 if ((unused &= -Align) && !IS_LEAF2(src)) {
2052 memcpy(dst, src, (lower + MP_HIBASE + (Align-1)) & -Align);
2053 memcpy((pgno_t *)((char *)dst+upper+MP_HIBASE), (pgno_t *)((char *)src+upper+MP_HIBASE),
2054 psize - upper - MP_HIBASE);
2056 memcpy(dst, src, psize - unused);
2060 /** Pull a page off the txn's spill list, if present.
2061 * If a page being referenced was spilled to disk in this txn, bring
2062 * it back and make it dirty/writable again.
2063 * @param[in] txn the transaction handle.
2064 * @param[in] mp the page being referenced. It must not be dirty.
2065 * @param[out] ret the writable page, if any. ret is unchanged if
2066 * mp wasn't spilled.
2069 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2071 MDB_env *env = txn->mt_env;
2074 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2076 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2077 if (!tx2->mt_spill_pgs)
2079 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2080 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2083 if (txn->mt_dirty_room == 0)
2084 return MDB_TXN_FULL;
2085 if (IS_OVERFLOW(mp))
2089 if (env->me_flags & MDB_WRITEMAP) {
2092 np = mdb_page_malloc(txn, num);
2096 memcpy(np, mp, num * env->me_psize);
2098 mdb_page_copy(np, mp, env->me_psize);
2101 /* If in current txn, this page is no longer spilled.
2102 * If it happens to be the last page, truncate the spill list.
2103 * Otherwise mark it as deleted by setting the LSB.
2105 if (x == txn->mt_spill_pgs[0])
2106 txn->mt_spill_pgs[0]--;
2108 txn->mt_spill_pgs[x] |= 1;
2109 } /* otherwise, if belonging to a parent txn, the
2110 * page remains spilled until child commits
2113 mdb_page_dirty(txn, np);
2114 np->mp_flags |= P_DIRTY;
2122 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2123 * @param[in] mc cursor pointing to the page to be touched
2124 * @return 0 on success, non-zero on failure.
2127 mdb_page_touch(MDB_cursor *mc)
2129 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2130 MDB_txn *txn = mc->mc_txn;
2131 MDB_cursor *m2, *m3;
2135 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2136 if (txn->mt_flags & MDB_TXN_SPILLS) {
2138 rc = mdb_page_unspill(txn, mp, &np);
2144 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2145 (rc = mdb_page_alloc(mc, 1, &np)))
2148 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2149 mp->mp_pgno, pgno));
2150 mdb_cassert(mc, mp->mp_pgno != pgno);
2151 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2152 /* Update the parent page, if any, to point to the new page */
2154 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2155 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2156 SETPGNO(node, pgno);
2158 mc->mc_db->md_root = pgno;
2160 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2161 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2163 /* If txn has a parent, make sure the page is in our
2167 unsigned x = mdb_mid2l_search(dl, pgno);
2168 if (x <= dl[0].mid && dl[x].mid == pgno) {
2169 if (mp != dl[x].mptr) { /* bad cursor? */
2170 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2171 txn->mt_flags |= MDB_TXN_ERROR;
2172 return MDB_CORRUPTED;
2177 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2179 np = mdb_page_malloc(txn, 1);
2184 rc = mdb_mid2l_insert(dl, &mid);
2185 mdb_cassert(mc, rc == 0);
2190 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2192 np->mp_flags |= P_DIRTY;
2195 /* Adjust cursors pointing to mp */
2196 mc->mc_pg[mc->mc_top] = np;
2197 m2 = txn->mt_cursors[mc->mc_dbi];
2198 if (mc->mc_flags & C_SUB) {
2199 for (; m2; m2=m2->mc_next) {
2200 m3 = &m2->mc_xcursor->mx_cursor;
2201 if (m3->mc_snum < mc->mc_snum) continue;
2202 if (m3->mc_pg[mc->mc_top] == mp)
2203 m3->mc_pg[mc->mc_top] = np;
2206 for (; m2; m2=m2->mc_next) {
2207 if (m2->mc_snum < mc->mc_snum) continue;
2208 if (m2->mc_pg[mc->mc_top] == mp) {
2209 m2->mc_pg[mc->mc_top] = np;
2210 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2212 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2214 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2215 if (!(leaf->mn_flags & F_SUBDATA))
2216 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2224 txn->mt_flags |= MDB_TXN_ERROR;
2229 mdb_env_sync(MDB_env *env, int force)
2232 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2233 if (env->me_flags & MDB_WRITEMAP) {
2234 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2235 ? MS_ASYNC : MS_SYNC;
2236 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2239 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2243 if (MDB_FDATASYNC(env->me_fd))
2250 /** Back up parent txn's cursors, then grab the originals for tracking */
2252 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2254 MDB_cursor *mc, *bk;
2259 for (i = src->mt_numdbs; --i >= 0; ) {
2260 if ((mc = src->mt_cursors[i]) != NULL) {
2261 size = sizeof(MDB_cursor);
2263 size += sizeof(MDB_xcursor);
2264 for (; mc; mc = bk->mc_next) {
2270 mc->mc_db = &dst->mt_dbs[i];
2271 /* Kill pointers into src - and dst to reduce abuse: The
2272 * user may not use mc until dst ends. Otherwise we'd...
2274 mc->mc_txn = NULL; /* ...set this to dst */
2275 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2276 if ((mx = mc->mc_xcursor) != NULL) {
2277 *(MDB_xcursor *)(bk+1) = *mx;
2278 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2280 mc->mc_next = dst->mt_cursors[i];
2281 dst->mt_cursors[i] = mc;
2288 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2289 * @param[in] txn the transaction handle.
2290 * @param[in] merge true to keep changes to parent cursors, false to revert.
2291 * @return 0 on success, non-zero on failure.
2294 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2296 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2300 for (i = txn->mt_numdbs; --i >= 0; ) {
2301 for (mc = cursors[i]; mc; mc = next) {
2303 if ((bk = mc->mc_backup) != NULL) {
2305 /* Commit changes to parent txn */
2306 mc->mc_next = bk->mc_next;
2307 mc->mc_backup = bk->mc_backup;
2308 mc->mc_txn = bk->mc_txn;
2309 mc->mc_db = bk->mc_db;
2310 mc->mc_dbflag = bk->mc_dbflag;
2311 if ((mx = mc->mc_xcursor) != NULL)
2312 mx->mx_cursor.mc_txn = bk->mc_txn;
2314 /* Abort nested txn */
2316 if ((mx = mc->mc_xcursor) != NULL)
2317 *mx = *(MDB_xcursor *)(bk+1);
2321 /* Only malloced cursors are permanently tracked. */
2329 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2332 mdb_txn_reset0(MDB_txn *txn, const char *act);
2334 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2340 Pidset = F_SETLK, Pidcheck = F_GETLK
2344 /** Set or check a pid lock. Set returns 0 on success.
2345 * Check returns 0 if the process is certainly dead, nonzero if it may
2346 * be alive (the lock exists or an error happened so we do not know).
2348 * On Windows Pidset is a no-op, we merely check for the existence
2349 * of the process with the given pid. On POSIX we use a single byte
2350 * lock on the lockfile, set at an offset equal to the pid.
2353 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2355 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2358 if (op == Pidcheck) {
2359 h = OpenProcess(env->me_pidquery, FALSE, pid);
2360 /* No documented "no such process" code, but other program use this: */
2362 return ErrCode() != ERROR_INVALID_PARAMETER;
2363 /* A process exists until all handles to it close. Has it exited? */
2364 ret = WaitForSingleObject(h, 0) != 0;
2371 struct flock lock_info;
2372 memset(&lock_info, 0, sizeof(lock_info));
2373 lock_info.l_type = F_WRLCK;
2374 lock_info.l_whence = SEEK_SET;
2375 lock_info.l_start = pid;
2376 lock_info.l_len = 1;
2377 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2378 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2380 } else if ((rc = ErrCode()) == EINTR) {
2388 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2389 * @param[in] txn the transaction handle to initialize
2390 * @return 0 on success, non-zero on failure.
2393 mdb_txn_renew0(MDB_txn *txn)
2395 MDB_env *env = txn->mt_env;
2396 MDB_txninfo *ti = env->me_txns;
2400 int rc, new_notls = 0;
2403 txn->mt_numdbs = env->me_numdbs;
2404 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2406 if (txn->mt_flags & MDB_TXN_RDONLY) {
2408 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2409 txn->mt_txnid = meta->mm_txnid;
2410 txn->mt_u.reader = NULL;
2412 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2413 pthread_getspecific(env->me_txkey);
2415 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2416 return MDB_BAD_RSLOT;
2418 MDB_PID_T pid = env->me_pid;
2419 MDB_THR_T tid = pthread_self();
2421 if (!env->me_live_reader) {
2422 rc = mdb_reader_pid(env, Pidset, pid);
2425 env->me_live_reader = 1;
2429 nr = ti->mti_numreaders;
2430 for (i=0; i<nr; i++)
2431 if (ti->mti_readers[i].mr_pid == 0)
2433 if (i == env->me_maxreaders) {
2434 UNLOCK_MUTEX_R(env);
2435 return MDB_READERS_FULL;
2437 ti->mti_readers[i].mr_pid = pid;
2438 ti->mti_readers[i].mr_tid = tid;
2440 ti->mti_numreaders = ++nr;
2441 /* Save numreaders for un-mutexed mdb_env_close() */
2442 env->me_numreaders = nr;
2443 UNLOCK_MUTEX_R(env);
2445 r = &ti->mti_readers[i];
2446 new_notls = (env->me_flags & MDB_NOTLS);
2447 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2452 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2453 txn->mt_u.reader = r;
2454 meta = env->me_metas[txn->mt_txnid & 1];
2460 txn->mt_txnid = ti->mti_txnid;
2461 meta = env->me_metas[txn->mt_txnid & 1];
2463 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2464 txn->mt_txnid = meta->mm_txnid;
2468 if (txn->mt_txnid == mdb_debug_start)
2471 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2472 txn->mt_u.dirty_list = env->me_dirty_list;
2473 txn->mt_u.dirty_list[0].mid = 0;
2474 txn->mt_free_pgs = env->me_free_pgs;
2475 txn->mt_free_pgs[0] = 0;
2476 txn->mt_spill_pgs = NULL;
2480 /* Copy the DB info and flags */
2481 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2483 /* Moved to here to avoid a data race in read TXNs */
2484 txn->mt_next_pgno = meta->mm_last_pg+1;
2486 for (i=2; i<txn->mt_numdbs; i++) {
2487 x = env->me_dbflags[i];
2488 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2489 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2491 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2493 if (env->me_maxpg < txn->mt_next_pgno) {
2494 mdb_txn_reset0(txn, "renew0-mapfail");
2496 txn->mt_u.reader->mr_pid = 0;
2497 txn->mt_u.reader = NULL;
2499 return MDB_MAP_RESIZED;
2506 mdb_txn_renew(MDB_txn *txn)
2510 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2513 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2514 DPUTS("environment had fatal error, must shutdown!");
2518 rc = mdb_txn_renew0(txn);
2519 if (rc == MDB_SUCCESS) {
2520 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2521 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2522 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2528 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2532 int rc, size, tsize = sizeof(MDB_txn);
2534 if (env->me_flags & MDB_FATAL_ERROR) {
2535 DPUTS("environment had fatal error, must shutdown!");
2538 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2541 /* Nested transactions: Max 1 child, write txns only, no writemap */
2542 if (parent->mt_child ||
2543 (flags & MDB_RDONLY) ||
2544 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2545 (env->me_flags & MDB_WRITEMAP))
2547 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2549 tsize = sizeof(MDB_ntxn);
2551 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2552 if (!(flags & MDB_RDONLY))
2553 size += env->me_maxdbs * sizeof(MDB_cursor *);
2555 if ((txn = calloc(1, size)) == NULL) {
2556 DPRINTF(("calloc: %s", strerror(ErrCode())));
2559 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2560 if (flags & MDB_RDONLY) {
2561 txn->mt_flags |= MDB_TXN_RDONLY;
2562 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2564 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2565 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2571 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2572 if (!txn->mt_u.dirty_list ||
2573 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2575 free(txn->mt_u.dirty_list);
2579 txn->mt_txnid = parent->mt_txnid;
2580 txn->mt_dirty_room = parent->mt_dirty_room;
2581 txn->mt_u.dirty_list[0].mid = 0;
2582 txn->mt_spill_pgs = NULL;
2583 txn->mt_next_pgno = parent->mt_next_pgno;
2584 parent->mt_child = txn;
2585 txn->mt_parent = parent;
2586 txn->mt_numdbs = parent->mt_numdbs;
2587 txn->mt_flags = parent->mt_flags;
2588 txn->mt_dbxs = parent->mt_dbxs;
2589 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2590 /* Copy parent's mt_dbflags, but clear DB_NEW */
2591 for (i=0; i<txn->mt_numdbs; i++)
2592 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2594 ntxn = (MDB_ntxn *)txn;
2595 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2596 if (env->me_pghead) {
2597 size = MDB_IDL_SIZEOF(env->me_pghead);
2598 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2600 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2605 rc = mdb_cursor_shadow(parent, txn);
2607 mdb_txn_reset0(txn, "beginchild-fail");
2609 rc = mdb_txn_renew0(txn);
2615 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2616 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2617 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2624 mdb_txn_env(MDB_txn *txn)
2626 if(!txn) return NULL;
2630 /** Export or close DBI handles opened in this txn. */
2632 mdb_dbis_update(MDB_txn *txn, int keep)
2635 MDB_dbi n = txn->mt_numdbs;
2636 MDB_env *env = txn->mt_env;
2637 unsigned char *tdbflags = txn->mt_dbflags;
2639 for (i = n; --i >= 2;) {
2640 if (tdbflags[i] & DB_NEW) {
2642 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2644 char *ptr = env->me_dbxs[i].md_name.mv_data;
2645 env->me_dbxs[i].md_name.mv_data = NULL;
2646 env->me_dbxs[i].md_name.mv_size = 0;
2647 env->me_dbflags[i] = 0;
2652 if (keep && env->me_numdbs < n)
2656 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2657 * May be called twice for readonly txns: First reset it, then abort.
2658 * @param[in] txn the transaction handle to reset
2659 * @param[in] act why the transaction is being reset
2662 mdb_txn_reset0(MDB_txn *txn, const char *act)
2664 MDB_env *env = txn->mt_env;
2666 /* Close any DBI handles opened in this txn */
2667 mdb_dbis_update(txn, 0);
2669 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2670 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2671 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2673 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2674 if (txn->mt_u.reader) {
2675 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2676 if (!(env->me_flags & MDB_NOTLS))
2677 txn->mt_u.reader = NULL; /* txn does not own reader */
2679 txn->mt_numdbs = 0; /* close nothing if called again */
2680 txn->mt_dbxs = NULL; /* mark txn as reset */
2682 mdb_cursors_close(txn, 0);
2684 if (!(env->me_flags & MDB_WRITEMAP)) {
2685 mdb_dlist_free(txn);
2687 mdb_midl_free(env->me_pghead);
2689 if (txn->mt_parent) {
2690 txn->mt_parent->mt_child = NULL;
2691 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2692 mdb_midl_free(txn->mt_free_pgs);
2693 mdb_midl_free(txn->mt_spill_pgs);
2694 free(txn->mt_u.dirty_list);
2698 if (mdb_midl_shrink(&txn->mt_free_pgs))
2699 env->me_free_pgs = txn->mt_free_pgs;
2700 env->me_pghead = NULL;
2704 /* The writer mutex was locked in mdb_txn_begin. */
2706 UNLOCK_MUTEX_W(env);
2711 mdb_txn_reset(MDB_txn *txn)
2716 /* This call is only valid for read-only txns */
2717 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2720 mdb_txn_reset0(txn, "reset");
2724 mdb_txn_abort(MDB_txn *txn)
2730 mdb_txn_abort(txn->mt_child);
2732 mdb_txn_reset0(txn, "abort");
2733 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2734 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2735 txn->mt_u.reader->mr_pid = 0;
2740 /** Save the freelist as of this transaction to the freeDB.
2741 * This changes the freelist. Keep trying until it stabilizes.
2744 mdb_freelist_save(MDB_txn *txn)
2746 /* env->me_pghead[] can grow and shrink during this call.
2747 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2748 * Page numbers cannot disappear from txn->mt_free_pgs[].
2751 MDB_env *env = txn->mt_env;
2752 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2753 txnid_t pglast = 0, head_id = 0;
2754 pgno_t freecnt = 0, *free_pgs, *mop;
2755 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2757 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2759 if (env->me_pghead) {
2760 /* Make sure first page of freeDB is touched and on freelist */
2761 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2762 if (rc && rc != MDB_NOTFOUND)
2766 /* Dispose of loose pages. Usually they will have all
2767 * been used up by the time we get here.
2769 if (txn->mt_loose_pgs) {
2770 MDB_page *mp = txn->mt_loose_pgs;
2771 /* Just return them to freeDB */
2772 if (env->me_pghead) {
2774 mop = env->me_pghead;
2775 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2776 pgno_t pg = mp->mp_pgno;
2778 for (i = mop[0]; i && mop[i] < pg; i--)
2784 /* Oh well, they were wasted. Put on freelist */
2785 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2786 mdb_midl_append(&txn->mt_free_pgs, mp->mp_pgno);
2789 txn->mt_loose_pgs = NULL;
2792 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2793 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2794 ? SSIZE_MAX : maxfree_1pg;
2797 /* Come back here after each Put() in case freelist changed */
2802 /* If using records from freeDB which we have not yet
2803 * deleted, delete them and any we reserved for me_pghead.
