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"
153 /** @defgroup internal LMDB Internals
156 /** @defgroup compat Compatibility Macros
157 * A bunch of macros to minimize the amount of platform-specific ifdefs
158 * needed throughout the rest of the code. When the features this library
159 * needs are similar enough to POSIX to be hidden in a one-or-two line
160 * replacement, this macro approach is used.
164 /** Wrapper around __func__, which is a C99 feature */
165 #if __STDC_VERSION__ >= 199901L
166 # define mdb_func_ __func__
167 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
168 # define mdb_func_ __FUNCTION__
170 /* If a debug message says <mdb_unknown>(), update the #if statements above */
171 # define mdb_func_ "<mdb_unknown>"
175 #define MDB_USE_HASH 1
176 #define MDB_PIDLOCK 0
177 #define THREAD_RET DWORD
178 #define pthread_t HANDLE
179 #define pthread_mutex_t HANDLE
180 #define pthread_cond_t HANDLE
181 #define pthread_key_t DWORD
182 #define pthread_self() GetCurrentThreadId()
183 #define pthread_key_create(x,y) \
184 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
185 #define pthread_key_delete(x) TlsFree(x)
186 #define pthread_getspecific(x) TlsGetValue(x)
187 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
188 #define pthread_mutex_unlock(x) ReleaseMutex(x)
189 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
190 #define pthread_cond_signal(x) SetEvent(*x)
191 #define pthread_cond_wait(cond,mutex) SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE)
192 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
193 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
194 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
195 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
196 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
197 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
198 #define getpid() GetCurrentProcessId()
199 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
200 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
201 #define ErrCode() GetLastError()
202 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
203 #define close(fd) (CloseHandle(fd) ? 0 : -1)
204 #define munmap(ptr,len) UnmapViewOfFile(ptr)
205 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
206 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
208 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
212 #define THREAD_RET void *
213 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
214 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
215 #define Z "z" /**< printf format modifier for size_t */
217 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
218 #define MDB_PIDLOCK 1
220 #ifdef MDB_USE_POSIX_SEM
222 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
223 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
224 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
225 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
228 mdb_sem_wait(sem_t *sem)
231 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
236 /** Lock the reader mutex.
238 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
239 /** Unlock the reader mutex.
241 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
243 /** Lock the writer mutex.
244 * Only a single write transaction is allowed at a time. Other writers
245 * will block waiting for this mutex.
247 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
248 /** Unlock the writer mutex.
250 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
251 #endif /* MDB_USE_POSIX_SEM */
253 /** Get the error code for the last failed system function.
255 #define ErrCode() errno
257 /** An abstraction for a file handle.
258 * On POSIX systems file handles are small integers. On Windows
259 * they're opaque pointers.
263 /** A value for an invalid file handle.
264 * Mainly used to initialize file variables and signify that they are
267 #define INVALID_HANDLE_VALUE (-1)
269 /** Get the size of a memory page for the system.
270 * This is the basic size that the platform's memory manager uses, and is
271 * fundamental to the use of memory-mapped files.
273 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
276 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
279 #define MNAME_LEN (sizeof(pthread_mutex_t))
285 /** A flag for opening a file and requesting synchronous data writes.
286 * This is only used when writing a meta page. It's not strictly needed;
287 * we could just do a normal write and then immediately perform a flush.
288 * But if this flag is available it saves us an extra system call.
290 * @note If O_DSYNC is undefined but exists in /usr/include,
291 * preferably set some compiler flag to get the definition.
292 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
295 # define MDB_DSYNC O_DSYNC
299 /** Function for flushing the data of a file. Define this to fsync
300 * if fdatasync() is not supported.
302 #ifndef MDB_FDATASYNC
303 # define MDB_FDATASYNC fdatasync
307 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
318 /** A page number in the database.
319 * Note that 64 bit page numbers are overkill, since pages themselves
320 * already represent 12-13 bits of addressable memory, and the OS will
321 * always limit applications to a maximum of 63 bits of address space.
323 * @note In the #MDB_node structure, we only store 48 bits of this value,
324 * which thus limits us to only 60 bits of addressable data.
326 typedef MDB_ID pgno_t;
328 /** A transaction ID.
329 * See struct MDB_txn.mt_txnid for details.
331 typedef MDB_ID txnid_t;
333 /** @defgroup debug Debug Macros
337 /** Enable debug output. Needs variable argument macros (a C99 feature).
338 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
339 * read from and written to the database (used for free space management).
345 static int mdb_debug;
346 static txnid_t mdb_debug_start;
348 /** Print a debug message with printf formatting.
349 * Requires double parenthesis around 2 or more args.
351 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
352 # define DPRINTF0(fmt, ...) \
353 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
355 # define DPRINTF(args) ((void) 0)
357 /** Print a debug string.
358 * The string is printed literally, with no format processing.
360 #define DPUTS(arg) DPRINTF(("%s", arg))
361 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
363 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
366 /** @brief The maximum size of a database page.
368 * This is 32k, since it must fit in #MDB_page.%mp_upper.
370 * LMDB will use database pages < OS pages if needed.
371 * That causes more I/O in write transactions: The OS must
372 * know (read) the whole page before writing a partial page.
374 * Note that we don't currently support Huge pages. On Linux,
375 * regular data files cannot use Huge pages, and in general
376 * Huge pages aren't actually pageable. We rely on the OS
377 * demand-pager to read our data and page it out when memory
378 * pressure from other processes is high. So until OSs have
379 * actual paging support for Huge pages, they're not viable.
381 #define MAX_PAGESIZE 0x8000
383 /** The minimum number of keys required in a database page.
384 * Setting this to a larger value will place a smaller bound on the
385 * maximum size of a data item. Data items larger than this size will
386 * be pushed into overflow pages instead of being stored directly in
387 * the B-tree node. This value used to default to 4. With a page size
388 * of 4096 bytes that meant that any item larger than 1024 bytes would
389 * go into an overflow page. That also meant that on average 2-3KB of
390 * each overflow page was wasted space. The value cannot be lower than
391 * 2 because then there would no longer be a tree structure. With this
392 * value, items larger than 2KB will go into overflow pages, and on
393 * average only 1KB will be wasted.
395 #define MDB_MINKEYS 2
397 /** A stamp that identifies a file as an LMDB file.
398 * There's nothing special about this value other than that it is easily
399 * recognizable, and it will reflect any byte order mismatches.
401 #define MDB_MAGIC 0xBEEFC0DE
403 /** The version number for a database's datafile format. */
404 #define MDB_DATA_VERSION 1
405 /** The version number for a database's lockfile format. */
406 #define MDB_LOCK_VERSION 1
408 /** @brief The max size of a key we can write, or 0 for dynamic max.
410 * Define this as 0 to compute the max from the page size. 511
411 * is default for backwards compat: liblmdb <= 0.9.10 can break
412 * when modifying a DB with keys/dupsort data bigger than its max.
414 * Data items in an #MDB_DUPSORT database are also limited to
415 * this size, since they're actually keys of a sub-DB. Keys and
416 * #MDB_DUPSORT data items must fit on a node in a regular page.
418 #ifndef MDB_MAXKEYSIZE
419 #define MDB_MAXKEYSIZE 511
422 /** The maximum size of a key we can write to the environment. */
424 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
426 #define ENV_MAXKEY(env) ((env)->me_maxkey)
429 /** @brief The maximum size of a data item.
431 * We only store a 32 bit value for node sizes.
433 #define MAXDATASIZE 0xffffffffUL
436 /** Key size which fits in a #DKBUF.
439 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
442 * This is used for printing a hex dump of a key's contents.
444 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
445 /** Display a key in hex.
447 * Invoke a function to display a key in hex.
449 #define DKEY(x) mdb_dkey(x, kbuf)
455 /** An invalid page number.
456 * Mainly used to denote an empty tree.
458 #define P_INVALID (~(pgno_t)0)
460 /** Test if the flags \b f are set in a flag word \b w. */
461 #define F_ISSET(w, f) (((w) & (f)) == (f))
463 /** Round \b n up to an even number. */
464 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
466 /** Used for offsets within a single page.
467 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
470 typedef uint16_t indx_t;
472 /** Default size of memory map.
473 * This is certainly too small for any actual applications. Apps should always set
474 * the size explicitly using #mdb_env_set_mapsize().