2805 while (pglast < env->me_pglast) {
2806 rc = mdb_cursor_first(&mc, &key, NULL);
2809 pglast = head_id = *(txnid_t *)key.mv_data;
2810 total_room = head_room = 0;
2811 mdb_tassert(txn, pglast <= env->me_pglast);
2812 rc = mdb_cursor_del(&mc, 0);
2817 /* Save the IDL of pages freed by this txn, to a single record */
2818 if (freecnt < txn->mt_free_pgs[0]) {
2820 /* Make sure last page of freeDB is touched and on freelist */
2821 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2822 if (rc && rc != MDB_NOTFOUND)
2825 free_pgs = txn->mt_free_pgs;
2826 /* Write to last page of freeDB */
2827 key.mv_size = sizeof(txn->mt_txnid);
2828 key.mv_data = &txn->mt_txnid;
2830 freecnt = free_pgs[0];
2831 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2832 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2835 /* Retry if mt_free_pgs[] grew during the Put() */
2836 free_pgs = txn->mt_free_pgs;
2837 } while (freecnt < free_pgs[0]);
2838 mdb_midl_sort(free_pgs);
2839 memcpy(data.mv_data, free_pgs, data.mv_size);
2842 unsigned int i = free_pgs[0];
2843 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2844 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2846 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2852 mop = env->me_pghead;
2853 mop_len = mop ? mop[0] : 0;
2855 /* Reserve records for me_pghead[]. Split it if multi-page,
2856 * to avoid searching freeDB for a page range. Use keys in
2857 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2859 if (total_room >= mop_len) {
2860 if (total_room == mop_len || --more < 0)
2862 } else if (head_room >= maxfree_1pg && head_id > 1) {
2863 /* Keep current record (overflow page), add a new one */
2867 /* (Re)write {key = head_id, IDL length = head_room} */
2868 total_room -= head_room;
2869 head_room = mop_len - total_room;
2870 if (head_room > maxfree_1pg && head_id > 1) {
2871 /* Overflow multi-page for part of me_pghead */
2872 head_room /= head_id; /* amortize page sizes */
2873 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2874 } else if (head_room < 0) {
2875 /* Rare case, not bothering to delete this record */
2878 key.mv_size = sizeof(head_id);
2879 key.mv_data = &head_id;
2880 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2881 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2884 /* IDL is initially empty, zero out at least the length */
2885 pgs = (pgno_t *)data.mv_data;
2886 j = head_room > clean_limit ? head_room : 0;
2890 total_room += head_room;
2893 /* Fill in the reserved me_pghead records */
2899 rc = mdb_cursor_first(&mc, &key, &data);
2900 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2901 txnid_t id = *(txnid_t *)key.mv_data;
2902 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2905 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2907 if (len > mop_len) {
2909 data.mv_size = (len + 1) * sizeof(MDB_ID);
2911 data.mv_data = mop -= len;
2914 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2916 if (rc || !(mop_len -= len))
2923 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2924 * @param[in] txn the transaction that's being committed
2925 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2926 * @return 0 on success, non-zero on failure.
2929 mdb_page_flush(MDB_txn *txn, int keep)
2931 MDB_env *env = txn->mt_env;
2932 MDB_ID2L dl = txn->mt_u.dirty_list;
2933 unsigned psize = env->me_psize, j;
2934 int i, pagecount = dl[0].mid, rc;
2935 size_t size = 0, pos = 0;
2937 MDB_page *dp = NULL;
2941 struct iovec iov[MDB_COMMIT_PAGES];
2942 ssize_t wpos = 0, wsize = 0, wres;
2943 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2949 if (env->me_flags & MDB_WRITEMAP) {
2950 /* Clear dirty flags */
2951 while (++i <= pagecount) {
2953 /* Don't flush this page yet */
2954 if (dp->mp_flags & P_KEEP) {
2955 dp->mp_flags ^= P_KEEP;
2959 dp->mp_flags &= ~P_DIRTY;
2964 /* Write the pages */
2966 if (++i <= pagecount) {
2968 /* Don't flush this page yet */
2969 if (dp->mp_flags & P_KEEP) {
2970 dp->mp_flags ^= P_KEEP;
2975 /* clear dirty flag */
2976 dp->mp_flags &= ~P_DIRTY;
2979 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2984 /* Windows actually supports scatter/gather I/O, but only on
2985 * unbuffered file handles. Since we're relying on the OS page
2986 * cache for all our data, that's self-defeating. So we just
2987 * write pages one at a time. We use the ov structure to set
2988 * the write offset, to at least save the overhead of a Seek
2991 DPRINTF(("committing page %"Z"u", pgno));
2992 memset(&ov, 0, sizeof(ov));
2993 ov.Offset = pos & 0xffffffff;
2994 ov.OffsetHigh = pos >> 16 >> 16;
2995 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2997 DPRINTF(("WriteFile: %d", rc));
3001 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3002 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3004 /* Write previous page(s) */
3005 #ifdef MDB_USE_PWRITEV
3006 wres = pwritev(env->me_fd, iov, n, wpos);
3009 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3011 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3013 DPRINTF(("lseek: %s", strerror(rc)));
3016 wres = writev(env->me_fd, iov, n);
3019 if (wres != wsize) {
3022 DPRINTF(("Write error: %s", strerror(rc)));
3024 rc = EIO; /* TODO: Use which error code? */
3025 DPUTS("short write, filesystem full?");
3036 DPRINTF(("committing page %"Z"u", pgno));
3037 next_pos = pos + size;
3038 iov[n].iov_len = size;
3039 iov[n].iov_base = (char *)dp;
3045 for (i = keep; ++i <= pagecount; ) {
3047 /* This is a page we skipped above */
3050 dl[j].mid = dp->mp_pgno;
3053 mdb_dpage_free(env, dp);
3058 txn->mt_dirty_room += i - j;
3064 mdb_txn_commit(MDB_txn *txn)
3070 if (txn == NULL || txn->mt_env == NULL)
3073 if (txn->mt_child) {
3074 rc = mdb_txn_commit(txn->mt_child);
3075 txn->mt_child = NULL;
3082 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3083 mdb_dbis_update(txn, 1);
3084 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3089 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3090 DPUTS("error flag is set, can't commit");
3092 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3097 if (txn->mt_parent) {
3098 MDB_txn *parent = txn->mt_parent;
3101 unsigned x, y, len, ps_len;
3103 /* Append our free list to parent's */
3104 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3107 mdb_midl_free(txn->mt_free_pgs);
3108 /* Failures after this must either undo the changes
3109 * to the parent or set MDB_TXN_ERROR in the parent.
3112 parent->mt_next_pgno = txn->mt_next_pgno;
3113 parent->mt_flags = txn->mt_flags;
3115 /* Merge our cursors into parent's and close them */
3116 mdb_cursors_close(txn, 1);
3118 /* Update parent's DB table. */
3119 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3120 parent->mt_numdbs = txn->mt_numdbs;
3121 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3122 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3123 for (i=2; i<txn->mt_numdbs; i++) {
3124 /* preserve parent's DB_NEW status */
3125 x = parent->mt_dbflags[i] & DB_NEW;
3126 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3129 dst = parent->mt_u.dirty_list;
3130 src = txn->mt_u.dirty_list;
3131 /* Remove anything in our dirty list from parent's spill list */
3132 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3134 pspill[0] = (pgno_t)-1;
3135 /* Mark our dirty pages as deleted in parent spill list */
3136 for (i=0, len=src[0].mid; ++i <= len; ) {
3137 MDB_ID pn = src[i].mid << 1;
3138 while (pn > pspill[x])
3140 if (pn == pspill[x]) {
3145 /* Squash deleted pagenums if we deleted any */
3146 for (x=y; ++x <= ps_len; )
3147 if (!(pspill[x] & 1))
3148 pspill[++y] = pspill[x];
3152 /* Find len = length of merging our dirty list with parent's */
3154 dst[0].mid = 0; /* simplify loops */
3155 if (parent->mt_parent) {
3156 len = x + src[0].mid;
3157 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3158 for (i = x; y && i; y--) {
3159 pgno_t yp = src[y].mid;
3160 while (yp < dst[i].mid)
3162 if (yp == dst[i].mid) {
3167 } else { /* Simplify the above for single-ancestor case */
3168 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3170 /* Merge our dirty list with parent's */
3172 for (i = len; y; dst[i--] = src[y--]) {
3173 pgno_t yp = src[y].mid;
3174 while (yp < dst[x].mid)
3175 dst[i--] = dst[x--];
3176 if (yp == dst[x].mid)
3177 free(dst[x--].mptr);
3179 mdb_tassert(txn, i == x);
3181 free(txn->mt_u.dirty_list);
3182 parent->mt_dirty_room = txn->mt_dirty_room;
3183 if (txn->mt_spill_pgs) {
3184 if (parent->mt_spill_pgs) {
3185 /* TODO: Prevent failure here, so parent does not fail */
3186 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3188 parent->mt_flags |= MDB_TXN_ERROR;
3189 mdb_midl_free(txn->mt_spill_pgs);
3190 mdb_midl_sort(parent->mt_spill_pgs);
3192 parent->mt_spill_pgs = txn->mt_spill_pgs;
3196 parent->mt_child = NULL;
3197 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3202 if (txn != env->me_txn) {
3203 DPUTS("attempt to commit unknown transaction");
3208 mdb_cursors_close(txn, 0);
3210 if (!txn->mt_u.dirty_list[0].mid &&
3211 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3214 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3215 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3217 /* Update DB root pointers */
3218 if (txn->mt_numdbs > 2) {
3222 data.mv_size = sizeof(MDB_db);
3224 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3225 for (i = 2; i < txn->mt_numdbs; i++) {
3226 if (txn->mt_dbflags[i] & DB_DIRTY) {
3227 data.mv_data = &txn->mt_dbs[i];
3228 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3235 rc = mdb_freelist_save(txn);
3239 mdb_midl_free(env->me_pghead);
3240 env->me_pghead = NULL;
3241 if (mdb_midl_shrink(&txn->mt_free_pgs))
3242 env->me_free_pgs = txn->mt_free_pgs;
3248 if ((rc = mdb_page_flush(txn, 0)) ||
3249 (rc = mdb_env_sync(env, 0)) ||
3250 (rc = mdb_env_write_meta(txn)))
3256 mdb_dbis_update(txn, 1);
3259 UNLOCK_MUTEX_W(env);
3269 /** Read the environment parameters of a DB environment before
3270 * mapping it into memory.
3271 * @param[in] env the environment handle
3272 * @param[out] meta address of where to store the meta information
3273 * @return 0 on success, non-zero on failure.
3276 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3282 enum { Size = sizeof(pbuf) };
3284 /* We don't know the page size yet, so use a minimum value.
3285 * Read both meta pages so we can use the latest one.
3288 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3292 memset(&ov, 0, sizeof(ov));
3294 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3295 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3298 rc = pread(env->me_fd, &pbuf, Size, off);
3301 if (rc == 0 && off == 0)
3303 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3304 DPRINTF(("read: %s", mdb_strerror(rc)));
3308 p = (MDB_page *)&pbuf;
3310 if (!F_ISSET(p->mp_flags, P_META)) {
3311 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3316 if (m->mm_magic != MDB_MAGIC) {
3317 DPUTS("meta has invalid magic");
3321 if (m->mm_version != MDB_DATA_VERSION) {
3322 DPRINTF(("database is version %u, expected version %u",
3323 m->mm_version, MDB_DATA_VERSION));
3324 return MDB_VERSION_MISMATCH;
3327 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3334 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3336 meta->mm_magic = MDB_MAGIC;
3337 meta->mm_version = MDB_DATA_VERSION;
3338 meta->mm_mapsize = env->me_mapsize;
3339 meta->mm_psize = env->me_psize;
3340 meta->mm_last_pg = 1;
3341 meta->mm_flags = env->me_flags & 0xffff;
3342 meta->mm_flags |= MDB_INTEGERKEY;
3343 meta->mm_dbs[0].md_root = P_INVALID;
3344 meta->mm_dbs[1].md_root = P_INVALID;
3347 /** Write the environment parameters of a freshly created DB environment.
3348 * @param[in] env the environment handle
3349 * @param[out] meta address of where to store the meta information
3350 * @return 0 on success, non-zero on failure.
3353 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3361 memset(&ov, 0, sizeof(ov));
3362 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3364 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3367 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3368 len = pwrite(fd, ptr, size, pos); \
3369 rc = (len >= 0); } while(0)
3372 DPUTS("writing new meta page");
3374 psize = env->me_psize;
3376 mdb_env_init_meta0(env, meta);
3378 p = calloc(2, psize);
3380 p->mp_flags = P_META;
3381 *(MDB_meta *)METADATA(p) = *meta;
3383 q = (MDB_page *)((char *)p + psize);
3385 q->mp_flags = P_META;
3386 *(MDB_meta *)METADATA(q) = *meta;
3388 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3391 else if ((unsigned) len == psize * 2)
3399 /** Update the environment info to commit a transaction.
3400 * @param[in] txn the transaction that's being committed
3401 * @return 0 on success, non-zero on failure.
3404 mdb_env_write_meta(MDB_txn *txn)
3407 MDB_meta meta, metab, *mp;
3409 int rc, len, toggle;
3418 toggle = txn->mt_txnid & 1;
3419 DPRINTF(("writing meta page %d for root page %"Z"u",
3420 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3423 mp = env->me_metas[toggle];
3425 if (env->me_flags & MDB_WRITEMAP) {
3426 /* Persist any increases of mapsize config */
3427 if (env->me_mapsize > mp->mm_mapsize)
3428 mp->mm_mapsize = env->me_mapsize;
3429 mp->mm_dbs[0] = txn->mt_dbs[0];
3430 mp->mm_dbs[1] = txn->mt_dbs[1];
3431 mp->mm_last_pg = txn->mt_next_pgno - 1;
3432 mp->mm_txnid = txn->mt_txnid;
3433 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3434 unsigned meta_size = env->me_psize;
3435 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3438 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3439 if (meta_size < env->me_os_psize)
3440 meta_size += meta_size;
3445 if (MDB_MSYNC(ptr, meta_size, rc)) {
3452 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3453 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3455 ptr = (char *)&meta;
3456 if (env->me_mapsize > mp->mm_mapsize) {
3457 /* Persist any increases of mapsize config */
3458 meta.mm_mapsize = env->me_mapsize;
3459 off = offsetof(MDB_meta, mm_mapsize);
3461 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3463 len = sizeof(MDB_meta) - off;
3466 meta.mm_dbs[0] = txn->mt_dbs[0];
3467 meta.mm_dbs[1] = txn->mt_dbs[1];
3468 meta.mm_last_pg = txn->mt_next_pgno - 1;
3469 meta.mm_txnid = txn->mt_txnid;
3472 off += env->me_psize;
3475 /* Write to the SYNC fd */
3476 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3477 env->me_fd : env->me_mfd;
3480 memset(&ov, 0, sizeof(ov));
3482 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3486 rc = pwrite(mfd, ptr, len, off);
3489 rc = rc < 0 ? ErrCode() : EIO;
3490 DPUTS("write failed, disk error?");
3491 /* On a failure, the pagecache still contains the new data.
3492 * Write some old data back, to prevent it from being used.
3493 * Use the non-SYNC fd; we know it will fail anyway.
3495 meta.mm_last_pg = metab.mm_last_pg;
3496 meta.mm_txnid = metab.mm_txnid;
3498 memset(&ov, 0, sizeof(ov));
3500 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3502 r2 = pwrite(env->me_fd, ptr, len, off);
3503 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3506 env->me_flags |= MDB_FATAL_ERROR;
3510 /* Memory ordering issues are irrelevant; since the entire writer
3511 * is wrapped by wmutex, all of these changes will become visible
3512 * after the wmutex is unlocked. Since the DB is multi-version,
3513 * readers will get consistent data regardless of how fresh or
3514 * how stale their view of these values is.
3517 env->me_txns->mti_txnid = txn->mt_txnid;
3522 /** Check both meta pages to see which one is newer.
3523 * @param[in] env the environment handle
3524 * @return meta toggle (0 or 1).
3527 mdb_env_pick_meta(const MDB_env *env)
3529 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3533 mdb_env_create(MDB_env **env)
3537 e = calloc(1, sizeof(MDB_env));
3541 e->me_maxreaders = DEFAULT_READERS;
3542 e->me_maxdbs = e->me_numdbs = 2;
3543 e->me_fd = INVALID_HANDLE_VALUE;
3544 e->me_lfd = INVALID_HANDLE_VALUE;
3545 e->me_mfd = INVALID_HANDLE_VALUE;
3546 #ifdef MDB_USE_POSIX_SEM
3547 e->me_rmutex = SEM_FAILED;
3548 e->me_wmutex = SEM_FAILED;
3550 e->me_pid = getpid();
3551 GET_PAGESIZE(e->me_os_psize);
3552 VGMEMP_CREATE(e,0,0);
3558 mdb_env_map(MDB_env *env, void *addr, int newsize)
3561 unsigned int flags = env->me_flags;
3565 LONG sizelo, sizehi;
3568 if (flags & MDB_RDONLY) {
3573 msize = env->me_mapsize;
3574 sizelo = msize & 0xffffffff;
3575 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3578 /* Windows won't create mappings for zero length files.
3579 * Just allocate the maxsize right now.
3582 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3583 || !SetEndOfFile(env->me_fd)
3584 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3587 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3588 PAGE_READWRITE : PAGE_READONLY,
3589 sizehi, sizelo, NULL);
3592 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3593 FILE_MAP_WRITE : FILE_MAP_READ,
3595 rc = env->me_map ? 0 : ErrCode();
3600 int prot = PROT_READ;
3601 if (flags & MDB_WRITEMAP) {
3603 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3606 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3608 if (env->me_map == MAP_FAILED) {
3613 if (flags & MDB_NORDAHEAD) {
3614 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3616 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3618 #ifdef POSIX_MADV_RANDOM
3619 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3620 #endif /* POSIX_MADV_RANDOM */
3621 #endif /* MADV_RANDOM */
3625 /* Can happen because the address argument to mmap() is just a
3626 * hint. mmap() can pick another, e.g. if the range is in use.
3627 * The MAP_FIXED flag would prevent that, but then mmap could
3628 * instead unmap existing pages to make room for the new map.
3630 if (addr && env->me_map != addr)
3631 return EBUSY; /* TODO: Make a new MDB_* error code? */
3633 p = (MDB_page *)env->me_map;
3634 env->me_metas[0] = METADATA(p);
3635 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3641 mdb_env_set_mapsize(MDB_env *env, size_t size)
3643 /* If env is already open, caller is responsible for making
3644 * sure there are no active txns.
3652 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3653 else if (size < env->me_mapsize) {
3654 /* If the configured size is smaller, make sure it's
3655 * still big enough. Silently round up to minimum if not.
3657 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3661 munmap(env->me_map, env->me_mapsize);
3662 env->me_mapsize = size;
3663 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3664 rc = mdb_env_map(env, old, 1);
3668 env->me_mapsize = size;
3670 env->me_maxpg = env->me_mapsize / env->me_psize;
3675 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3679 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3684 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3686 if (env->me_map || readers < 1)
3688 env->me_maxreaders = readers;
3693 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3695 if (!env || !readers)
3697 *readers = env->me_maxreaders;
3701 /** Further setup required for opening an LMDB environment
3704 mdb_env_open2(MDB_env *env)
3706 unsigned int flags = env->me_flags;
3707 int i, newenv = 0, rc;
3711 /* See if we should use QueryLimited */
3713 if ((rc & 0xff) > 5)
3714 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3716 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3719 memset(&meta, 0, sizeof(meta));
3721 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3724 DPUTS("new mdbenv");
3726 env->me_psize = env->me_os_psize;
3727 if (env->me_psize > MAX_PAGESIZE)
3728 env->me_psize = MAX_PAGESIZE;
3730 env->me_psize = meta.mm_psize;
3733 /* Was a mapsize configured? */
3734 if (!env->me_mapsize) {
3735 /* If this is a new environment, take the default,
3736 * else use the size recorded in the existing env.