476 #define DEFAULT_MAPSIZE 1048576
478 /** @defgroup readers Reader Lock Table
479 * Readers don't acquire any locks for their data access. Instead, they
480 * simply record their transaction ID in the reader table. The reader
481 * mutex is needed just to find an empty slot in the reader table. The
482 * slot's address is saved in thread-specific data so that subsequent read
483 * transactions started by the same thread need no further locking to proceed.
485 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
487 * No reader table is used if the database is on a read-only filesystem, or
488 * if #MDB_NOLOCK is set.
490 * Since the database uses multi-version concurrency control, readers don't
491 * actually need any locking. This table is used to keep track of which
492 * readers are using data from which old transactions, so that we'll know
493 * when a particular old transaction is no longer in use. Old transactions
494 * that have discarded any data pages can then have those pages reclaimed
495 * for use by a later write transaction.
497 * The lock table is constructed such that reader slots are aligned with the
498 * processor's cache line size. Any slot is only ever used by one thread.
499 * This alignment guarantees that there will be no contention or cache
500 * thrashing as threads update their own slot info, and also eliminates
501 * any need for locking when accessing a slot.
503 * A writer thread will scan every slot in the table to determine the oldest
504 * outstanding reader transaction. Any freed pages older than this will be
505 * reclaimed by the writer. The writer doesn't use any locks when scanning
506 * this table. This means that there's no guarantee that the writer will
507 * see the most up-to-date reader info, but that's not required for correct
508 * operation - all we need is to know the upper bound on the oldest reader,
509 * we don't care at all about the newest reader. So the only consequence of
510 * reading stale information here is that old pages might hang around a
511 * while longer before being reclaimed. That's actually good anyway, because
512 * the longer we delay reclaiming old pages, the more likely it is that a
513 * string of contiguous pages can be found after coalescing old pages from
514 * many old transactions together.
517 /** Number of slots in the reader table.
518 * This value was chosen somewhat arbitrarily. 126 readers plus a
519 * couple mutexes fit exactly into 8KB on my development machine.
520 * Applications should set the table size using #mdb_env_set_maxreaders().
522 #define DEFAULT_READERS 126
524 /** The size of a CPU cache line in bytes. We want our lock structures
525 * aligned to this size to avoid false cache line sharing in the
527 * This value works for most CPUs. For Itanium this should be 128.
533 /** The information we store in a single slot of the reader table.
534 * In addition to a transaction ID, we also record the process and
535 * thread ID that owns a slot, so that we can detect stale information,
536 * e.g. threads or processes that went away without cleaning up.
537 * @note We currently don't check for stale records. We simply re-init
538 * the table when we know that we're the only process opening the
541 typedef struct MDB_rxbody {
542 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
543 * Multiple readers that start at the same time will probably have the
544 * same ID here. Again, it's not important to exclude them from
545 * anything; all we need to know is which version of the DB they
546 * started from so we can avoid overwriting any data used in that
547 * particular version.
550 /** The process ID of the process owning this reader txn. */
552 /** The thread ID of the thread owning this txn. */
556 /** The actual reader record, with cacheline padding. */
557 typedef struct MDB_reader {
560 /** shorthand for mrb_txnid */
561 #define mr_txnid mru.mrx.mrb_txnid
562 #define mr_pid mru.mrx.mrb_pid
563 #define mr_tid mru.mrx.mrb_tid
564 /** cache line alignment */
565 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
569 /** The header for the reader table.
570 * The table resides in a memory-mapped file. (This is a different file
571 * than is used for the main database.)
573 * For POSIX the actual mutexes reside in the shared memory of this
574 * mapped file. On Windows, mutexes are named objects allocated by the
575 * kernel; we store the mutex names in this mapped file so that other
576 * processes can grab them. This same approach is also used on
577 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
578 * process-shared POSIX mutexes. For these cases where a named object
579 * is used, the object name is derived from a 64 bit FNV hash of the
580 * environment pathname. As such, naming collisions are extremely
581 * unlikely. If a collision occurs, the results are unpredictable.
583 typedef struct MDB_txbody {
584 /** Stamp identifying this as an LMDB file. It must be set
587 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
589 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
590 char mtb_rmname[MNAME_LEN];
592 /** Mutex protecting access to this table.
593 * This is the reader lock that #LOCK_MUTEX_R acquires.
595 pthread_mutex_t mtb_mutex;
597 /** The ID of the last transaction committed to the database.
598 * This is recorded here only for convenience; the value can always
599 * be determined by reading the main database meta pages.
602 /** The number of slots that have been used in the reader table.
603 * This always records the maximum count, it is not decremented
604 * when readers release their slots.
606 unsigned mtb_numreaders;
609 /** The actual reader table definition. */
610 typedef struct MDB_txninfo {
613 #define mti_magic mt1.mtb.mtb_magic
614 #define mti_format mt1.mtb.mtb_format
615 #define mti_mutex mt1.mtb.mtb_mutex
616 #define mti_rmname mt1.mtb.mtb_rmname
617 #define mti_txnid mt1.mtb.mtb_txnid
618 #define mti_numreaders mt1.mtb.mtb_numreaders
619 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
622 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
623 char mt2_wmname[MNAME_LEN];
624 #define mti_wmname mt2.mt2_wmname
626 pthread_mutex_t mt2_wmutex;
627 #define mti_wmutex mt2.mt2_wmutex
629 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
631 MDB_reader mti_readers[1];
634 /** Lockfile format signature: version, features and field layout */
635 #define MDB_LOCK_FORMAT \
637 ((MDB_LOCK_VERSION) \
638 /* Flags which describe functionality */ \
639 + (((MDB_PIDLOCK) != 0) << 16)))
642 /** Common header for all page types.
643 * Overflow records occupy a number of contiguous pages with no
644 * headers on any page after the first.
646 typedef struct MDB_page {
647 #define mp_pgno mp_p.p_pgno
648 #define mp_next mp_p.p_next
650 pgno_t p_pgno; /**< page number */
651 void * p_next; /**< for in-memory list of freed structs */
654 /** @defgroup mdb_page Page Flags
656 * Flags for the page headers.
659 #define P_BRANCH 0x01 /**< branch page */
660 #define P_LEAF 0x02 /**< leaf page */
661 #define P_OVERFLOW 0x04 /**< overflow page */
662 #define P_META 0x08 /**< meta page */
663 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
664 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
665 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
666 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
667 #define P_KEEP 0x8000 /**< leave this page alone during spill */
669 uint16_t mp_flags; /**< @ref mdb_page */
670 #define mp_lower mp_pb.pb.pb_lower
671 #define mp_upper mp_pb.pb.pb_upper
672 #define mp_pages mp_pb.pb_pages
675 indx_t pb_lower; /**< lower bound of free space */
676 indx_t pb_upper; /**< upper bound of free space */
678 uint32_t pb_pages; /**< number of overflow pages */
680 indx_t mp_ptrs[1]; /**< dynamic size */
683 /** Size of the page header, excluding dynamic data at the end */
684 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
686 /** Address of first usable data byte in a page, after the header */
687 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
689 /** Number of nodes on a page */
690 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
692 /** The amount of space remaining in the page */
693 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
695 /** The percentage of space used in the page, in tenths of a percent. */
696 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
697 ((env)->me_psize - PAGEHDRSZ))
698 /** The minimum page fill factor, in tenths of a percent.
699 * Pages emptier than this are candidates for merging.
701 #define FILL_THRESHOLD 250
703 /** Test if a page is a leaf page */
704 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
705 /** Test if a page is a LEAF2 page */
706 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
707 /** Test if a page is a branch page */
708 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
709 /** Test if a page is an overflow page */
710 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
711 /** Test if a page is a sub page */
712 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
714 /** The number of overflow pages needed to store the given size. */
715 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
717 /** Link in #MDB_txn.%mt_loose_pages list */
718 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)METADATA(p))
720 /** Header for a single key/data pair within a page.
721 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
722 * We guarantee 2-byte alignment for 'MDB_node's.
724 typedef struct MDB_node {
725 /** lo and hi are used for data size on leaf nodes and for
726 * child pgno on branch nodes. On 64 bit platforms, flags
727 * is also used for pgno. (Branch nodes have no flags).
728 * They are in host byte order in case that lets some
729 * accesses be optimized into a 32-bit word access.
731 #if BYTE_ORDER == LITTLE_ENDIAN
732 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
734 unsigned short mn_hi, mn_lo;
736 /** @defgroup mdb_node Node Flags
738 * Flags for node headers.