3738 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3739 } else if (env->me_mapsize < meta.mm_mapsize) {
3740 /* If the configured size is smaller, make sure it's
3741 * still big enough. Silently round up to minimum if not.
3743 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3744 if (env->me_mapsize < minsize)
3745 env->me_mapsize = minsize;
3748 rc = mdb_env_map(env, meta.mm_address, newenv || env->me_mapsize != meta.mm_mapsize);
3753 if (flags & MDB_FIXEDMAP)
3754 meta.mm_address = env->me_map;
3755 i = mdb_env_init_meta(env, &meta);
3756 if (i != MDB_SUCCESS) {
3761 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3762 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3764 #if !(MDB_MAXKEYSIZE)
3765 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3767 env->me_maxpg = env->me_mapsize / env->me_psize;
3771 int toggle = mdb_env_pick_meta(env);
3772 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3774 DPRINTF(("opened database version %u, pagesize %u",
3775 env->me_metas[0]->mm_version, env->me_psize));
3776 DPRINTF(("using meta page %d", toggle));
3777 DPRINTF(("depth: %u", db->md_depth));
3778 DPRINTF(("entries: %"Z"u", db->md_entries));
3779 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3780 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3781 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3782 DPRINTF(("root: %"Z"u", db->md_root));
3790 /** Release a reader thread's slot in the reader lock table.
3791 * This function is called automatically when a thread exits.
3792 * @param[in] ptr This points to the slot in the reader lock table.
3795 mdb_env_reader_dest(void *ptr)
3797 MDB_reader *reader = ptr;
3803 /** Junk for arranging thread-specific callbacks on Windows. This is
3804 * necessarily platform and compiler-specific. Windows supports up
3805 * to 1088 keys. Let's assume nobody opens more than 64 environments
3806 * in a single process, for now. They can override this if needed.
3808 #ifndef MAX_TLS_KEYS
3809 #define MAX_TLS_KEYS 64
3811 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3812 static int mdb_tls_nkeys;
3814 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3818 case DLL_PROCESS_ATTACH: break;
3819 case DLL_THREAD_ATTACH: break;
3820 case DLL_THREAD_DETACH:
3821 for (i=0; i<mdb_tls_nkeys; i++) {
3822 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3824 mdb_env_reader_dest(r);
3828 case DLL_PROCESS_DETACH: break;
3833 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3835 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3839 /* Force some symbol references.
3840 * _tls_used forces the linker to create the TLS directory if not already done
3841 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3843 #pragma comment(linker, "/INCLUDE:_tls_used")
3844 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3845 #pragma const_seg(".CRT$XLB")
3846 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3847 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3850 #pragma comment(linker, "/INCLUDE:__tls_used")
3851 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3852 #pragma data_seg(".CRT$XLB")
3853 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3855 #endif /* WIN 32/64 */
3856 #endif /* !__GNUC__ */
3859 /** Downgrade the exclusive lock on the region back to shared */
3861 mdb_env_share_locks(MDB_env *env, int *excl)
3863 int rc = 0, toggle = mdb_env_pick_meta(env);
3865 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3870 /* First acquire a shared lock. The Unlock will
3871 * then release the existing exclusive lock.
3873 memset(&ov, 0, sizeof(ov));
3874 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3877 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3883 struct flock lock_info;
3884 /* The shared lock replaces the existing lock */
3885 memset((void *)&lock_info, 0, sizeof(lock_info));
3886 lock_info.l_type = F_RDLCK;
3887 lock_info.l_whence = SEEK_SET;
3888 lock_info.l_start = 0;
3889 lock_info.l_len = 1;
3890 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3891 (rc = ErrCode()) == EINTR) ;
3892 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3899 /** Try to get exlusive lock, otherwise shared.
3900 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3903 mdb_env_excl_lock(MDB_env *env, int *excl)
3907 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3911 memset(&ov, 0, sizeof(ov));
3912 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3919 struct flock lock_info;
3920 memset((void *)&lock_info, 0, sizeof(lock_info));
3921 lock_info.l_type = F_WRLCK;
3922 lock_info.l_whence = SEEK_SET;
3923 lock_info.l_start = 0;
3924 lock_info.l_len = 1;
3925 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3926 (rc = ErrCode()) == EINTR) ;
3930 # ifdef MDB_USE_POSIX_SEM
3931 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3934 lock_info.l_type = F_RDLCK;
3935 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3936 (rc = ErrCode()) == EINTR) ;
3946 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3948 * @(#) $Revision: 5.1 $
3949 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3950 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3952 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3956 * Please do not copyright this code. This code is in the public domain.
3958 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3959 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3960 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3961 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3962 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3963 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3964 * PERFORMANCE OF THIS SOFTWARE.
3967 * chongo <Landon Curt Noll> /\oo/\
3968 * http://www.isthe.com/chongo/
3970 * Share and Enjoy! :-)
3973 typedef unsigned long long mdb_hash_t;
3974 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3976 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3977 * @param[in] val value to hash
3978 * @param[in] hval initial value for hash
3979 * @return 64 bit hash
3981 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3982 * hval arg on the first call.
3985 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3987 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3988 unsigned char *end = s + val->mv_size;
3990 * FNV-1a hash each octet of the string
3993 /* xor the bottom with the current octet */
3994 hval ^= (mdb_hash_t)*s++;
3996 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3997 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3998 (hval << 7) + (hval << 8) + (hval << 40);
4000 /* return our new hash value */
4004 /** Hash the string and output the encoded hash.
4005 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4006 * very short name limits. We don't care about the encoding being reversible,
4007 * we just want to preserve as many bits of the input as possible in a
4008 * small printable string.
4009 * @param[in] str string to hash
4010 * @param[out] encbuf an array of 11 chars to hold the hash
4012 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4015 mdb_pack85(unsigned long l, char *out)
4019 for (i=0; i<5; i++) {
4020 *out++ = mdb_a85[l % 85];
4026 mdb_hash_enc(MDB_val *val, char *encbuf)
4028 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4030 mdb_pack85(h, encbuf);
4031 mdb_pack85(h>>32, encbuf+5);
4036 /** Open and/or initialize the lock region for the environment.
4037 * @param[in] env The LMDB environment.
4038 * @param[in] lpath The pathname of the file used for the lock region.
4039 * @param[in] mode The Unix permissions for the file, if we create it.
4040 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4041 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4042 * @return 0 on success, non-zero on failure.
4045 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4048 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4050 # define MDB_ERRCODE_ROFS EROFS
4051 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4052 # define MDB_CLOEXEC O_CLOEXEC
4055 # define MDB_CLOEXEC 0
4062 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4063 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4064 FILE_ATTRIBUTE_NORMAL, NULL);
4066 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4068 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4070 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4075 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4076 /* Lose record locks when exec*() */
4077 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4078 fcntl(env->me_lfd, F_SETFD, fdflags);
4081 if (!(env->me_flags & MDB_NOTLS)) {
4082 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4085 env->me_flags |= MDB_ENV_TXKEY;
4087 /* Windows TLS callbacks need help finding their TLS info. */
4088 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4092 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4096 /* Try to get exclusive lock. If we succeed, then
4097 * nobody is using the lock region and we should initialize it.
4099 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4102 size = GetFileSize(env->me_lfd, NULL);
4104 size = lseek(env->me_lfd, 0, SEEK_END);
4105 if (size == -1) goto fail_errno;
4107 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4108 if (size < rsize && *excl > 0) {
4110 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4111 || !SetEndOfFile(env->me_lfd))
4114 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4118 size = rsize - sizeof(MDB_txninfo);
4119 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4124 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4126 if (!mh) goto fail_errno;
4127 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4129 if (!env->me_txns) goto fail_errno;
4131 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4133 if (m == MAP_FAILED) goto fail_errno;
4139 BY_HANDLE_FILE_INFORMATION stbuf;
4148 if (!mdb_sec_inited) {
4149 InitializeSecurityDescriptor(&mdb_null_sd,
4150 SECURITY_DESCRIPTOR_REVISION);
4151 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4152 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4153 mdb_all_sa.bInheritHandle = FALSE;
4154 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4157 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4158 idbuf.volume = stbuf.dwVolumeSerialNumber;
4159 idbuf.nhigh = stbuf.nFileIndexHigh;
4160 idbuf.nlow = stbuf.nFileIndexLow;
4161 val.mv_data = &idbuf;
4162 val.mv_size = sizeof(idbuf);
4163 mdb_hash_enc(&val, encbuf);
4164 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4165 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4166 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4167 if (!env->me_rmutex) goto fail_errno;
4168 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4169 if (!env->me_wmutex) goto fail_errno;
4170 #elif defined(MDB_USE_POSIX_SEM)
4179 #if defined(__NetBSD__)
4180 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4182 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4183 idbuf.dev = stbuf.st_dev;
4184 idbuf.ino = stbuf.st_ino;
4185 val.mv_data = &idbuf;
4186 val.mv_size = sizeof(idbuf);
4187 mdb_hash_enc(&val, encbuf);
4188 #ifdef MDB_SHORT_SEMNAMES
4189 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4191 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4192 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4193 /* Clean up after a previous run, if needed: Try to
4194 * remove both semaphores before doing anything else.
4196 sem_unlink(env->me_txns->mti_rmname);
4197 sem_unlink(env->me_txns->mti_wmname);
4198 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4199 O_CREAT|O_EXCL, mode, 1);
4200 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4201 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4202 O_CREAT|O_EXCL, mode, 1);
4203 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4204 #else /* MDB_USE_POSIX_SEM */
4205 pthread_mutexattr_t mattr;
4207 if ((rc = pthread_mutexattr_init(&mattr))
4208 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4209 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4210 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4212 pthread_mutexattr_destroy(&mattr);
4213 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4215 env->me_txns->mti_magic = MDB_MAGIC;
4216 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4217 env->me_txns->mti_txnid = 0;
4218 env->me_txns->mti_numreaders = 0;
4221 if (env->me_txns->mti_magic != MDB_MAGIC) {
4222 DPUTS("lock region has invalid magic");
4226 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4227 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4228 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4229 rc = MDB_VERSION_MISMATCH;
4233 if (rc && rc != EACCES && rc != EAGAIN) {
4237 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4238 if (!env->me_rmutex) goto fail_errno;
4239 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4240 if (!env->me_wmutex) goto fail_errno;
4241 #elif defined(MDB_USE_POSIX_SEM)
4242 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4243 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4244 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4245 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4256 /** The name of the lock file in the DB environment */
4257 #define LOCKNAME "/lock.mdb"
4258 /** The name of the data file in the DB environment */
4259 #define DATANAME "/data.mdb"
4260 /** The suffix of the lock file when no subdir is used */
4261 #define LOCKSUFF "-lock"
4262 /** Only a subset of the @ref mdb_env flags can be changed
4263 * at runtime. Changing other flags requires closing the
4264 * environment and re-opening it with the new flags.
4266 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4267 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4268 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4270 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4271 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4275 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4277 int oflags, rc, len, excl = -1;
4278 char *lpath, *dpath;
4280 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4284 if (flags & MDB_NOSUBDIR) {
4285 rc = len + sizeof(LOCKSUFF) + len + 1;
4287 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4292 if (flags & MDB_NOSUBDIR) {
4293 dpath = lpath + len + sizeof(LOCKSUFF);
4294 sprintf(lpath, "%s" LOCKSUFF, path);
4295 strcpy(dpath, path);
4297 dpath = lpath + len + sizeof(LOCKNAME);
4298 sprintf(lpath, "%s" LOCKNAME, path);
4299 sprintf(dpath, "%s" DATANAME, path);
4303 flags |= env->me_flags;
4304 if (flags & MDB_RDONLY) {
4305 /* silently ignore WRITEMAP when we're only getting read access */
4306 flags &= ~MDB_WRITEMAP;
4308 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4309 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4312 env->me_flags = flags |= MDB_ENV_ACTIVE;
4316 env->me_path = strdup(path);
4317 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4318 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4319 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4324 /* For RDONLY, get lockfile after we know datafile exists */
4325 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4326 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4332 if (F_ISSET(flags, MDB_RDONLY)) {
4333 oflags = GENERIC_READ;
4334 len = OPEN_EXISTING;
4336 oflags = GENERIC_READ|GENERIC_WRITE;
4339 mode = FILE_ATTRIBUTE_NORMAL;
4340 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4341 NULL, len, mode, NULL);
4343 if (F_ISSET(flags, MDB_RDONLY))
4346 oflags = O_RDWR | O_CREAT;
4348 env->me_fd = open(dpath, oflags, mode);
4350 if (env->me_fd == INVALID_HANDLE_VALUE) {
4355 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4356 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4361 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4362 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4363 env->me_mfd = env->me_fd;
4365 /* Synchronous fd for meta writes. Needed even with
4366 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4369 len = OPEN_EXISTING;
4370 env->me_mfd = CreateFile(dpath, oflags,
4371 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4372 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4375 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4377 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4382 DPRINTF(("opened dbenv %p", (void *) env));
4384 rc = mdb_env_share_locks(env, &excl);
4388 if (!((flags & MDB_RDONLY) ||
4389 (env->me_pbuf = calloc(1, env->me_psize))))
4395 mdb_env_close0(env, excl);
4401 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4403 mdb_env_close0(MDB_env *env, int excl)
4407 if (!(env->me_flags & MDB_ENV_ACTIVE))
4410 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4411 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4412 free(env->me_dbxs[i].md_name.mv_data);
4415 free(env->me_dbflags);
4418 free(env->me_dirty_list);
4419 mdb_midl_free(env->me_free_pgs);
4421 if (env->me_flags & MDB_ENV_TXKEY) {
4422 pthread_key_delete(env->me_txkey);
4424 /* Delete our key from the global list */
4425 for (i=0; i<mdb_tls_nkeys; i++)
4426 if (mdb_tls_keys[i] == env->me_txkey) {
4427 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4435 munmap(env->me_map, env->me_mapsize);
4437 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4438 (void) close(env->me_mfd);
4439 if (env->me_fd != INVALID_HANDLE_VALUE)
4440 (void) close(env->me_fd);
4442 MDB_PID_T pid = env->me_pid;
4443 /* Clearing readers is done in this function because
4444 * me_txkey with its destructor must be disabled first.
4446 for (i = env->me_numreaders; --i >= 0; )
4447 if (env->me_txns->mti_readers[i].mr_pid == pid)
4448 env->me_txns->mti_readers[i].mr_pid = 0;
4450 if (env->me_rmutex) {
4451 CloseHandle(env->me_rmutex);
4452 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4454 /* Windows automatically destroys the mutexes when
4455 * the last handle closes.
4457 #elif defined(MDB_USE_POSIX_SEM)
4458 if (env->me_rmutex != SEM_FAILED) {
4459 sem_close(env->me_rmutex);
4460 if (env->me_wmutex != SEM_FAILED)
4461 sem_close(env->me_wmutex);
4462 /* If we have the filelock: If we are the
4463 * only remaining user, clean up semaphores.
4466 mdb_env_excl_lock(env, &excl);
4468 sem_unlink(env->me_txns->mti_rmname);
4469 sem_unlink(env->me_txns->mti_wmname);
4473 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4475 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4478 /* Unlock the lockfile. Windows would have unlocked it
4479 * after closing anyway, but not necessarily at once.
4481 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4484 (void) close(env->me_lfd);
4487 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4492 mdb_env_close(MDB_env *env)
4499 VGMEMP_DESTROY(env);
4500 while ((dp = env->me_dpages) != NULL) {
4501 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4502 env->me_dpages = dp->mp_next;
4506 mdb_env_close0(env, 0);
4510 /** Compare two items pointing at aligned size_t's */
4512 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4514 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4515 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4518 /** Compare two items pointing at aligned unsigned int's */
4520 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4522 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4523 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4526 /** Compare two items pointing at unsigned ints of unknown alignment.
4527 * Nodes and keys are guaranteed to be 2-byte aligned.
4530 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4532 #if BYTE_ORDER == LITTLE_ENDIAN
4533 unsigned short *u, *c;
4536 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4537 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4540 } while(!x && u > (unsigned short *)a->mv_data);
4543 unsigned short *u, *c, *end;
4546 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4547 u = (unsigned short *)a->mv_data;
4548 c = (unsigned short *)b->mv_data;
4551 } while(!x && u < end);
4556 /** Compare two items pointing at size_t's of unknown alignment. */
4557 #ifdef MISALIGNED_OK
4558 # define mdb_cmp_clong mdb_cmp_long
4560 # define mdb_cmp_clong mdb_cmp_cint
4563 /** Compare two items lexically */
4565 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4572 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4578 diff = memcmp(a->mv_data, b->mv_data, len);
4579 return diff ? diff : len_diff<0 ? -1 : len_diff;
4582 /** Compare two items in reverse byte order */
4584 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4586 const unsigned char *p1, *p2, *p1_lim;
4590 p1_lim = (const unsigned char *)a->mv_data;
4591 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4592 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4594 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4600 while (p1 > p1_lim) {
4601 diff = *--p1 - *--p2;
4605 return len_diff<0 ? -1 : len_diff;
4608 /** Search for key within a page, using binary search.
4609 * Returns the smallest entry larger or equal to the key.
4610 * If exactp is non-null, stores whether the found entry was an exact match
4611 * in *exactp (1 or 0).
4612 * Updates the cursor index with the index of the found entry.
4613 * If no entry larger or equal to the key is found, returns NULL.
4616 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4618 unsigned int i = 0, nkeys;
4621 MDB_page *mp = mc->mc_pg[mc->mc_top];
4622 MDB_node *node = NULL;
4627 nkeys = NUMKEYS(mp);
4629 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4630 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4633 low = IS_LEAF(mp) ? 0 : 1;
4635 cmp = mc->mc_dbx->md_cmp;
4637 /* Branch pages have no data, so if using integer keys,
4638 * alignment is guaranteed. Use faster mdb_cmp_int.
4640 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4641 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4648 nodekey.mv_size = mc->mc_db->md_pad;
4649 node = NODEPTR(mp, 0); /* fake */
4650 while (low <= high) {
4651 i = (low + high) >> 1;
4652 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4653 rc = cmp(key, &nodekey);
4654 DPRINTF(("found leaf index %u [%s], rc = %i",
4655 i, DKEY(&nodekey), rc));
4664 while (low <= high) {
4665 i = (low + high) >> 1;
4667 node = NODEPTR(mp, i);
4668 nodekey.mv_size = NODEKSZ(node);
4669 nodekey.mv_data = NODEKEY(node);
4671 rc = cmp(key, &nodekey);
4674 DPRINTF(("found leaf index %u [%s], rc = %i",
4675 i, DKEY(&nodekey), rc));
4677 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4678 i, DKEY(&nodekey), NODEPGNO(node), rc));
4689 if (rc > 0) { /* Found entry is less than the key. */
4690 i++; /* Skip to get the smallest entry larger than key. */
4692 node = NODEPTR(mp, i);
4695 *exactp = (rc == 0 && nkeys > 0);
4696 /* store the key index */
4697 mc->mc_ki[mc->mc_top] = i;
4699 /* There is no entry larger or equal to the key. */
4702 /* nodeptr is fake for LEAF2 */
4708 mdb_cursor_adjust(MDB_cursor *mc, func)
4712 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4713 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4720 /** Pop a page off the top of the cursor's stack. */
4722 mdb_cursor_pop(MDB_cursor *mc)
4726 MDB_page *top = mc->mc_pg[mc->mc_top];
4732 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4733 DDBI(mc), (void *) mc));
4737 /** Push a page onto the top of the cursor's stack. */
4739 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4741 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4742 DDBI(mc), (void *) mc));
4744 if (mc->mc_snum >= CURSOR_STACK) {
4745 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4746 return MDB_CURSOR_FULL;
4749 mc->mc_top = mc->mc_snum++;
4750 mc->mc_pg[mc->mc_top] = mp;
4751 mc->mc_ki[mc->mc_top] = 0;
4756 /** Find the address of the page corresponding to a given page number.
4757 * @param[in] txn the transaction for this access.
4758 * @param[in] pgno the page number for the page to retrieve.
4759 * @param[out] ret address of a pointer where the page's address will be stored.
4760 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4761 * @return 0 on success, non-zero on failure.
4764 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4766 MDB_env *env = txn->mt_env;
4770 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4774 MDB_ID2L dl = tx2->mt_u.dirty_list;
4776 /* Spilled pages were dirtied in this txn and flushed
4777 * because the dirty list got full. Bring this page
4778 * back in from the map (but don't unspill it here,
4779 * leave that unless page_touch happens again).
4781 if (tx2->mt_spill_pgs) {
4782 MDB_ID pn = pgno << 1;
4783 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4784 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4785 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4790 unsigned x = mdb_mid2l_search(dl, pgno);
4791 if (x <= dl[0].mid && dl[x].mid == pgno) {
4797 } while ((tx2 = tx2->mt_parent) != NULL);
4800 if (pgno < txn->mt_next_pgno) {
4802 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4804 DPRINTF(("page %"Z"u not found", pgno));
4805 txn->mt_flags |= MDB_TXN_ERROR;
4806 return MDB_PAGE_NOTFOUND;
4816 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4817 * The cursor is at the root page, set up the rest of it.
4820 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4822 MDB_page *mp = mc->mc_pg[mc->mc_top];
4826 while (IS_BRANCH(mp)) {
4830 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4831 mdb_cassert(mc, NUMKEYS(mp) > 1);
4832 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4834 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4836 if (flags & MDB_PS_LAST)
4837 i = NUMKEYS(mp) - 1;
4840 node = mdb_node_search(mc, key, &exact);
4842 i = NUMKEYS(mp) - 1;
4844 i = mc->mc_ki[mc->mc_top];
4846 mdb_cassert(mc, i > 0);
4850 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4853 mdb_cassert(mc, i < NUMKEYS(mp));
4854 node = NODEPTR(mp, i);
4856 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4859 mc->mc_ki[mc->mc_top] = i;
4860 if ((rc = mdb_cursor_push(mc, mp)))
4863 if (flags & MDB_PS_MODIFY) {
4864 if ((rc = mdb_page_touch(mc)) != 0)
4866 mp = mc->mc_pg[mc->mc_top];
4871 DPRINTF(("internal error, index points to a %02X page!?",
4873 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4874 return MDB_CORRUPTED;
4877 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4878 key ? DKEY(key) : "null"));
4879 mc->mc_flags |= C_INITIALIZED;
4880 mc->mc_flags &= ~C_EOF;
4885 /** Search for the lowest key under the current branch page.
4886 * This just bypasses a NUMKEYS check in the current page
4887 * before calling mdb_page_search_root(), because the callers
4888 * are all in situations where the current page is known to
4892 mdb_page_search_lowest(MDB_cursor *mc)
4894 MDB_page *mp = mc->mc_pg[mc->mc_top];
4895 MDB_node *node = NODEPTR(mp, 0);
4898 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4901 mc->mc_ki[mc->mc_top] = 0;
4902 if ((rc = mdb_cursor_push(mc, mp)))
4904 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4907 /** Search for the page a given key should be in.
4908 * Push it and its parent pages on the cursor stack.
4909 * @param[in,out] mc the cursor for this operation.
4910 * @param[in] key the key to search for, or NULL for first/last page.
4911 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4912 * are touched (updated with new page numbers).
4913 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4914 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4915 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4916 * @return 0 on success, non-zero on failure.
4919 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4924 /* Make sure the txn is still viable, then find the root from
4925 * the txn's db table and set it as the root of the cursor's stack.
4927 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4928 DPUTS("transaction has failed, must abort");
4931 /* Make sure we're using an up-to-date root */
4932 if (*mc->mc_dbflag & DB_STALE) {
4934 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4935 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4942 MDB_node *leaf = mdb_node_search(&mc2,
4943 &mc->mc_dbx->md_name, &exact);
4945 return MDB_NOTFOUND;
4946 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4949 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4951 /* The txn may not know this DBI, or another process may
4952 * have dropped and recreated the DB with other flags.
4954 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4955 return MDB_INCOMPATIBLE;
4956 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4958 *mc->mc_dbflag &= ~DB_STALE;
4960 root = mc->mc_db->md_root;
4962 if (root == P_INVALID) { /* Tree is empty. */
4963 DPUTS("tree is empty");
4964 return MDB_NOTFOUND;
4968 mdb_cassert(mc, root > 1);
4969 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4970 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4976 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4977 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4979 if (flags & MDB_PS_MODIFY) {
4980 if ((rc = mdb_page_touch(mc)))
4984 if (flags & MDB_PS_ROOTONLY)
4987 return mdb_page_search_root(mc, key, flags);
4991 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4993 MDB_txn *txn = mc->mc_txn;
4994 pgno_t pg = mp->mp_pgno;
4995 unsigned x = 0, ovpages = mp->mp_pages;
4996 MDB_env *env = txn->mt_env;
4997 MDB_IDL sl = txn->mt_spill_pgs;
4998 MDB_ID pn = pg << 1;
5001 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5002 /* If the page is dirty or on the spill list we just acquired it,
5003 * so we should give it back to our current free list, if any.
5004 * Otherwise put it onto the list of pages we freed in this txn.
5006 * Won't create me_pghead: me_pglast must be inited along with it.
5007 * Unsupported in nested txns: They would need to hide the page
5008 * range in ancestor txns' dirty and spilled lists.
5010 if (env->me_pghead &&
5012 ((mp->mp_flags & P_DIRTY) ||
5013 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5017 MDB_ID2 *dl, ix, iy;
5018 rc = mdb_midl_need(&env->me_pghead, ovpages);
5021 if (!(mp->mp_flags & P_DIRTY)) {
5022 /* This page is no longer spilled */
5029 /* Remove from dirty list */
5030 dl = txn->mt_u.dirty_list;
5032 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5038 mdb_cassert(mc, x > 1);
5040 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5041 txn->mt_flags |= MDB_TXN_ERROR;
5042 return MDB_CORRUPTED;
5045 if (!(env->me_flags & MDB_WRITEMAP))
5046 mdb_dpage_free(env, mp);
5048 /* Insert in me_pghead */
5049 mop = env->me_pghead;
5050 j = mop[0] + ovpages;
5051 for (i = mop[0]; i && mop[i] < pg; i--)
5057 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5061 mc->mc_db->md_overflow_pages -= ovpages;
5065 /** Return the data associated with a given node.
5066 * @param[in] txn The transaction for this operation.
5067 * @param[in] leaf The node being read.
5068 * @param[out] data Updated to point to the node's data.
5069 * @return 0 on success, non-zero on failure.
5072 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5074 MDB_page *omp; /* overflow page */
5078 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5079 data->mv_size = NODEDSZ(leaf);
5080 data->mv_data = NODEDATA(leaf);
5084 /* Read overflow data.
5086 data->mv_size = NODEDSZ(leaf);
5087 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5088 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5089 DPRINTF(("read overflow page %"Z"u failed", pgno));
5092 data->mv_data = METADATA(omp);
5098 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5099 MDB_val *key, MDB_val *data)
5106 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5108 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5111 if (txn->mt_flags & MDB_TXN_ERROR)
5114 mdb_cursor_init(&mc, txn, dbi, &mx);
5115 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5118 /** Find a sibling for a page.
5119 * Replaces the page at the top of the cursor's stack with the
5120 * specified sibling, if one exists.
5121 * @param[in] mc The cursor for this operation.
5122 * @param[in] move_right Non-zero if the right sibling is requested,
5123 * otherwise the left sibling.
5124 * @return 0 on success, non-zero on failure.
5127 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5133 if (mc->mc_snum < 2) {
5134 return MDB_NOTFOUND; /* root has no siblings */
5138 DPRINTF(("parent page is page %"Z"u, index %u",
5139 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5141 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5142 : (mc->mc_ki[mc->mc_top] == 0)) {
5143 DPRINTF(("no more keys left, moving to %s sibling",
5144 move_right ? "right" : "left"));
5145 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5146 /* undo cursor_pop before returning */
5153 mc->mc_ki[mc->mc_top]++;
5155 mc->mc_ki[mc->mc_top]--;
5156 DPRINTF(("just moving to %s index key %u",
5157 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5159 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5161 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5162 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5163 /* mc will be inconsistent if caller does mc_snum++ as above */
5164 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5168 mdb_cursor_push(mc, mp);
5170 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5175 /** Move the cursor to the next data item. */
5177 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5183 if (mc->mc_flags & C_EOF) {
5184 return MDB_NOTFOUND;
5187 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5189 mp = mc->mc_pg[mc->mc_top];
5191 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5192 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5193 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5194 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5195 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5196 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5197 if (rc == MDB_SUCCESS)
5198 MDB_GET_KEY(leaf, key);
5203 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5204 if (op == MDB_NEXT_DUP)
5205 return MDB_NOTFOUND;
5209 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5210 mdb_dbg_pgno(mp), (void *) mc));
5211 if (mc->mc_flags & C_DEL)
5214 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5215 DPUTS("=====> move to next sibling page");
5216 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5217 mc->mc_flags |= C_EOF;
5220 mp = mc->mc_pg[mc->mc_top];
5221 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5223 mc->mc_ki[mc->mc_top]++;
5226 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5227 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5230 key->mv_size = mc->mc_db->md_pad;
5231 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5235 mdb_cassert(mc, IS_LEAF(mp));
5236 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5238 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5239 mdb_xcursor_init1(mc, leaf);
5242 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5245 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5246 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5247 if (rc != MDB_SUCCESS)
5252 MDB_GET_KEY(leaf, key);
5256 /** Move the cursor to the previous data item. */
5258 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5264 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5266 mp = mc->mc_pg[mc->mc_top];
5268 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5269 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5270 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5271 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5272 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5273 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5274 if (rc == MDB_SUCCESS)
5275 MDB_GET_KEY(leaf, key);
5279 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5280 if (op == MDB_PREV_DUP)
5281 return MDB_NOTFOUND;
5286 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5287 mdb_dbg_pgno(mp), (void *) mc));
5289 if (mc->mc_ki[mc->mc_top] == 0) {
5290 DPUTS("=====> move to prev sibling page");
5291 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5294 mp = mc->mc_pg[mc->mc_top];
5295 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5296 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5298 mc->mc_ki[mc->mc_top]--;
5300 mc->mc_flags &= ~C_EOF;
5302 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5303 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5306 key->mv_size = mc->mc_db->md_pad;
5307 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5311 mdb_cassert(mc, IS_LEAF(mp));
5312 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5314 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5315 mdb_xcursor_init1(mc, leaf);
5318 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5321 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5322 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5323 if (rc != MDB_SUCCESS)
5328 MDB_GET_KEY(leaf, key);
5332 /** Set the cursor on a specific data item. */
5334 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5335 MDB_cursor_op op, int *exactp)
5339 MDB_node *leaf = NULL;
5342 if (key->mv_size == 0)
5343 return MDB_BAD_VALSIZE;
5346 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5348 /* See if we're already on the right page */
5349 if (mc->mc_flags & C_INITIALIZED) {
5352 mp = mc->mc_pg[mc->mc_top];
5354 mc->mc_ki[mc->mc_top] = 0;
5355 return MDB_NOTFOUND;
5357 if (mp->mp_flags & P_LEAF2) {
5358 nodekey.mv_size = mc->mc_db->md_pad;
5359 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5361 leaf = NODEPTR(mp, 0);
5362 MDB_GET_KEY2(leaf, nodekey);
5364 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5366 /* Probably happens rarely, but first node on the page
5367 * was the one we wanted.
5369 mc->mc_ki[mc->mc_top] = 0;
5376 unsigned int nkeys = NUMKEYS(mp);
5378 if (mp->mp_flags & P_LEAF2) {
5379 nodekey.mv_data = LEAF2KEY(mp,
5380 nkeys-1, nodekey.mv_size);
5382 leaf = NODEPTR(mp, nkeys-1);
5383 MDB_GET_KEY2(leaf, nodekey);
5385 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5387 /* last node was the one we wanted */
5388 mc->mc_ki[mc->mc_top] = nkeys-1;
5394 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5395 /* This is definitely the right page, skip search_page */
5396 if (mp->mp_flags & P_LEAF2) {
5397 nodekey.mv_data = LEAF2KEY(mp,
5398 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5400 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5401 MDB_GET_KEY2(leaf, nodekey);
5403 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5405 /* current node was the one we wanted */
5415 /* If any parents have right-sibs, search.
5416 * Otherwise, there's nothing further.
5418 for (i=0; i<mc->mc_top; i++)
5420 NUMKEYS(mc->mc_pg[i])-1)
5422 if (i == mc->mc_top) {
5423 /* There are no other pages */
5424 mc->mc_ki[mc->mc_top] = nkeys;
5425 return MDB_NOTFOUND;
5429 /* There are no other pages */
5430 mc->mc_ki[mc->mc_top] = 0;
5431 if (op == MDB_SET_RANGE && !exactp) {
5435 return MDB_NOTFOUND;
5439 rc = mdb_page_search(mc, key, 0);
5440 if (rc != MDB_SUCCESS)
5443 mp = mc->mc_pg[mc->mc_top];
5444 mdb_cassert(mc, IS_LEAF(mp));
5447 leaf = mdb_node_search(mc, key, exactp);
5448 if (exactp != NULL && !*exactp) {
5449 /* MDB_SET specified and not an exact match. */
5450 return MDB_NOTFOUND;
5454 DPUTS("===> inexact leaf not found, goto sibling");
5455 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5456 return rc; /* no entries matched */
5457 mp = mc->mc_pg[mc->mc_top];
5458 mdb_cassert(mc, IS_LEAF(mp));
5459 leaf = NODEPTR(mp, 0);
5463 mc->mc_flags |= C_INITIALIZED;
5464 mc->mc_flags &= ~C_EOF;
5467 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5468 key->mv_size = mc->mc_db->md_pad;
5469 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5474 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5475 mdb_xcursor_init1(mc, leaf);
5478 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5479 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5480 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5483 if (op == MDB_GET_BOTH) {
5489 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5490 if (rc != MDB_SUCCESS)
5493 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5495 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5497 rc = mc->mc_dbx->md_dcmp(data, &d2);
5499 if (op == MDB_GET_BOTH || rc > 0)
5500 return MDB_NOTFOUND;
5507 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5508 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5513 /* The key already matches in all other cases */
5514 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5515 MDB_GET_KEY(leaf, key);
5516 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5521 /** Move the cursor to the first item in the database. */
5523 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5529 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5531 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5532 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5533 if (rc != MDB_SUCCESS)
5536 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5538 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5539 mc->mc_flags |= C_INITIALIZED;
5540 mc->mc_flags &= ~C_EOF;
5542 mc->mc_ki[mc->mc_top] = 0;
5544 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5545 key->mv_size = mc->mc_db->md_pad;
5546 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5551 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5552 mdb_xcursor_init1(mc, leaf);
5553 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5557 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5561 MDB_GET_KEY(leaf, key);
5565 /** Move the cursor to the last item in the database. */
5567 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5573 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5575 if (!(mc->mc_flags & C_EOF)) {
5577 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5578 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5579 if (rc != MDB_SUCCESS)
5582 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5585 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5586 mc->mc_flags |= C_INITIALIZED|C_EOF;
5587 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5589 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5590 key->mv_size = mc->mc_db->md_pad;
5591 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5596 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5597 mdb_xcursor_init1(mc, leaf);
5598 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5602 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5607 MDB_GET_KEY(leaf, key);
5612 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5617 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5622 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5626 case MDB_GET_CURRENT:
5627 if (!(mc->mc_flags & C_INITIALIZED)) {
5630 MDB_page *mp = mc->mc_pg[mc->mc_top];
5631 int nkeys = NUMKEYS(mp);
5632 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5633 mc->mc_ki[mc->mc_top] = nkeys;
5639 key->mv_size = mc->mc_db->md_pad;
5640 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5642 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5643 MDB_GET_KEY(leaf, key);
5645 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5646 if (mc->mc_flags & C_DEL)
5647 mdb_xcursor_init1(mc, leaf);
5648 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5650 rc = mdb_node_read(mc->mc_txn, leaf, data);
5657 case MDB_GET_BOTH_RANGE:
5662 if (mc->mc_xcursor == NULL) {
5663 rc = MDB_INCOMPATIBLE;
5673 rc = mdb_cursor_set(mc, key, data, op,
5674 op == MDB_SET_RANGE ? NULL : &exact);
5677 case MDB_GET_MULTIPLE:
5678 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5682 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5683 rc = MDB_INCOMPATIBLE;
5687 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5688 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5691 case MDB_NEXT_MULTIPLE:
5696 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5697 rc = MDB_INCOMPATIBLE;
5700 if (!(mc->mc_flags & C_INITIALIZED))
5701 rc = mdb_cursor_first(mc, key, data);
5703 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5704 if (rc == MDB_SUCCESS) {
5705 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5708 mx = &mc->mc_xcursor->mx_cursor;
5709 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5711 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5712 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5720 case MDB_NEXT_NODUP:
5721 if (!(mc->mc_flags & C_INITIALIZED))
5722 rc = mdb_cursor_first(mc, key, data);
5724 rc = mdb_cursor_next(mc, key, data, op);
5728 case MDB_PREV_NODUP:
5729 if (!(mc->mc_flags & C_INITIALIZED)) {
5730 rc = mdb_cursor_last(mc, key, data);
5733 mc->mc_flags |= C_INITIALIZED;
5734 mc->mc_ki[mc->mc_top]++;
5736 rc = mdb_cursor_prev(mc, key, data, op);
5739 rc = mdb_cursor_first(mc, key, data);
5742 mfunc = mdb_cursor_first;
5744 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5748 if (mc->mc_xcursor == NULL) {
5749 rc = MDB_INCOMPATIBLE;
5752 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5756 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5759 rc = mdb_cursor_last(mc, key, data);
5762 mfunc = mdb_cursor_last;
5765 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5770 if (mc->mc_flags & C_DEL)
5771 mc->mc_flags ^= C_DEL;
5776 /** Touch all the pages in the cursor stack. Set mc_top.