741 #define F_BIGDATA 0x01 /**< data put on overflow page */
742 #define F_SUBDATA 0x02 /**< data is a sub-database */
743 #define F_DUPDATA 0x04 /**< data has duplicates */
745 /** valid flags for #mdb_node_add() */
746 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
749 unsigned short mn_flags; /**< @ref mdb_node */
750 unsigned short mn_ksize; /**< key size */
751 char mn_data[1]; /**< key and data are appended here */
754 /** Size of the node header, excluding dynamic data at the end */
755 #define NODESIZE offsetof(MDB_node, mn_data)
757 /** Bit position of top word in page number, for shifting mn_flags */
758 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
760 /** Size of a node in a branch page with a given key.
761 * This is just the node header plus the key, there is no data.
763 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
765 /** Size of a node in a leaf page with a given key and data.
766 * This is node header plus key plus data size.
768 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
770 /** Address of node \b i in page \b p */
771 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
773 /** Address of the key for the node */
774 #define NODEKEY(node) (void *)((node)->mn_data)
776 /** Address of the data for a node */
777 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
779 /** Get the page number pointed to by a branch node */
780 #define NODEPGNO(node) \
781 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
782 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
783 /** Set the page number in a branch node */
784 #define SETPGNO(node,pgno) do { \
785 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
786 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
788 /** Get the size of the data in a leaf node */
789 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
790 /** Set the size of the data for a leaf node */
791 #define SETDSZ(node,size) do { \
792 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
793 /** The size of a key in a node */
794 #define NODEKSZ(node) ((node)->mn_ksize)
796 /** Copy a page number from src to dst */
798 #define COPY_PGNO(dst,src) dst = src
800 #if SIZE_MAX > 4294967295UL
801 #define COPY_PGNO(dst,src) do { \
802 unsigned short *s, *d; \
803 s = (unsigned short *)&(src); \
804 d = (unsigned short *)&(dst); \
811 #define COPY_PGNO(dst,src) do { \
812 unsigned short *s, *d; \
813 s = (unsigned short *)&(src); \
814 d = (unsigned short *)&(dst); \
820 /** The address of a key in a LEAF2 page.
821 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
822 * There are no node headers, keys are stored contiguously.
824 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
826 /** Set the \b node's key into \b keyptr, if requested. */
827 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
828 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
830 /** Set the \b node's key into \b key. */
831 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
833 /** Information about a single database in the environment. */
834 typedef struct MDB_db {
835 uint32_t md_pad; /**< also ksize for LEAF2 pages */
836 uint16_t md_flags; /**< @ref mdb_dbi_open */
837 uint16_t md_depth; /**< depth of this tree */
838 pgno_t md_branch_pages; /**< number of internal pages */
839 pgno_t md_leaf_pages; /**< number of leaf pages */
840 pgno_t md_overflow_pages; /**< number of overflow pages */
841 size_t md_entries; /**< number of data items */
842 pgno_t md_root; /**< the root page of this tree */
845 /** mdb_dbi_open flags */
846 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
847 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
848 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
849 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
851 /** Handle for the DB used to track free pages. */
853 /** Handle for the default DB. */
856 /** Meta page content.
857 * A meta page is the start point for accessing a database snapshot.
858 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
860 typedef struct MDB_meta {
861 /** Stamp identifying this as an LMDB file. It must be set
864 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
866 void *mm_address; /**< address for fixed mapping */
867 size_t mm_mapsize; /**< size of mmap region */
868 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
869 /** The size of pages used in this DB */
870 #define mm_psize mm_dbs[0].md_pad
871 /** Any persistent environment flags. @ref mdb_env */
872 #define mm_flags mm_dbs[0].md_flags
873 pgno_t mm_last_pg; /**< last used page in file */
874 txnid_t mm_txnid; /**< txnid that committed this page */
877 /** Buffer for a stack-allocated meta page.
878 * The members define size and alignment, and silence type
879 * aliasing warnings. They are not used directly; that could
880 * mean incorrectly using several union members in parallel.
882 typedef union MDB_metabuf {
885 char mm_pad[PAGEHDRSZ];
890 /** Auxiliary DB info.
891 * The information here is mostly static/read-only. There is
892 * only a single copy of this record in the environment.