5777 * Makes sure all the pages are writable, before attempting a write operation.
5778 * @param[in] mc The cursor to operate on.
5781 mdb_cursor_touch(MDB_cursor *mc)
5783 int rc = MDB_SUCCESS;
5785 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5788 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5789 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5792 *mc->mc_dbflag |= DB_DIRTY;
5797 rc = mdb_page_touch(mc);
5798 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5799 mc->mc_top = mc->mc_snum-1;
5804 /** Do not spill pages to disk if txn is getting full, may fail instead */
5805 #define MDB_NOSPILL 0x8000
5808 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5811 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5813 MDB_node *leaf = NULL;
5816 MDB_val xdata, *rdata, dkey, olddata;
5818 int do_sub = 0, insert_key, insert_data;
5819 unsigned int mcount = 0, dcount = 0, nospill;
5822 unsigned int nflags;
5825 if (mc == NULL || key == NULL)
5828 env = mc->mc_txn->mt_env;
5830 /* Check this first so counter will always be zero on any
5833 if (flags & MDB_MULTIPLE) {
5834 dcount = data[1].mv_size;
5835 data[1].mv_size = 0;
5836 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5837 return MDB_INCOMPATIBLE;
5840 nospill = flags & MDB_NOSPILL;
5841 flags &= ~MDB_NOSPILL;
5843 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5844 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5846 if (key->mv_size-1 >= ENV_MAXKEY(env))
5847 return MDB_BAD_VALSIZE;
5849 #if SIZE_MAX > MAXDATASIZE
5850 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5851 return MDB_BAD_VALSIZE;
5853 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5854 return MDB_BAD_VALSIZE;
5857 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5858 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5862 if (flags == MDB_CURRENT) {
5863 if (!(mc->mc_flags & C_INITIALIZED))
5866 } else if (mc->mc_db->md_root == P_INVALID) {
5867 /* new database, cursor has nothing to point to */
5870 mc->mc_flags &= ~C_INITIALIZED;
5875 if (flags & MDB_APPEND) {
5877 rc = mdb_cursor_last(mc, &k2, &d2);
5879 rc = mc->mc_dbx->md_cmp(key, &k2);
5882 mc->mc_ki[mc->mc_top]++;
5884 /* new key is <= last key */
5889 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5891 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5892 DPRINTF(("duplicate key [%s]", DKEY(key)));
5894 return MDB_KEYEXIST;
5896 if (rc && rc != MDB_NOTFOUND)
5900 if (mc->mc_flags & C_DEL)
5901 mc->mc_flags ^= C_DEL;
5903 /* Cursor is positioned, check for room in the dirty list */
5905 if (flags & MDB_MULTIPLE) {
5907 xdata.mv_size = data->mv_size * dcount;
5911 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5915 if (rc == MDB_NO_ROOT) {
5917 /* new database, write a root leaf page */
5918 DPUTS("allocating new root leaf page");
5919 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5922 mdb_cursor_push(mc, np);
5923 mc->mc_db->md_root = np->mp_pgno;
5924 mc->mc_db->md_depth++;
5925 *mc->mc_dbflag |= DB_DIRTY;
5926 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5928 np->mp_flags |= P_LEAF2;
5929 mc->mc_flags |= C_INITIALIZED;
5931 /* make sure all cursor pages are writable */
5932 rc2 = mdb_cursor_touch(mc);
5937 insert_key = insert_data = rc;
5939 /* The key does not exist */
5940 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5941 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5942 LEAFSIZE(key, data) > env->me_nodemax)
5944 /* Too big for a node, insert in sub-DB */
5945 fp_flags = P_LEAF|P_DIRTY;
5947 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5948 fp->mp_lower = fp->mp_upper = olddata.mv_size = MP_LOBASE;
5952 /* there's only a key anyway, so this is a no-op */
5953 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5955 unsigned int ksize = mc->mc_db->md_pad;
5956 if (key->mv_size != ksize)
5957 return MDB_BAD_VALSIZE;
5958 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5959 memcpy(ptr, key->mv_data, ksize);
5964 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5965 olddata.mv_size = NODEDSZ(leaf);
5966 olddata.mv_data = NODEDATA(leaf);
5969 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5970 /* Prepare (sub-)page/sub-DB to accept the new item,
5971 * if needed. fp: old sub-page or a header faking
5972 * it. mp: new (sub-)page. offset: growth in page
5973 * size. xdata: node data with new page or DB.
5975 unsigned i, offset = 0;
5976 mp = fp = xdata.mv_data = env->me_pbuf;
5977 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5979 /* Was a single item before, must convert now */
5980 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5981 /* Just overwrite the current item */
5982 if (flags == MDB_CURRENT)
5985 #if UINT_MAX < SIZE_MAX
5986 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5987 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
5989 /* does data match? */
5990 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5991 if (flags & MDB_NODUPDATA)
5992 return MDB_KEYEXIST;
5997 /* Back up original data item */
5998 dkey.mv_size = olddata.mv_size;
5999 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6001 /* Make sub-page header for the dup items, with dummy body */
6002 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6003 fp->mp_lower = MP_LOBASE;
6004 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6005 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6006 fp->mp_flags |= P_LEAF2;
6007 fp->mp_pad = data->mv_size;
6008 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6010 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6011 (dkey.mv_size & 1) + (data->mv_size & 1);
6013 fp->mp_upper = xdata.mv_size - MP_HIBASE;
6014 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6015 } else if (leaf->mn_flags & F_SUBDATA) {
6016 /* Data is on sub-DB, just store it */
6017 flags |= F_DUPDATA|F_SUBDATA;
6020 /* Data is on sub-page */
6021 fp = olddata.mv_data;
6024 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6025 offset = EVEN(NODESIZE + sizeof(indx_t) +
6029 offset = fp->mp_pad;
6030 if (SIZELEFT(fp) < offset) {
6031 offset *= 4; /* space for 4 more */
6034 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6036 fp->mp_flags |= P_DIRTY;
6037 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6038 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6042 xdata.mv_size = olddata.mv_size + offset;
6045 fp_flags = fp->mp_flags;
6046 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6047 /* Too big for a sub-page, convert to sub-DB */
6048 fp_flags &= ~P_SUBP;
6050 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6051 fp_flags |= P_LEAF2;
6052 dummy.md_pad = fp->mp_pad;
6053 dummy.md_flags = MDB_DUPFIXED;
6054 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6055 dummy.md_flags |= MDB_INTEGERKEY;
6061 dummy.md_branch_pages = 0;
6062 dummy.md_leaf_pages = 1;
6063 dummy.md_overflow_pages = 0;
6064 dummy.md_entries = NUMKEYS(fp);
6065 xdata.mv_size = sizeof(MDB_db);
6066 xdata.mv_data = &dummy;
6067 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6069 offset = env->me_psize - olddata.mv_size;
6070 flags |= F_DUPDATA|F_SUBDATA;
6071 dummy.md_root = mp->mp_pgno;
6074 mp->mp_flags = fp_flags | P_DIRTY;
6075 mp->mp_pad = fp->mp_pad;
6076 mp->mp_lower = fp->mp_lower;
6077 mp->mp_upper = fp->mp_upper + offset;
6078 if (fp_flags & P_LEAF2) {
6079 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6081 memcpy((char *)mp + mp->mp_upper + MP_HIBASE, (char *)fp + fp->mp_upper + MP_HIBASE,
6082 olddata.mv_size - fp->mp_upper);
6083 for (i=0; i<NUMKEYS(fp); i++)
6084 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6092 mdb_node_del(mc, 0);
6096 /* overflow page overwrites need special handling */
6097 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6100 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6102 memcpy(&pg, olddata.mv_data, sizeof(pg));
6103 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6105 ovpages = omp->mp_pages;
6107 /* Is the ov page large enough? */
6108 if (ovpages >= dpages) {
6109 if (!(omp->mp_flags & P_DIRTY) &&
6110 (level || (env->me_flags & MDB_WRITEMAP)))
6112 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6115 level = 0; /* dirty in this txn or clean */
6118 if (omp->mp_flags & P_DIRTY) {
6119 /* yes, overwrite it. Note in this case we don't
6120 * bother to try shrinking the page if the new data
6121 * is smaller than the overflow threshold.
6124 /* It is writable only in a parent txn */
6125 size_t sz = (size_t) env->me_psize * ovpages, off;
6126 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6132 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6133 mdb_cassert(mc, rc2 == 0);
6134 if (!(flags & MDB_RESERVE)) {
6135 /* Copy end of page, adjusting alignment so
6136 * compiler may copy words instead of bytes.
6138 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6139 memcpy((size_t *)((char *)np + off),
6140 (size_t *)((char *)omp + off), sz - off);
6143 memcpy(np, omp, sz); /* Copy beginning of page */
6146 SETDSZ(leaf, data->mv_size);
6147 if (F_ISSET(flags, MDB_RESERVE))
6148 data->mv_data = METADATA(omp);
6150 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6154 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6156 } else if (data->mv_size == olddata.mv_size) {
6157 /* same size, just replace it. Note that we could
6158 * also reuse this node if the new data is smaller,
6159 * but instead we opt to shrink the node in that case.
6161 if (F_ISSET(flags, MDB_RESERVE))
6162 data->mv_data = olddata.mv_data;
6163 else if (!(mc->mc_flags & C_SUB))
6164 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6166 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6169 mdb_node_del(mc, 0);
6175 nflags = flags & NODE_ADD_FLAGS;
6176 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6177 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6178 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6179 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6181 nflags |= MDB_SPLIT_REPLACE;
6182 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6184 /* There is room already in this leaf page. */
6185 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6186 if (rc == 0 && insert_key) {
6187 /* Adjust other cursors pointing to mp */
6188 MDB_cursor *m2, *m3;
6189 MDB_dbi dbi = mc->mc_dbi;
6190 unsigned i = mc->mc_top;
6191 MDB_page *mp = mc->mc_pg[i];
6193 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6194 if (mc->mc_flags & C_SUB)
6195 m3 = &m2->mc_xcursor->mx_cursor;
6198 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6199 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6206 if (rc == MDB_SUCCESS) {
6207 /* Now store the actual data in the child DB. Note that we're
6208 * storing the user data in the keys field, so there are strict
6209 * size limits on dupdata. The actual data fields of the child
6210 * DB are all zero size.
6218 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6219 if (flags & MDB_CURRENT) {
6220 xflags = MDB_CURRENT|MDB_NOSPILL;
6222 mdb_xcursor_init1(mc, leaf);
6223 xflags = (flags & MDB_NODUPDATA) ?
6224 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6226 /* converted, write the original data first */
6228 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6232 /* Adjust other cursors pointing to mp */
6234 unsigned i = mc->mc_top;
6235 MDB_page *mp = mc->mc_pg[i];
6237 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6238 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6239 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6240 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6241 mdb_xcursor_init1(m2, leaf);
6245 /* we've done our job */
6248 ecount = mc->mc_xcursor->mx_db.md_entries;
6249 if (flags & MDB_APPENDDUP)
6250 xflags |= MDB_APPEND;
6251 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6252 if (flags & F_SUBDATA) {
6253 void *db = NODEDATA(leaf);
6254 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6256 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6258 /* Increment count unless we just replaced an existing item. */
6260 mc->mc_db->md_entries++;
6262 /* Invalidate txn if we created an empty sub-DB */
6265 /* If we succeeded and the key didn't exist before,
6266 * make sure the cursor is marked valid.
6268 mc->mc_flags |= C_INITIALIZED;
6270 if (flags & MDB_MULTIPLE) {
6273 /* let caller know how many succeeded, if any */
6274 data[1].mv_size = mcount;
6275 if (mcount < dcount) {
6276 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6277 insert_key = insert_data = 0;
6284 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6287 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6292 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6298 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6299 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6301 if (!(mc->mc_flags & C_INITIALIZED))
6304 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6305 return MDB_NOTFOUND;
6307 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6310 rc = mdb_cursor_touch(mc);
6314 mp = mc->mc_pg[mc->mc_top];
6317 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6319 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6320 if (flags & MDB_NODUPDATA) {
6321 /* mdb_cursor_del0() will subtract the final entry */
6322 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6324 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6325 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6327 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6330 /* If sub-DB still has entries, we're done */
6331 if (mc->mc_xcursor->mx_db.md_entries) {
6332 if (leaf->mn_flags & F_SUBDATA) {
6333 /* update subDB info */
6334 void *db = NODEDATA(leaf);
6335 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6338 /* shrink fake page */
6339 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6340 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6341 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6342 /* fix other sub-DB cursors pointed at this fake page */
6343 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6344 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6345 if (m2->mc_pg[mc->mc_top] == mp &&
6346 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6347 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6350 mc->mc_db->md_entries--;
6351 mc->mc_flags |= C_DEL;
6354 /* otherwise fall thru and delete the sub-DB */
6355 mc->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
6358 if (leaf->mn_flags & F_SUBDATA) {
6359 /* add all the child DB's pages to the free list */
6360 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6366 /* add overflow pages to free list */
6367 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6371 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6372 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6373 (rc = mdb_ovpage_free(mc, omp)))
6378 return mdb_cursor_del0(mc);
6381 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6385 /** Allocate and initialize new pages for a database.
6386 * @param[in] mc a cursor on the database being added to.
6387 * @param[in] flags flags defining what type of page is being allocated.
6388 * @param[in] num the number of pages to allocate. This is usually 1,
6389 * unless allocating overflow pages for a large record.
6390 * @param[out] mp Address of a page, or NULL on failure.
6391 * @return 0 on success, non-zero on failure.
6394 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6399 if ((rc = mdb_page_alloc(mc, num, &np)))
6401 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6402 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6403 np->mp_flags = flags | P_DIRTY;
6404 np->mp_lower = MP_LOBASE;
6405 np->mp_upper = mc->mc_txn->mt_env->me_psize - MP_HIBASE;
6408 mc->mc_db->md_branch_pages++;
6409 else if (IS_LEAF(np))
6410 mc->mc_db->md_leaf_pages++;
6411 else if (IS_OVERFLOW(np)) {
6412 mc->mc_db->md_overflow_pages += num;
6420 /** Calculate the size of a leaf node.
6421 * The size depends on the environment's page size; if a data item
6422 * is too large it will be put onto an overflow page and the node
6423 * size will only include the key and not the data. Sizes are always
6424 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6425 * of the #MDB_node headers.
6426 * @param[in] env The environment handle.
6427 * @param[in] key The key for the node.
6428 * @param[in] data The data for the node.
6429 * @return The number of bytes needed to store the node.
6432 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6436 sz = LEAFSIZE(key, data);
6437 if (sz > env->me_nodemax) {
6438 /* put on overflow page */
6439 sz -= data->mv_size - sizeof(pgno_t);
6442 return EVEN(sz + sizeof(indx_t));
6445 /** Calculate the size of a branch node.
6446 * The size should depend on the environment's page size but since
6447 * we currently don't support spilling large keys onto overflow
6448 * pages, it's simply the size of the #MDB_node header plus the
6449 * size of the key. Sizes are always rounded up to an even number
6450 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6451 * @param[in] env The environment handle.
6452 * @param[in] key The key for the node.
6453 * @return The number of bytes needed to store the node.
6456 mdb_branch_size(MDB_env *env, MDB_val *key)
6461 if (sz > env->me_nodemax) {
6462 /* put on overflow page */
6463 /* not implemented */
6464 /* sz -= key->size - sizeof(pgno_t); */
6467 return sz + sizeof(indx_t);
6470 /** Add a node to the page pointed to by the cursor.
6471 * @param[in] mc The cursor for this operation.
6472 * @param[in] indx The index on the page where the new node should be added.
6473 * @param[in] key The key for the new node.
6474 * @param[in] data The data for the new node, if any.
6475 * @param[in] pgno The page number, if adding a branch node.
6476 * @param[in] flags Flags for the node.
6477 * @return 0 on success, non-zero on failure. Possible errors are:
6479 * <li>ENOMEM - failed to allocate overflow pages for the node.
6480 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6481 * should never happen since all callers already calculate the
6482 * page's free space before calling this function.
6486 mdb_node_add(MDB_cursor *mc, indx_t indx,
6487 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6490 size_t node_size = NODESIZE;
6494 MDB_page *mp = mc->mc_pg[mc->mc_top];
6495 MDB_page *ofp = NULL; /* overflow page */
6498 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6500 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6501 IS_LEAF(mp) ? "leaf" : "branch",
6502 IS_SUBP(mp) ? "sub-" : "",
6503 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6504 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6507 /* Move higher keys up one slot. */
6508 int ksize = mc->mc_db->md_pad, dif;
6509 char *ptr = LEAF2KEY(mp, indx, ksize);
6510 dif = NUMKEYS(mp) - indx;
6512 memmove(ptr+ksize, ptr, dif*ksize);
6513 /* insert new key */
6514 memcpy(ptr, key->mv_data, ksize);
6516 /* Just using these for counting */
6517 mp->mp_lower += sizeof(indx_t);
6518 mp->mp_upper -= ksize - sizeof(indx_t);
6522 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6524 node_size += key->mv_size;
6526 mdb_cassert(mc, data);
6527 if (F_ISSET(flags, F_BIGDATA)) {
6528 /* Data already on overflow page. */
6529 node_size += sizeof(pgno_t);
6530 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6531 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6533 /* Put data on overflow page. */
6534 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6535 data->mv_size, node_size+data->mv_size));
6536 node_size = EVEN(node_size + sizeof(pgno_t));
6537 if ((ssize_t)node_size > room)
6539 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6541 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6545 node_size += data->mv_size;
6548 node_size = EVEN(node_size);
6549 if ((ssize_t)node_size > room)
6553 /* Move higher pointers up one slot. */
6554 for (i = NUMKEYS(mp); i > indx; i--)
6555 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6557 /* Adjust free space offsets. */
6558 ofs = mp->mp_upper - node_size;
6559 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6560 mp->mp_ptrs[indx] = ofs;
6562 mp->mp_lower += sizeof(indx_t);
6564 /* Write the node data. */
6565 node = NODEPTR(mp, indx);
6566 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6567 node->mn_flags = flags;
6569 SETDSZ(node,data->mv_size);
6574 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6577 mdb_cassert(mc, key);
6579 if (F_ISSET(flags, F_BIGDATA))
6580 memcpy(node->mn_data + key->mv_size, data->mv_data,
6582 else if (F_ISSET(flags, MDB_RESERVE))
6583 data->mv_data = node->mn_data + key->mv_size;
6585 memcpy(node->mn_data + key->mv_size, data->mv_data,
6588 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6590 if (F_ISSET(flags, MDB_RESERVE))
6591 data->mv_data = METADATA(ofp);
6593 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6600 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6601 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6602 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6603 DPRINTF(("node size = %"Z"u", node_size));
6604 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6605 return MDB_PAGE_FULL;
6608 /** Delete the specified node from a page.
6609 * @param[in] mc Cursor pointing to the node to delete.
6610 * @param[in] ksize The size of a node. Only used if the page is
6611 * part of a #MDB_DUPFIXED database.
6614 mdb_node_del(MDB_cursor *mc, int ksize)
6616 MDB_page *mp = mc->mc_pg[mc->mc_top];
6617 indx_t indx = mc->mc_ki[mc->mc_top];
6619 indx_t i, j, numkeys, ptr;
6623 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6624 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6625 numkeys = NUMKEYS(mp);
6626 mdb_cassert(mc, indx < numkeys);
6629 int x = numkeys - 1 - indx;
6630 base = LEAF2KEY(mp, indx, ksize);
6632 memmove(base, base + ksize, x * ksize);
6633 mp->mp_lower -= sizeof(indx_t);
6634 mp->mp_upper += ksize - sizeof(indx_t);
6638 node = NODEPTR(mp, indx);
6639 sz = NODESIZE + node->mn_ksize;
6641 if (F_ISSET(node->mn_flags, F_BIGDATA))
6642 sz += sizeof(pgno_t);
6644 sz += NODEDSZ(node);
6648 ptr = mp->mp_ptrs[indx];
6649 for (i = j = 0; i < numkeys; i++) {
6651 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6652 if (mp->mp_ptrs[i] < ptr)
6653 mp->mp_ptrs[j] += sz;
6658 base = (char *)mp + mp->mp_upper + MP_HIBASE;
6659 memmove(base + sz, base, ptr - mp->mp_upper);
6661 mp->mp_lower -= sizeof(indx_t);
6665 /** Compact the main page after deleting a node on a subpage.
6666 * @param[in] mp The main page to operate on.
6667 * @param[in] indx The index of the subpage on the main page.
6670 mdb_node_shrink(MDB_page *mp, indx_t indx)
6676 indx_t i, numkeys, ptr;
6678 node = NODEPTR(mp, indx);
6679 sp = (MDB_page *)NODEDATA(node);
6680 delta = SIZELEFT(sp);
6681 xp = (MDB_page *)((char *)sp + delta);
6683 /* shift subpage upward */
6685 nsize = NUMKEYS(sp) * sp->mp_pad;
6687 return; /* do not make the node uneven-sized */
6688 memmove(METADATA(xp), METADATA(sp), nsize);
6691 numkeys = NUMKEYS(sp);
6692 for (i=numkeys-1; i>=0; i--)
6693 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6695 xp->mp_upper = sp->mp_lower;
6696 xp->mp_lower = sp->mp_lower;
6697 xp->mp_flags = sp->mp_flags;
6698 xp->mp_pad = sp->mp_pad;
6699 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6701 nsize = NODEDSZ(node) - delta;
6702 SETDSZ(node, nsize);
6704 /* shift lower nodes upward */
6705 ptr = mp->mp_ptrs[indx];
6706 numkeys = NUMKEYS(mp);
6707 for (i = 0; i < numkeys; i++) {
6708 if (mp->mp_ptrs[i] <= ptr)
6709 mp->mp_ptrs[i] += delta;
6712 base = (char *)mp + mp->mp_upper + MP_HIBASE;
6713 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6714 mp->mp_upper += delta;
6717 /** Initial setup of a sorted-dups cursor.
6718 * Sorted duplicates are implemented as a sub-database for the given key.
6719 * The duplicate data items are actually keys of the sub-database.
6720 * Operations on the duplicate data items are performed using a sub-cursor
6721 * initialized when the sub-database is first accessed. This function does
6722 * the preliminary setup of the sub-cursor, filling in the fields that
6723 * depend only on the parent DB.
6724 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6727 mdb_xcursor_init0(MDB_cursor *mc)
6729 MDB_xcursor *mx = mc->mc_xcursor;
6731 mx->mx_cursor.mc_xcursor = NULL;
6732 mx->mx_cursor.mc_txn = mc->mc_txn;
6733 mx->mx_cursor.mc_db = &mx->mx_db;
6734 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6735 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6736 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6737 mx->mx_cursor.mc_snum = 0;
6738 mx->mx_cursor.mc_top = 0;
6739 mx->mx_cursor.mc_flags = C_SUB;
6740 mx->mx_dbx.md_name.mv_size = 0;
6741 mx->mx_dbx.md_name.mv_data = NULL;
6742 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6743 mx->mx_dbx.md_dcmp = NULL;
6744 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6747 /** Final setup of a sorted-dups cursor.
6748 * Sets up the fields that depend on the data from the main cursor.
6749 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6750 * @param[in] node The data containing the #MDB_db record for the
6751 * sorted-dup database.
6754 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6756 MDB_xcursor *mx = mc->mc_xcursor;
6758 if (node->mn_flags & F_SUBDATA) {
6759 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6760 mx->mx_cursor.mc_pg[0] = 0;
6761 mx->mx_cursor.mc_snum = 0;
6762 mx->mx_cursor.mc_top = 0;
6763 mx->mx_cursor.mc_flags = C_SUB;
6765 MDB_page *fp = NODEDATA(node);
6766 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6767 mx->mx_db.md_flags = 0;
6768 mx->mx_db.md_depth = 1;
6769 mx->mx_db.md_branch_pages = 0;
6770 mx->mx_db.md_leaf_pages = 1;
6771 mx->mx_db.md_overflow_pages = 0;
6772 mx->mx_db.md_entries = NUMKEYS(fp);
6773 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6774 mx->mx_cursor.mc_snum = 1;
6775 mx->mx_cursor.mc_top = 0;
6776 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6777 mx->mx_cursor.mc_pg[0] = fp;
6778 mx->mx_cursor.mc_ki[0] = 0;
6779 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6780 mx->mx_db.md_flags = MDB_DUPFIXED;
6781 mx->mx_db.md_pad = fp->mp_pad;
6782 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6783 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6786 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6787 mx->mx_db.md_root));
6788 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6789 #if UINT_MAX < SIZE_MAX
6790 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6791 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6795 /** Initialize a cursor for a given transaction and database. */
6797 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6800 mc->mc_backup = NULL;
6803 mc->mc_db = &txn->mt_dbs[dbi];
6804 mc->mc_dbx = &txn->mt_dbxs[dbi];
6805 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6810 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6811 mdb_tassert(txn, mx != NULL);
6812 mc->mc_xcursor = mx;
6813 mdb_xcursor_init0(mc);
6815 mc->mc_xcursor = NULL;
6817 if (*mc->mc_dbflag & DB_STALE) {
6818 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6823 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6826 size_t size = sizeof(MDB_cursor);
6828 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6831 if (txn->mt_flags & MDB_TXN_ERROR)
6834 /* Allow read access to the freelist */
6835 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6838 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6839 size += sizeof(MDB_xcursor);
6841 if ((mc = malloc(size)) != NULL) {
6842 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6843 if (txn->mt_cursors) {
6844 mc->mc_next = txn->mt_cursors[dbi];
6845 txn->mt_cursors[dbi] = mc;
6846 mc->mc_flags |= C_UNTRACK;
6858 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6860 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
6863 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6866 if (txn->mt_flags & MDB_TXN_ERROR)
6869 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6873 /* Return the count of duplicate data items for the current key */
6875 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6879 if (mc == NULL || countp == NULL)
6882 if (mc->mc_xcursor == NULL)
6883 return MDB_INCOMPATIBLE;
6885 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6888 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6889 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6892 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6895 *countp = mc->mc_xcursor->mx_db.md_entries;
6901 mdb_cursor_close(MDB_cursor *mc)
6903 if (mc && !mc->mc_backup) {
6904 /* remove from txn, if tracked */
6905 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6906 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6907 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6909 *prev = mc->mc_next;
6916 mdb_cursor_txn(MDB_cursor *mc)
6918 if (!mc) return NULL;
6923 mdb_cursor_dbi(MDB_cursor *mc)
6928 /** Replace the key for a branch node with a new key.
6929 * @param[in] mc Cursor pointing to the node to operate on.
6930 * @param[in] key The new key to use.
6931 * @return 0 on success, non-zero on failure.
6934 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6940 int delta, ksize, oksize;
6941 indx_t ptr, i, numkeys, indx;
6944 indx = mc->mc_ki[mc->mc_top];
6945 mp = mc->mc_pg[mc->mc_top];
6946 node = NODEPTR(mp, indx);
6947 ptr = mp->mp_ptrs[indx];
6951 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6952 k2.mv_data = NODEKEY(node);
6953 k2.mv_size = node->mn_ksize;
6954 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6956 mdb_dkey(&k2, kbuf2),
6962 /* Sizes must be 2-byte aligned. */
6963 ksize = EVEN(key->mv_size);
6964 oksize = EVEN(node->mn_ksize);
6965 delta = ksize - oksize;
6967 /* Shift node contents if EVEN(key length) changed. */
6969 if (delta > 0 && SIZELEFT(mp) < delta) {
6971 /* not enough space left, do a delete and split */
6972 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6973 pgno = NODEPGNO(node);
6974 mdb_node_del(mc, 0);
6975 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6978 numkeys = NUMKEYS(mp);
6979 for (i = 0; i < numkeys; i++) {
6980 if (mp->mp_ptrs[i] <= ptr)
6981 mp->mp_ptrs[i] -= delta;
6984 base = (char *)mp + mp->mp_upper + MP_HIBASE;
6985 len = ptr - mp->mp_upper + NODESIZE;
6986 memmove(base - delta, base, len);
6987 mp->mp_upper -= delta;
6989 node = NODEPTR(mp, indx);
6992 /* But even if no shift was needed, update ksize */
6993 if (node->mn_ksize != key->mv_size)
6994 node->mn_ksize = key->mv_size;
6997 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7003 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7005 /** Move a node from csrc to cdst.
7008 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7015 unsigned short flags;
7019 /* Mark src and dst as dirty. */
7020 if ((rc = mdb_page_touch(csrc)) ||
7021 (rc = mdb_page_touch(cdst)))
7024 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7025 key.mv_size = csrc->mc_db->md_pad;
7026 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7028 data.mv_data = NULL;
7032 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7033 mdb_cassert(csrc, !((size_t)srcnode & 1));
7034 srcpg = NODEPGNO(srcnode);
7035 flags = srcnode->mn_flags;
7036 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7037 unsigned int snum = csrc->mc_snum;
7039 /* must find the lowest key below src */
7040 rc = mdb_page_search_lowest(csrc);
7043 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7044 key.mv_size = csrc->mc_db->md_pad;
7045 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7047 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7048 key.mv_size = NODEKSZ(s2);
7049 key.mv_data = NODEKEY(s2);
7051 csrc->mc_snum = snum--;
7052 csrc->mc_top = snum;
7054 key.mv_size = NODEKSZ(srcnode);
7055 key.mv_data = NODEKEY(srcnode);
7057 data.mv_size = NODEDSZ(srcnode);
7058 data.mv_data = NODEDATA(srcnode);
7060 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7061 unsigned int snum = cdst->mc_snum;
7064 /* must find the lowest key below dst */
7065 mdb_cursor_copy(cdst, &mn);
7066 rc = mdb_page_search_lowest(&mn);
7069 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7070 bkey.mv_size = mn.mc_db->md_pad;
7071 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7073 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7074 bkey.mv_size = NODEKSZ(s2);
7075 bkey.mv_data = NODEKEY(s2);
7077 mn.mc_snum = snum--;
7080 rc = mdb_update_key(&mn, &bkey);
7085 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7086 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7087 csrc->mc_ki[csrc->mc_top],
7089 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7090 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7092 /* Add the node to the destination page.
7094 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7095 if (rc != MDB_SUCCESS)
7098 /* Delete the node from the source page.
7100 mdb_node_del(csrc, key.mv_size);
7103 /* Adjust other cursors pointing to mp */
7104 MDB_cursor *m2, *m3;
7105 MDB_dbi dbi = csrc->mc_dbi;
7106 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7108 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7109 if (csrc->mc_flags & C_SUB)
7110 m3 = &m2->mc_xcursor->mx_cursor;
7113 if (m3 == csrc) continue;
7114 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7115 csrc->mc_ki[csrc->mc_top]) {
7116 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7117 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7122 /* Update the parent separators.
7124 if (csrc->mc_ki[csrc->mc_top] == 0) {
7125 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7126 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7127 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7129 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7130 key.mv_size = NODEKSZ(srcnode);
7131 key.mv_data = NODEKEY(srcnode);
7133 DPRINTF(("update separator for source page %"Z"u to [%s]",
7134 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7135 mdb_cursor_copy(csrc, &mn);
7138 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7141 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7143 indx_t ix = csrc->mc_ki[csrc->mc_top];
7144 nullkey.mv_size = 0;
7145 csrc->mc_ki[csrc->mc_top] = 0;
7146 rc = mdb_update_key(csrc, &nullkey);
7147 csrc->mc_ki[csrc->mc_top] = ix;
7148 mdb_cassert(csrc, rc == MDB_SUCCESS);
7152 if (cdst->mc_ki[cdst->mc_top] == 0) {
7153 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7154 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7155 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7157 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7158 key.mv_size = NODEKSZ(srcnode);
7159 key.mv_data = NODEKEY(srcnode);
7161 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7162 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7163 mdb_cursor_copy(cdst, &mn);
7166 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7169 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7171 indx_t ix = cdst->mc_ki[cdst->mc_top];
7172 nullkey.mv_size = 0;
7173 cdst->mc_ki[cdst->mc_top] = 0;
7174 rc = mdb_update_key(cdst, &nullkey);
7175 cdst->mc_ki[cdst->mc_top] = ix;
7176 mdb_cassert(csrc, rc == MDB_SUCCESS);
7183 /** Merge one page into another.
7184 * The nodes from the page pointed to by \b csrc will
7185 * be copied to the page pointed to by \b cdst and then
7186 * the \b csrc page will be freed.
7187 * @param[in] csrc Cursor pointing to the source page.
7188 * @param[in] cdst Cursor pointing to the destination page.
7189 * @return 0 on success, non-zero on failure.
7192 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7194 MDB_page *psrc, *pdst;
7201 psrc = csrc->mc_pg[csrc->mc_top];
7202 pdst = cdst->mc_pg[cdst->mc_top];
7204 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7206 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7207 mdb_cassert(csrc, cdst->mc_snum > 1);
7209 /* Mark dst as dirty. */
7210 if ((rc = mdb_page_touch(cdst)))
7213 /* Move all nodes from src to dst.
7215 j = nkeys = NUMKEYS(pdst);
7216 if (IS_LEAF2(psrc)) {
7217 key.mv_size = csrc->mc_db->md_pad;
7218 key.mv_data = METADATA(psrc);
7219 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7220 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7221 if (rc != MDB_SUCCESS)
7223 key.mv_data = (char *)key.mv_data + key.mv_size;
7226 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7227 srcnode = NODEPTR(psrc, i);
7228 if (i == 0 && IS_BRANCH(psrc)) {
7231 mdb_cursor_copy(csrc, &mn);
7232 /* must find the lowest key below src */
7233 rc = mdb_page_search_lowest(&mn);
7236 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7237 key.mv_size = mn.mc_db->md_pad;
7238 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7240 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7241 key.mv_size = NODEKSZ(s2);
7242 key.mv_data = NODEKEY(s2);
7245 key.mv_size = srcnode->mn_ksize;
7246 key.mv_data = NODEKEY(srcnode);
7249 data.mv_size = NODEDSZ(srcnode);
7250 data.mv_data = NODEDATA(srcnode);
7251 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7252 if (rc != MDB_SUCCESS)
7257 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7258 pdst->mp_pgno, NUMKEYS(pdst),
7259 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7261 /* Unlink the src page from parent and add to free list.
7264 mdb_node_del(csrc, 0);
7265 if (csrc->mc_ki[csrc->mc_top] == 0) {
7267 rc = mdb_update_key(csrc, &key);
7275 psrc = csrc->mc_pg[csrc->mc_top];
7276 /* If not operating on FreeDB, allow this page to be reused
7277 * in this txn. Otherwise just add to free list.
7279 rc = mdb_page_loose(csrc, psrc);
7283 csrc->mc_db->md_leaf_pages--;
7285 csrc->mc_db->md_branch_pages--;
7287 /* Adjust other cursors pointing to mp */
7288 MDB_cursor *m2, *m3;
7289 MDB_dbi dbi = csrc->mc_dbi;
7291 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7292 if (csrc->mc_flags & C_SUB)
7293 m3 = &m2->mc_xcursor->mx_cursor;
7296 if (m3 == csrc) continue;
7297 if (m3->mc_snum < csrc->mc_snum) continue;
7298 if (m3->mc_pg[csrc->mc_top] == psrc) {
7299 m3->mc_pg[csrc->mc_top] = pdst;
7300 m3->mc_ki[csrc->mc_top] += nkeys;
7305 unsigned int snum = cdst->mc_snum;
7306 uint16_t depth = cdst->mc_db->md_depth;
7307 mdb_cursor_pop(cdst);
7308 rc = mdb_rebalance(cdst);
7309 /* Did the tree shrink? */
7310 if (depth > cdst->mc_db->md_depth)
7312 cdst->mc_snum = snum;
7313 cdst->mc_top = snum-1;
7318 /** Copy the contents of a cursor.
7319 * @param[in] csrc The cursor to copy from.
7320 * @param[out] cdst The cursor to copy to.