894 typedef struct MDB_dbx {
895 MDB_val md_name; /**< name of the database */
896 MDB_cmp_func *md_cmp; /**< function for comparing keys */
897 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
898 MDB_rel_func *md_rel; /**< user relocate function */
899 void *md_relctx; /**< user-provided context for md_rel */
902 /** A database transaction.
903 * Every operation requires a transaction handle.
906 MDB_txn *mt_parent; /**< parent of a nested txn */
907 MDB_txn *mt_child; /**< nested txn under this txn */
908 pgno_t mt_next_pgno; /**< next unallocated page */
909 /** The ID of this transaction. IDs are integers incrementing from 1.
910 * Only committed write transactions increment the ID. If a transaction
911 * aborts, the ID may be re-used by the next writer.
914 MDB_env *mt_env; /**< the DB environment */
915 /** The list of pages that became unused during this transaction.
918 /** The list of loose pages that became unused and may be reused
919 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
921 MDB_page *mt_loose_pgs;
922 /** The sorted list of dirty pages we temporarily wrote to disk
923 * because the dirty list was full. page numbers in here are
924 * shifted left by 1, deleted slots have the LSB set.
926 MDB_IDL mt_spill_pgs;
928 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
930 /** For read txns: This thread/txn's reader table slot, or NULL. */
933 /** Array of records for each DB known in the environment. */
935 /** Array of MDB_db records for each known DB */
937 /** @defgroup mt_dbflag Transaction DB Flags
941 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
942 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
943 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
944 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
946 /** In write txns, array of cursors for each DB */
947 MDB_cursor **mt_cursors;
948 /** Array of flags for each DB */
949 unsigned char *mt_dbflags;
950 /** Number of DB records in use. This number only ever increments;
951 * we don't decrement it when individual DB handles are closed.
955 /** @defgroup mdb_txn Transaction Flags
959 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
960 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
961 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
962 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
964 unsigned int mt_flags; /**< @ref mdb_txn */
965 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
966 * Includes ancestor txns' dirty pages not hidden by other txns'
967 * dirty/spilled pages. Thus commit(nested txn) has room to merge
968 * dirty_list into mt_parent after freeing hidden mt_parent pages.
970 unsigned int mt_dirty_room;
973 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
974 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
975 * raise this on a 64 bit machine.
977 #define CURSOR_STACK 32
981 /** Cursors are used for all DB operations.
982 * A cursor holds a path of (page pointer, key index) from the DB
983 * root to a position in the DB, plus other state. #MDB_DUPSORT
984 * cursors include an xcursor to the current data item. Write txns
985 * track their cursors and keep them up to date when data moves.
986 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
987 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
990 /** Next cursor on this DB in this txn */
992 /** Backup of the original cursor if this cursor is a shadow */
993 MDB_cursor *mc_backup;
994 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
995 struct MDB_xcursor *mc_xcursor;
996 /** The transaction that owns this cursor */
998 /** The database handle this cursor operates on */
1000 /** The database record for this cursor */
1002 /** The database auxiliary record for this cursor */
1004 /** The @ref mt_dbflag for this database */
1005 unsigned char *mc_dbflag;
1006 unsigned short mc_snum; /**< number of pushed pages */
1007 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1008 /** @defgroup mdb_cursor Cursor Flags
1010 * Cursor state flags.
1013 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1014 #define C_EOF 0x02 /**< No more data */
1015 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1016 #define C_DEL 0x08 /**< last op was a cursor_del */
1017 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1018 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1020 unsigned int mc_flags; /**< @ref mdb_cursor */
1021 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1022 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1025 /** Context for sorted-dup records.