7323 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7327 cdst->mc_txn = csrc->mc_txn;
7328 cdst->mc_dbi = csrc->mc_dbi;
7329 cdst->mc_db = csrc->mc_db;
7330 cdst->mc_dbx = csrc->mc_dbx;
7331 cdst->mc_snum = csrc->mc_snum;
7332 cdst->mc_top = csrc->mc_top;
7333 cdst->mc_flags = csrc->mc_flags;
7335 for (i=0; i<csrc->mc_snum; i++) {
7336 cdst->mc_pg[i] = csrc->mc_pg[i];
7337 cdst->mc_ki[i] = csrc->mc_ki[i];
7341 /** Rebalance the tree after a delete operation.
7342 * @param[in] mc Cursor pointing to the page where rebalancing
7344 * @return 0 on success, non-zero on failure.
7347 mdb_rebalance(MDB_cursor *mc)
7351 unsigned int ptop, minkeys;
7355 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7356 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7357 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7358 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7359 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7361 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7362 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7363 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7364 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7368 if (mc->mc_snum < 2) {
7369 MDB_page *mp = mc->mc_pg[0];
7371 DPUTS("Can't rebalance a subpage, ignoring");
7374 if (NUMKEYS(mp) == 0) {
7375 DPUTS("tree is completely empty");
7376 mc->mc_db->md_root = P_INVALID;
7377 mc->mc_db->md_depth = 0;
7378 mc->mc_db->md_leaf_pages = 0;
7379 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7382 /* Adjust cursors pointing to mp */
7385 mc->mc_flags &= ~C_INITIALIZED;
7387 MDB_cursor *m2, *m3;
7388 MDB_dbi dbi = mc->mc_dbi;
7390 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7391 if (mc->mc_flags & C_SUB)
7392 m3 = &m2->mc_xcursor->mx_cursor;
7395 if (m3->mc_snum < mc->mc_snum) continue;
7396 if (m3->mc_pg[0] == mp) {
7399 m3->mc_flags &= ~C_INITIALIZED;
7403 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7405 DPUTS("collapsing root page!");
7406 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7409 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7410 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7413 mc->mc_db->md_depth--;
7414 mc->mc_db->md_branch_pages--;
7415 mc->mc_ki[0] = mc->mc_ki[1];
7416 for (i = 1; i<mc->mc_db->md_depth; i++) {
7417 mc->mc_pg[i] = mc->mc_pg[i+1];
7418 mc->mc_ki[i] = mc->mc_ki[i+1];
7421 /* Adjust other cursors pointing to mp */
7422 MDB_cursor *m2, *m3;
7423 MDB_dbi dbi = mc->mc_dbi;
7425 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7426 if (mc->mc_flags & C_SUB)
7427 m3 = &m2->mc_xcursor->mx_cursor;
7430 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7431 if (m3->mc_pg[0] == mp) {
7434 for (i=0; i<m3->mc_snum; i++) {
7435 m3->mc_pg[i] = m3->mc_pg[i+1];
7436 m3->mc_ki[i] = m3->mc_ki[i+1];
7442 DPUTS("root page doesn't need rebalancing");
7446 /* The parent (branch page) must have at least 2 pointers,
7447 * otherwise the tree is invalid.
7449 ptop = mc->mc_top-1;
7450 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7452 /* Leaf page fill factor is below the threshold.
7453 * Try to move keys from left or right neighbor, or
7454 * merge with a neighbor page.
7459 mdb_cursor_copy(mc, &mn);
7460 mn.mc_xcursor = NULL;
7462 oldki = mc->mc_ki[mc->mc_top];
7463 if (mc->mc_ki[ptop] == 0) {
7464 /* We're the leftmost leaf in our parent.
7466 DPUTS("reading right neighbor");
7468 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7469 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7472 mn.mc_ki[mn.mc_top] = 0;
7473 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7475 /* There is at least one neighbor to the left.
7477 DPUTS("reading left neighbor");
7479 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7480 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7483 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7484 mc->mc_ki[mc->mc_top] = 0;
7487 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7488 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7489 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7491 /* If the neighbor page is above threshold and has enough keys,
7492 * move one key from it. Otherwise we should try to merge them.
7493 * (A branch page must never have less than 2 keys.)
7495 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7496 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7497 rc = mdb_node_move(&mn, mc);
7498 if (mc->mc_ki[ptop]) {
7502 if (mc->mc_ki[ptop] == 0) {
7503 rc = mdb_page_merge(&mn, mc);
7505 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7506 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7507 rc = mdb_page_merge(mc, &mn);
7508 mdb_cursor_copy(&mn, mc);
7510 mc->mc_flags &= ~C_EOF;
7512 mc->mc_ki[mc->mc_top] = oldki;
7516 /** Complete a delete operation started by #mdb_cursor_del(). */
7518 mdb_cursor_del0(MDB_cursor *mc)
7525 ki = mc->mc_ki[mc->mc_top];
7526 mdb_node_del(mc, mc->mc_db->md_pad);
7527 mc->mc_db->md_entries--;
7528 rc = mdb_rebalance(mc);
7530 if (rc == MDB_SUCCESS) {
7531 MDB_cursor *m2, *m3;
7532 MDB_dbi dbi = mc->mc_dbi;
7534 mp = mc->mc_pg[mc->mc_top];
7535 nkeys = NUMKEYS(mp);
7537 /* if mc points past last node in page, find next sibling */
7538 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7539 rc = mdb_cursor_sibling(mc, 1);
7540 if (rc == MDB_NOTFOUND)
7544 /* Adjust other cursors pointing to mp */
7545 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7546 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7547 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7549 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7551 if (m3->mc_pg[mc->mc_top] == mp) {
7552 if (m3->mc_ki[mc->mc_top] >= ki) {
7553 m3->mc_flags |= C_DEL;
7554 if (m3->mc_ki[mc->mc_top] > ki)
7555 m3->mc_ki[mc->mc_top]--;
7557 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7558 rc = mdb_cursor_sibling(m3, 1);
7559 if (rc == MDB_NOTFOUND)
7564 mc->mc_flags |= C_DEL;
7568 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7573 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7574 MDB_val *key, MDB_val *data)
7576 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7579 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7580 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7582 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7583 /* must ignore any data */
7587 return mdb_del0(txn, dbi, key, data, 0);
7591 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7592 MDB_val *key, MDB_val *data, unsigned flags)
7597 MDB_val rdata, *xdata;
7601 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7603 mdb_cursor_init(&mc, txn, dbi, &mx);
7612 flags |= MDB_NODUPDATA;
7614 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7616 /* let mdb_page_split know about this cursor if needed:
7617 * delete will trigger a rebalance; if it needs to move
7618 * a node from one page to another, it will have to
7619 * update the parent's separator key(s). If the new sepkey
7620 * is larger than the current one, the parent page may
7621 * run out of space, triggering a split. We need this
7622 * cursor to be consistent until the end of the rebalance.
7624 mc.mc_flags |= C_UNTRACK;
7625 mc.mc_next = txn->mt_cursors[dbi];
7626 txn->mt_cursors[dbi] = &mc;
7627 rc = mdb_cursor_del(&mc, flags);
7628 txn->mt_cursors[dbi] = mc.mc_next;
7633 /** Split a page and insert a new node.
7634 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7635 * The cursor will be updated to point to the actual page and index where
7636 * the node got inserted after the split.
7637 * @param[in] newkey The key for the newly inserted node.
7638 * @param[in] newdata The data for the newly inserted node.
7639 * @param[in] newpgno The page number, if the new node is a branch node.
7640 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7641 * @return 0 on success, non-zero on failure.
7644 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7645 unsigned int nflags)
7648 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7651 int i, j, split_indx, nkeys, pmax;
7652 MDB_env *env = mc->mc_txn->mt_env;
7654 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7655 MDB_page *copy = NULL;
7656 MDB_page *mp, *rp, *pp;
7661 mp = mc->mc_pg[mc->mc_top];
7662 newindx = mc->mc_ki[mc->mc_top];
7663 nkeys = NUMKEYS(mp);
7665 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7666 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7667 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7669 /* Create a right sibling. */
7670 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7672 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7674 if (mc->mc_snum < 2) {
7675 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7677 /* shift current top to make room for new parent */
7678 mc->mc_pg[1] = mc->mc_pg[0];
7679 mc->mc_ki[1] = mc->mc_ki[0];
7682 mc->mc_db->md_root = pp->mp_pgno;
7683 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7684 mc->mc_db->md_depth++;
7687 /* Add left (implicit) pointer. */
7688 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7689 /* undo the pre-push */
7690 mc->mc_pg[0] = mc->mc_pg[1];
7691 mc->mc_ki[0] = mc->mc_ki[1];
7692 mc->mc_db->md_root = mp->mp_pgno;
7693 mc->mc_db->md_depth--;
7700 ptop = mc->mc_top-1;
7701 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7704 mc->mc_flags |= C_SPLITTING;
7705 mdb_cursor_copy(mc, &mn);
7706 mn.mc_pg[mn.mc_top] = rp;
7707 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7709 if (nflags & MDB_APPEND) {
7710 mn.mc_ki[mn.mc_top] = 0;
7712 split_indx = newindx;
7716 split_indx = (nkeys+1) / 2;
7721 unsigned int lsize, rsize, ksize;
7722 /* Move half of the keys to the right sibling */
7723 x = mc->mc_ki[mc->mc_top] - split_indx;
7724 ksize = mc->mc_db->md_pad;
7725 split = LEAF2KEY(mp, split_indx, ksize);
7726 rsize = (nkeys - split_indx) * ksize;
7727 lsize = (nkeys - split_indx) * sizeof(indx_t);
7728 mp->mp_lower -= lsize;
7729 rp->mp_lower += lsize;
7730 mp->mp_upper += rsize - lsize;
7731 rp->mp_upper -= rsize - lsize;
7732 sepkey.mv_size = ksize;
7733 if (newindx == split_indx) {
7734 sepkey.mv_data = newkey->mv_data;
7736 sepkey.mv_data = split;
7739 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7740 memcpy(rp->mp_ptrs, split, rsize);
7741 sepkey.mv_data = rp->mp_ptrs;
7742 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7743 memcpy(ins, newkey->mv_data, ksize);
7744 mp->mp_lower += sizeof(indx_t);
7745 mp->mp_upper -= ksize - sizeof(indx_t);
7748 memcpy(rp->mp_ptrs, split, x * ksize);
7749 ins = LEAF2KEY(rp, x, ksize);
7750 memcpy(ins, newkey->mv_data, ksize);
7751 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7752 rp->mp_lower += sizeof(indx_t);
7753 rp->mp_upper -= ksize - sizeof(indx_t);
7754 mc->mc_ki[mc->mc_top] = x;
7755 mc->mc_pg[mc->mc_top] = rp;
7758 int psize, nsize, k;
7759 /* Maximum free space in an empty page */
7760 pmax = env->me_psize - PAGEHDRSZ;
7762 nsize = mdb_leaf_size(env, newkey, newdata);
7764 nsize = mdb_branch_size(env, newkey);
7765 nsize = EVEN(nsize);
7767 /* grab a page to hold a temporary copy */
7768 copy = mdb_page_malloc(mc->mc_txn, 1);
7773 copy->mp_pgno = mp->mp_pgno;
7774 copy->mp_flags = mp->mp_flags;
7775 copy->mp_lower = MP_LOBASE;
7776 copy->mp_upper = env->me_psize - MP_HIBASE;
7778 /* prepare to insert */
7779 for (i=0, j=0; i<nkeys; i++) {
7781 copy->mp_ptrs[j++] = 0;
7783 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7786 /* When items are relatively large the split point needs
7787 * to be checked, because being off-by-one will make the
7788 * difference between success or failure in mdb_node_add.
7790 * It's also relevant if a page happens to be laid out
7791 * such that one half of its nodes are all "small" and
7792 * the other half of its nodes are "large." If the new
7793 * item is also "large" and falls on the half with
7794 * "large" nodes, it also may not fit.
7796 * As a final tweak, if the new item goes on the last
7797 * spot on the page (and thus, onto the new page), bias
7798 * the split so the new page is emptier than the old page.
7799 * This yields better packing during sequential inserts.
7801 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7802 /* Find split point */
7804 if (newindx <= split_indx || newindx >= nkeys) {
7806 k = newindx >= nkeys ? nkeys : split_indx+2;
7811 for (; i!=k; i+=j) {
7816 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + MP_HIBASE);
7817 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7819 if (F_ISSET(node->mn_flags, F_BIGDATA))
7820 psize += sizeof(pgno_t);
7822 psize += NODEDSZ(node);
7824 psize = EVEN(psize);
7826 if (psize > pmax || i == k-j) {
7827 split_indx = i + (j<0);
7832 if (split_indx == newindx) {
7833 sepkey.mv_size = newkey->mv_size;
7834 sepkey.mv_data = newkey->mv_data;
7836 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + MP_HIBASE);
7837 sepkey.mv_size = node->mn_ksize;
7838 sepkey.mv_data = NODEKEY(node);
7843 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7845 /* Copy separator key to the parent.
7847 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7851 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7856 if (mn.mc_snum == mc->mc_snum) {
7857 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7858 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7859 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7860 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7865 /* Right page might now have changed parent.
7866 * Check if left page also changed parent.
7868 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7869 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7870 for (i=0; i<ptop; i++) {
7871 mc->mc_pg[i] = mn.mc_pg[i];
7872 mc->mc_ki[i] = mn.mc_ki[i];
7874 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7875 if (mn.mc_ki[ptop]) {
7876 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7878 /* find right page's left sibling */
7879 mc->mc_ki[ptop] = mn.mc_ki[ptop];
7880 mdb_cursor_sibling(mc, 0);
7885 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7888 mc->mc_flags ^= C_SPLITTING;
7889 if (rc != MDB_SUCCESS) {
7892 if (nflags & MDB_APPEND) {
7893 mc->mc_pg[mc->mc_top] = rp;
7894 mc->mc_ki[mc->mc_top] = 0;
7895 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7898 for (i=0; i<mc->mc_top; i++)
7899 mc->mc_ki[i] = mn.mc_ki[i];
7900 } else if (!IS_LEAF2(mp)) {
7902 mc->mc_pg[mc->mc_top] = rp;
7907 rkey.mv_data = newkey->mv_data;
7908 rkey.mv_size = newkey->mv_size;
7914 /* Update index for the new key. */
7915 mc->mc_ki[mc->mc_top] = j;
7917 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + MP_HIBASE);
7918 rkey.mv_data = NODEKEY(node);
7919 rkey.mv_size = node->mn_ksize;
7921 xdata.mv_data = NODEDATA(node);
7922 xdata.mv_size = NODEDSZ(node);
7925 pgno = NODEPGNO(node);
7926 flags = node->mn_flags;
7929 if (!IS_LEAF(mp) && j == 0) {
7930 /* First branch index doesn't need key data. */
7934 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7940 mc->mc_pg[mc->mc_top] = copy;
7945 } while (i != split_indx);
7947 nkeys = NUMKEYS(copy);
7948 for (i=0; i<nkeys; i++)
7949 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7950 mp->mp_lower = copy->mp_lower;
7951 mp->mp_upper = copy->mp_upper;
7952 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7953 env->me_psize - copy->mp_upper - MP_HIBASE);
7955 /* reset back to original page */
7956 if (newindx < split_indx) {
7957 mc->mc_pg[mc->mc_top] = mp;
7958 if (nflags & MDB_RESERVE) {
7959 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7960 if (!(node->mn_flags & F_BIGDATA))
7961 newdata->mv_data = NODEDATA(node);
7964 mc->mc_pg[mc->mc_top] = rp;
7966 /* Make sure mc_ki is still valid.
7968 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7969 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7970 for (i=0; i<=ptop; i++) {
7971 mc->mc_pg[i] = mn.mc_pg[i];
7972 mc->mc_ki[i] = mn.mc_ki[i];
7979 /* Adjust other cursors pointing to mp */
7980 MDB_cursor *m2, *m3;
7981 MDB_dbi dbi = mc->mc_dbi;
7982 int fixup = NUMKEYS(mp);
7984 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7985 if (mc->mc_flags & C_SUB)
7986 m3 = &m2->mc_xcursor->mx_cursor;
7991 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7993 if (m3->mc_flags & C_SPLITTING)
7998 for (k=m3->mc_top; k>=0; k--) {
7999 m3->mc_ki[k+1] = m3->mc_ki[k];
8000 m3->mc_pg[k+1] = m3->mc_pg[k];
8002 if (m3->mc_ki[0] >= split_indx) {
8007 m3->mc_pg[0] = mc->mc_pg[0];
8011 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8012 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8013 m3->mc_ki[mc->mc_top]++;
8014 if (m3->mc_ki[mc->mc_top] >= fixup) {
8015 m3->mc_pg[mc->mc_top] = rp;
8016 m3->mc_ki[mc->mc_top] -= fixup;
8017 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8019 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8020 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8025 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8028 if (copy) /* tmp page */
8029 mdb_page_free(env, copy);
8031 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8036 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8037 MDB_val *key, MDB_val *data, unsigned int flags)
8042 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8045 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8048 mdb_cursor_init(&mc, txn, dbi, &mx);
8049 return mdb_cursor_put(&mc, key, data, flags);
8053 #define MDB_WBUF (1024*1024)
8056 /** State needed for a compacting copy. */
8057 typedef struct mdb_copy {
8058 pthread_mutex_t mc_mutex;
8059 pthread_cond_t mc_cond;
8066 pgno_t mc_next_pgno;
8069 volatile int mc_new;
8074 /** Dedicated writer thread for compacting copy. */
8075 static THREAD_RET ESECT
8076 mdb_env_copythr(void *arg)
8080 int toggle = 0, wsize, rc;
8083 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8086 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8089 pthread_mutex_lock(&my->mc_mutex);
8091 pthread_cond_signal(&my->mc_cond);
8094 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8095 if (my->mc_new < 0) {
8100 wsize = my->mc_wlen[toggle];
8101 ptr = my->mc_wbuf[toggle];
8104 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8108 } else if (len > 0) {
8122 /* If there's an overflow page tail, write it too */
8123 if (my->mc_olen[toggle]) {
8124 wsize = my->mc_olen[toggle];
8125 ptr = my->mc_over[toggle];
8126 my->mc_olen[toggle] = 0;
8129 my->mc_wlen[toggle] = 0;
8131 pthread_cond_signal(&my->mc_cond);
8133 pthread_cond_signal(&my->mc_cond);
8134 pthread_mutex_unlock(&my->mc_mutex);
8135 return (THREAD_RET)0;
8139 /** Tell the writer thread there's a buffer ready to write */
8141 mdb_env_cthr_toggle(mdb_copy *my, int st)
8143 int toggle = my->mc_toggle ^ 1;
8144 pthread_mutex_lock(&my->mc_mutex);
8145 if (my->mc_status) {
8146 pthread_mutex_unlock(&my->mc_mutex);
8147 return my->mc_status;
8149 while (my->mc_new == 1)
8150 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8152 my->mc_toggle = toggle;
8153 pthread_cond_signal(&my->mc_cond);
8154 pthread_mutex_unlock(&my->mc_mutex);
8158 /** Depth-first tree traversal for compacting copy. */
8160 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8163 MDB_txn *txn = my->mc_txn;
8165 MDB_page *mo, *mp, *leaf;
8170 /* Empty DB, nothing to do */
8171 if (*pg == P_INVALID)
8178 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8181 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8185 /* Make cursor pages writable */
8186 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8190 for (i=0; i<mc.mc_top; i++) {
8191 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8192 mc.mc_pg[i] = (MDB_page *)ptr;
8193 ptr += my->mc_env->me_psize;
8196 /* This is writable space for a leaf page. Usually not needed. */
8197 leaf = (MDB_page *)ptr;
8199 toggle = my->mc_toggle;
8200 while (mc.mc_snum > 0) {
8202 mp = mc.mc_pg[mc.mc_top];
8206 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8207 for (i=0; i<n; i++) {
8208 ni = NODEPTR(mp, i);
8209 if (ni->mn_flags & F_BIGDATA) {
8213 /* Need writable leaf */
8215 mc.mc_pg[mc.mc_top] = leaf;
8216 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8218 ni = NODEPTR(mp, i);
8221 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8222 rc = mdb_page_get(txn, pg, &omp, NULL);
8225 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8226 rc = mdb_env_cthr_toggle(my, 1);
8229 toggle = my->mc_toggle;
8231 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8232 memcpy(mo, omp, my->mc_env->me_psize);
8233 mo->mp_pgno = my->mc_next_pgno;
8234 my->mc_next_pgno += omp->mp_pages;
8235 my->mc_wlen[toggle] += my->mc_env->me_psize;
8236 if (omp->mp_pages > 1) {
8237 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8238 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8239 rc = mdb_env_cthr_toggle(my, 1);
8242 toggle = my->mc_toggle;
8244 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8245 } else if (ni->mn_flags & F_SUBDATA) {
8248 /* Need writable leaf */
8250 mc.mc_pg[mc.mc_top] = leaf;
8251 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8253 ni = NODEPTR(mp, i);
8256 memcpy(&db, NODEDATA(ni), sizeof(db));
8257 my->mc_toggle = toggle;
8258 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8261 toggle = my->mc_toggle;
8262 memcpy(NODEDATA(ni), &db, sizeof(db));
8267 mc.mc_ki[mc.mc_top]++;
8268 if (mc.mc_ki[mc.mc_top] < n) {
8271 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8273 rc = mdb_page_get(txn, pg, &mp, NULL);
8278 mc.mc_ki[mc.mc_top] = 0;
8279 if (IS_BRANCH(mp)) {
8280 /* Whenever we advance to a sibling branch page,
8281 * we must proceed all the way down to its first leaf.
8283 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8286 mc.mc_pg[mc.mc_top] = mp;
8290 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8291 rc = mdb_env_cthr_toggle(my, 1);
8294 toggle = my->mc_toggle;
8296 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8297 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8298 mo->mp_pgno = my->mc_next_pgno++;
8299 my->mc_wlen[toggle] += my->mc_env->me_psize;
8301 /* Update parent if there is one */
8302 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8303 SETPGNO(ni, mo->mp_pgno);
8304 mdb_cursor_pop(&mc);
8306 /* Otherwise we're done */
8316 /** Copy environment with compaction. */
8318 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8323 MDB_txn *txn = NULL;
8328 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8329 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8330 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_psize);
8331 if (my.mc_wbuf[0] == NULL)
8334 pthread_mutex_init(&my.mc_mutex, NULL);
8335 pthread_cond_init(&my.mc_cond, NULL);
8336 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_psize, MDB_WBUF*2);
8340 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8345 my.mc_next_pgno = 2;
8351 THREAD_CREATE(thr, mdb_env_copythr, &my);
8353 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8357 mp = (MDB_page *)my.mc_wbuf[0];
8358 memset(mp, 0, 2*env->me_psize);
8360 mp->mp_flags = P_META;
8361 mm = (MDB_meta *)METADATA(mp);
8362 mdb_env_init_meta0(env, mm);
8363 mm->mm_address = env->me_metas[0]->mm_address;
8365 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8367 mp->mp_flags = P_META;
8368 *(MDB_meta *)METADATA(mp) = *mm;
8369 mm = (MDB_meta *)METADATA(mp);
8371 /* Count the number of free pages, subtract from lastpg to find
8372 * number of active pages
8375 MDB_ID freecount = 0;
8378 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8379 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8380 freecount += *(MDB_ID *)data.mv_data;
8381 freecount += txn->mt_dbs[0].md_branch_pages +
8382 txn->mt_dbs[0].md_leaf_pages +
8383 txn->mt_dbs[0].md_overflow_pages;
8385 /* Set metapage 1 */
8386 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8387 mm->mm_dbs[1] = txn->mt_dbs[1];
8388 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8391 my.mc_wlen[0] = env->me_psize * 2;
8393 pthread_mutex_lock(&my.mc_mutex);
8395 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8396 pthread_mutex_unlock(&my.mc_mutex);
8397 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8398 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8399 rc = mdb_env_cthr_toggle(&my, 1);
8400 mdb_env_cthr_toggle(&my, -1);
8401 pthread_mutex_lock(&my.mc_mutex);
8403 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8404 pthread_mutex_unlock(&my.mc_mutex);
8409 CloseHandle(my.mc_cond);
8410 CloseHandle(my.mc_mutex);
8411 _aligned_free(my.mc_wbuf[0]);
8413 pthread_cond_destroy(&my.mc_cond);
8414 pthread_mutex_destroy(&my.mc_mutex);
8415 free(my.mc_wbuf[0]);
8420 /** Copy environment as-is. */
8422 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8424 MDB_txn *txn = NULL;
8430 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8434 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8437 /* Do the lock/unlock of the reader mutex before starting the
8438 * write txn. Otherwise other read txns could block writers.
8440 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8445 /* We must start the actual read txn after blocking writers */
8446 mdb_txn_reset0(txn, "reset-stage1");
8448 /* Temporarily block writers until we snapshot the meta pages */
8451 rc = mdb_txn_renew0(txn);
8453 UNLOCK_MUTEX_W(env);
8458 wsize = env->me_psize * 2;
8462 DO_WRITE(rc, fd, ptr, w2, len);
8466 } else if (len > 0) {
8472 /* Non-blocking or async handles are not supported */
8478 UNLOCK_MUTEX_W(env);
8483 w2 = txn->mt_next_pgno * env->me_psize;
8486 LARGE_INTEGER fsize;
8487 GetFileSizeEx(env->me_fd, &fsize);
8488 if (w2 > fsize.QuadPart)
8489 w2 = fsize.QuadPart;
8494 fstat(env->me_fd, &st);
8495 if (w2 > (size_t)st.st_size)
8501 if (wsize > MAX_WRITE)
8505 DO_WRITE(rc, fd, ptr, w2, len);
8509 } else if (len > 0) {
8526 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8528 if (flags & MDB_CP_COMPACT)
8529 return mdb_env_copyfd1(env, fd);
8531 return mdb_env_copyfd0(env, fd);
8535 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8537 return mdb_env_copyfd2(env, fd, 0);
8541 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8545 HANDLE newfd = INVALID_HANDLE_VALUE;
8547 if (env->me_flags & MDB_NOSUBDIR) {
8548 lpath = (char *)path;
8551 len += sizeof(DATANAME);
8552 lpath = malloc(len);
8555 sprintf(lpath, "%s" DATANAME, path);
8558 /* The destination path must exist, but the destination file must not.
8559 * We don't want the OS to cache the writes, since the source data is
8560 * already in the OS cache.
8563 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8564 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8566 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8568 if (newfd == INVALID_HANDLE_VALUE) {
8574 /* Set O_DIRECT if the file system supports it */
8575 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8576 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8578 #ifdef F_NOCACHE /* __APPLE__ */
8579 rc = fcntl(newfd, F_NOCACHE, 1);
8586 rc = mdb_env_copyfd2(env, newfd, flags);
8589 if (!(env->me_flags & MDB_NOSUBDIR))
8591 if (newfd != INVALID_HANDLE_VALUE)
8592 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8599 mdb_env_copy(MDB_env *env, const char *path)
8601 return mdb_env_copy2(env, path, 0);
8605 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8607 if ((flag & CHANGEABLE) != flag)
8610 env->me_flags |= flag;
8612 env->me_flags &= ~flag;
8617 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8622 *arg = env->me_flags;
8627 mdb_env_set_userctx(MDB_env *env, void *ctx)
8631 env->me_userctx = ctx;
8636 mdb_env_get_userctx(MDB_env *env)
8638 return env ? env->me_userctx : NULL;
8642 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8647 env->me_assert_func = func;
8653 mdb_env_get_path(MDB_env *env, const char **arg)
8658 *arg = env->me_path;
8663 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8672 /** Common code for #mdb_stat() and #mdb_env_stat().
8673 * @param[in] env the environment to operate in.
8674 * @param[in] db the #MDB_db record containing the stats to return.
8675 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8676 * @return 0, this function always succeeds.
8679 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8681 arg->ms_psize = env->me_psize;
8682 arg->ms_depth = db->md_depth;
8683 arg->ms_branch_pages = db->md_branch_pages;
8684 arg->ms_leaf_pages = db->md_leaf_pages;
8685 arg->ms_overflow_pages = db->md_overflow_pages;
8686 arg->ms_entries = db->md_entries;
8692 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8696 if (env == NULL || arg == NULL)
8699 toggle = mdb_env_pick_meta(env);
8701 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8705 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8709 if (env == NULL || arg == NULL)
8712 toggle = mdb_env_pick_meta(env);
8713 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8714 arg->me_mapsize = env->me_mapsize;
8715 arg->me_maxreaders = env->me_maxreaders;
8717 /* me_numreaders may be zero if this process never used any readers. Use
8718 * the shared numreader count if it exists.
8720 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8722 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8723 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8727 /** Set the default comparison functions for a database.
8728 * Called immediately after a database is opened to set the defaults.
8729 * The user can then override them with #mdb_set_compare() or
8730 * #mdb_set_dupsort().
8731 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8732 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8735 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8737 uint16_t f = txn->mt_dbs[dbi].md_flags;
8739 txn->mt_dbxs[dbi].md_cmp =
8740 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8741 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8743 txn->mt_dbxs[dbi].md_dcmp =
8744 !(f & MDB_DUPSORT) ? 0 :
8745 ((f & MDB_INTEGERDUP)
8746 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8747 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8750 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8755 int rc, dbflag, exact;
8756 unsigned int unused = 0;
8759 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8760 mdb_default_cmp(txn, FREE_DBI);
8763 if ((flags & VALID_FLAGS) != flags)
8765 if (txn->mt_flags & MDB_TXN_ERROR)
8771 if (flags & PERSISTENT_FLAGS) {
8772 uint16_t f2 = flags & PERSISTENT_FLAGS;
8773 /* make sure flag changes get committed */
8774 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8775 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8776 txn->mt_flags |= MDB_TXN_DIRTY;
8779 mdb_default_cmp(txn, MAIN_DBI);
8783 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8784 mdb_default_cmp(txn, MAIN_DBI);
8787 /* Is the DB already open? */
8789 for (i=2; i<txn->mt_numdbs; i++) {
8790 if (!txn->mt_dbxs[i].md_name.mv_size) {
8791 /* Remember this free slot */
8792 if (!unused) unused = i;
8795 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8796 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8802 /* If no free slot and max hit, fail */
8803 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8804 return MDB_DBS_FULL;
8806 /* Cannot mix named databases with some mainDB flags */
8807 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8808 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8810 /* Find the DB info */
8811 dbflag = DB_NEW|DB_VALID;
8814 key.mv_data = (void *)name;
8815 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8816 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8817 if (rc == MDB_SUCCESS) {
8818 /* make sure this is actually a DB */
8819 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8820 if (!(node->mn_flags & F_SUBDATA))
8821 return MDB_INCOMPATIBLE;
8822 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8823 /* Create if requested */
8825 data.mv_size = sizeof(MDB_db);
8826 data.mv_data = &dummy;
8827 memset(&dummy, 0, sizeof(dummy));
8828 dummy.md_root = P_INVALID;
8829 dummy.md_flags = flags & PERSISTENT_FLAGS;
8830 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8834 /* OK, got info, add to table */
8835 if (rc == MDB_SUCCESS) {
8836 unsigned int slot = unused ? unused : txn->mt_numdbs;
8837 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8838 txn->mt_dbxs[slot].md_name.mv_size = len;
8839 txn->mt_dbxs[slot].md_rel = NULL;
8840 txn->mt_dbflags[slot] = dbflag;
8841 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8843 mdb_default_cmp(txn, slot);
8852 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8854 if (!arg || !TXN_DBI_EXIST(txn, dbi))
8857 if (txn->mt_flags & MDB_TXN_ERROR)
8860 if (txn->mt_dbflags[dbi] & DB_STALE) {
8863 /* Stale, must read the DB's root. cursor_init does it for us. */
8864 mdb_cursor_init(&mc, txn, dbi, &mx);
8866 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8869 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8872 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8874 ptr = env->me_dbxs[dbi].md_name.mv_data;
8875 env->me_dbxs[dbi].md_name.mv_data = NULL;
8876 env->me_dbxs[dbi].md_name.mv_size = 0;
8877 env->me_dbflags[dbi] = 0;
8881 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8883 /* We could return the flags for the FREE_DBI too but what's the point? */
8884 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8886 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8890 /** Add all the DB's pages to the free list.
8891 * @param[in] mc Cursor on the DB to free.
8892 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8893 * @return 0 on success, non-zero on failure.
8896 mdb_drop0(MDB_cursor *mc, int subs)
8900 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8901 if (rc == MDB_SUCCESS) {
8902 MDB_txn *txn = mc->mc_txn;
8907 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8908 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8911 mdb_cursor_copy(mc, &mx);
8912 while (mc->mc_snum > 0) {
8913 MDB_page *mp = mc->mc_pg[mc->mc_top];
8914 unsigned n = NUMKEYS(mp);
8916 for (i=0; i<n; i++) {
8917 ni = NODEPTR(mp, i);
8918 if (ni->mn_flags & F_BIGDATA) {
8921 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8922 rc = mdb_page_get(txn, pg, &omp, NULL);
8925 mdb_cassert(mc, IS_OVERFLOW(omp));
8926 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8930 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8931 mdb_xcursor_init1(mc, ni);
8932 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8938 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8940 for (i=0; i<n; i++) {
8942 ni = NODEPTR(mp, i);
8945 mdb_midl_xappend(txn->mt_free_pgs, pg);
8950 mc->mc_ki[mc->mc_top] = i;
8951 rc = mdb_cursor_sibling(mc, 1);
8953 if (rc != MDB_NOTFOUND)
8955 /* no more siblings, go back to beginning
8956 * of previous level.
8960 for (i=1; i<mc->mc_snum; i++) {
8962 mc->mc_pg[i] = mx.mc_pg[i];
8967 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8970 txn->mt_flags |= MDB_TXN_ERROR;
8971 } else if (rc == MDB_NOTFOUND) {
8977 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8979 MDB_cursor *mc, *m2;
8982 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8985 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8988 rc = mdb_cursor_open(txn, dbi, &mc);
8992 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8993 /* Invalidate the dropped DB's cursors */
8994 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8995 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8999 /* Can't delete the main DB */
9000 if (del && dbi > MAIN_DBI) {
9001 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9003 txn->mt_dbflags[dbi] = DB_STALE;
9004 mdb_dbi_close(txn->mt_env, dbi);
9006 txn->mt_flags |= MDB_TXN_ERROR;
9009 /* reset the DB record, mark it dirty */
9010 txn->mt_dbflags[dbi] |= DB_DIRTY;
9011 txn->mt_dbs[dbi].md_depth = 0;
9012 txn->mt_dbs[dbi].md_branch_pages = 0;
9013 txn->mt_dbs[dbi].md_leaf_pages = 0;
9014 txn->mt_dbs[dbi].md_overflow_pages = 0;
9015 txn->mt_dbs[dbi].md_entries = 0;
9016 txn->mt_dbs[dbi].md_root = P_INVALID;
9018 txn->mt_flags |= MDB_TXN_DIRTY;
9021 mdb_cursor_close(mc);
9025 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9027 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9030 txn->mt_dbxs[dbi].md_cmp = cmp;
9034 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9036 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9039 txn->mt_dbxs[dbi].md_dcmp = cmp;
9043 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9045 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9048 txn->mt_dbxs[dbi].md_rel = rel;
9052 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9054 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9057 txn->mt_dbxs[dbi].md_relctx = ctx;
9062 mdb_env_get_maxkeysize(MDB_env *env)
9064 return ENV_MAXKEY(env);
9068 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9070 unsigned int i, rdrs;
9073 int rc = 0, first = 1;
9077 if (!env->me_txns) {
9078 return func("(no reader locks)\n", ctx);
9080 rdrs = env->me_txns->mti_numreaders;
9081 mr = env->me_txns->mti_readers;
9082 for (i=0; i<rdrs; i++) {
9084 txnid_t txnid = mr[i].mr_txnid;
9085 sprintf(buf, txnid == (txnid_t)-1 ?
9086 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9087 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9090 rc = func(" pid thread txnid\n", ctx);
9094 rc = func(buf, ctx);
9100 rc = func("(no active readers)\n", ctx);
9105 /** Insert pid into list if not already present.
9106 * return -1 if already present.
9109 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9111 /* binary search of pid in list */
9113 unsigned cursor = 1;
9115 unsigned n = ids[0];
9118 unsigned pivot = n >> 1;
9119 cursor = base + pivot + 1;
9120 val = pid - ids[cursor];
9125 } else if ( val > 0 ) {
9130 /* found, so it's a duplicate */
9139 for (n = ids[0]; n > cursor; n--)
9146 mdb_reader_check(MDB_env *env, int *dead)
9148 unsigned int i, j, rdrs;
9150 MDB_PID_T *pids, pid;
9159 rdrs = env->me_txns->mti_numreaders;
9160 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9164 mr = env->me_txns->mti_readers;
9165 for (i=0; i<rdrs; i++) {
9166 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9168 if (mdb_pid_insert(pids, pid) == 0) {
9169 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9171 /* Recheck, a new process may have reused pid */
9172 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9173 for (j=i; j<rdrs; j++)
9174 if (mr[j].mr_pid == pid) {
9175 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9176 (unsigned) pid, mr[j].mr_txnid));
9181 UNLOCK_MUTEX_R(env);