1026 * We could have gone to a fully recursive design, with arbitrarily
1027 * deep nesting of sub-databases. But for now we only handle these
1028 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1030 typedef struct MDB_xcursor {
1031 /** A sub-cursor for traversing the Dup DB */
1032 MDB_cursor mx_cursor;
1033 /** The database record for this Dup DB */
1035 /** The auxiliary DB record for this Dup DB */
1037 /** The @ref mt_dbflag for this Dup DB */
1038 unsigned char mx_dbflag;
1041 /** State of FreeDB old pages, stored in the MDB_env */
1042 typedef struct MDB_pgstate {
1043 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1044 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1047 /** The database environment. */
1049 HANDLE me_fd; /**< The main data file */
1050 HANDLE me_lfd; /**< The lock file */
1051 HANDLE me_mfd; /**< just for writing the meta pages */
1052 /** Failed to update the meta page. Probably an I/O error. */
1053 #define MDB_FATAL_ERROR 0x80000000U
1054 /** Some fields are initialized. */
1055 #define MDB_ENV_ACTIVE 0x20000000U
1056 /** me_txkey is set */
1057 #define MDB_ENV_TXKEY 0x10000000U
1058 uint32_t me_flags; /**< @ref mdb_env */
1059 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1060 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1061 unsigned int me_maxreaders; /**< size of the reader table */
1062 unsigned int me_numreaders; /**< max numreaders set by this env */
1063 MDB_dbi me_numdbs; /**< number of DBs opened */
1064 MDB_dbi me_maxdbs; /**< size of the DB table */
1065 MDB_PID_T me_pid; /**< process ID of this env */
1066 char *me_path; /**< path to the DB files */
1067 char *me_map; /**< the memory map of the data file */
1068 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1069 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1070 void *me_pbuf; /**< scratch area for DUPSORT put() */
1071 MDB_txn *me_txn; /**< current write transaction */
1072 size_t me_mapsize; /**< size of the data memory map */
1073 off_t me_size; /**< current file size */
1074 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1075 MDB_dbx *me_dbxs; /**< array of static DB info */
1076 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1077 pthread_key_t me_txkey; /**< thread-key for readers */
1078 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1079 # define me_pglast me_pgstate.mf_pglast
1080 # define me_pghead me_pgstate.mf_pghead
1081 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1082 /** IDL of pages that became unused in a write txn */
1083 MDB_IDL me_free_pgs;
1084 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1085 MDB_ID2L me_dirty_list;
1086 /** Max number of freelist items that can fit in a single overflow page */
1088 /** Max size of a node on a page */
1089 unsigned int me_nodemax;
1090 #if !(MDB_MAXKEYSIZE)
1091 unsigned int me_maxkey; /**< max size of a key */
1093 int me_live_reader; /**< have liveness lock in reader table */
1095 int me_pidquery; /**< Used in OpenProcess */
1096 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1098 #elif defined(MDB_USE_POSIX_SEM)
1099 sem_t *me_rmutex; /* Shared mutexes are not supported */
1102 void *me_userctx; /**< User-settable context */
1103 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1106 /** Nested transaction */
1107 typedef struct MDB_ntxn {
1108 MDB_txn mnt_txn; /**< the transaction */
1109 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1112 /** max number of pages to commit in one writev() call */
1113 #define MDB_COMMIT_PAGES 64
1114 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1115 #undef MDB_COMMIT_PAGES
1116 #define MDB_COMMIT_PAGES IOV_MAX
1119 /** max bytes to write in one call */
1120 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1122 /** Check \b txn and \b dbi arguments to a function */
1123 #define TXN_DBI_EXIST(txn, dbi) \
1124 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1126 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1127 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1128 static int mdb_page_touch(MDB_cursor *mc);
1130 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1131 static int mdb_page_search_root(MDB_cursor *mc,
1132 MDB_val *key, int modify);
1133 #define MDB_PS_MODIFY 1
1134 #define MDB_PS_ROOTONLY 2
1135 #define MDB_PS_FIRST 4
1136 #define MDB_PS_LAST 8
1137 static int mdb_page_search(MDB_cursor *mc,
1138 MDB_val *key, int flags);
1139 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1141 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1142 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1143 pgno_t newpgno, unsigned int nflags);
1145 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1146 static int mdb_env_pick_meta(const MDB_env *env);
1147 static int mdb_env_write_meta(MDB_txn *txn);
1148 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1149 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1151 static void mdb_env_close0(MDB_env *env, int excl);
1153 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1154 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1155 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1156 static void mdb_node_del(MDB_cursor *mc, int ksize);
1157 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1158 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1159 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1160 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1161 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1163 static int mdb_rebalance(MDB_cursor *mc);
1164 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1166 static void mdb_cursor_pop(MDB_cursor *mc);
1167 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1169 static int mdb_cursor_del0(MDB_cursor *mc);
1170 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1171 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1172 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1173 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1174 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1176 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1177 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1179 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1180 static void mdb_xcursor_init0(MDB_cursor *mc);
1181 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1183 static int mdb_drop0(MDB_cursor *mc, int subs);
1184 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1187 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1191 static SECURITY_DESCRIPTOR mdb_null_sd;
1192 static SECURITY_ATTRIBUTES mdb_all_sa;
1193 static int mdb_sec_inited;
1196 /** Return the library version info. */
1198 mdb_version(int *major, int *minor, int *patch)
1200 if (major) *major = MDB_VERSION_MAJOR;
1201 if (minor) *minor = MDB_VERSION_MINOR;
1202 if (patch) *patch = MDB_VERSION_PATCH;
1203 return MDB_VERSION_STRING;
1206 /** Table of descriptions for LMDB @ref errors */
1207 static char *const mdb_errstr[] = {
1208 "MDB_KEYEXIST: Key/data pair already exists",
1209 "MDB_NOTFOUND: No matching key/data pair found",
1210 "MDB_PAGE_NOTFOUND: Requested page not found",
1211 "MDB_CORRUPTED: Located page was wrong type",
1212 "MDB_PANIC: Update of meta page failed",
1213 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1214 "MDB_INVALID: File is not an LMDB file",
1215 "MDB_MAP_FULL: Environment mapsize limit reached",
1216 "MDB_DBS_FULL: Environment maxdbs limit reached",
1217 "MDB_READERS_FULL: Environment maxreaders limit reached",
1218 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1219 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1220 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1221 "MDB_PAGE_FULL: Internal error - page has no more space",
1222 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1223 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1224 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1225 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1226 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1230 mdb_strerror(int err)
1234 return ("Successful return: 0");
1236 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1237 i = err - MDB_KEYEXIST;
1238 return mdb_errstr[i];
1241 return strerror(err);
1244 /** assert(3) variant in cursor context */
1245 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1246 /** assert(3) variant in transaction context */
1247 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1248 /** assert(3) variant in environment context */
1249 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1252 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1253 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1256 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1257 const char *func, const char *file, int line)
1260 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1261 file, line, expr_txt, func);
1262 if (env->me_assert_func)
1263 env->me_assert_func(env, buf);
1264 fprintf(stderr, "%s\n", buf);
1268 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1272 /** Return the page number of \b mp which may be sub-page, for debug output */
1274 mdb_dbg_pgno(MDB_page *mp)
1277 COPY_PGNO(ret, mp->mp_pgno);
1281 /** Display a key in hexadecimal and return the address of the result.
1282 * @param[in] key the key to display
1283 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1284 * @return The key in hexadecimal form.
1287 mdb_dkey(MDB_val *key, char *buf)
1290 unsigned char *c = key->mv_data;
1296 if (key->mv_size > DKBUF_MAXKEYSIZE)
1297 return "MDB_MAXKEYSIZE";
1298 /* may want to make this a dynamic check: if the key is mostly
1299 * printable characters, print it as-is instead of converting to hex.
1303 for (i=0; i<key->mv_size; i++)
1304 ptr += sprintf(ptr, "%02x", *c++);
1306 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1312 mdb_leafnode_type(MDB_node *n)
1314 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1315 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1316 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1319 /** Display all the keys in the page. */
1321 mdb_page_list(MDB_page *mp)
1323 pgno_t pgno = mdb_dbg_pgno(mp);
1324 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1326 unsigned int i, nkeys, nsize, total = 0;
1330 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1331 case P_BRANCH: type = "Branch page"; break;
1332 case P_LEAF: type = "Leaf page"; break;
1333 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1334 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1335 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1337 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1338 pgno, mp->mp_pages, state);
1341 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1342 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1345 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1349 nkeys = NUMKEYS(mp);
1350 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1352 for (i=0; i<nkeys; i++) {
1353 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1354 key.mv_size = nsize = mp->mp_pad;
1355 key.mv_data = LEAF2KEY(mp, i, nsize);
1357 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1360 node = NODEPTR(mp, i);
1361 key.mv_size = node->mn_ksize;
1362 key.mv_data = node->mn_data;
1363 nsize = NODESIZE + key.mv_size;
1364 if (IS_BRANCH(mp)) {
1365 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1369 if (F_ISSET(node->mn_flags, F_BIGDATA))
1370 nsize += sizeof(pgno_t);
1372 nsize += NODEDSZ(node);
1374 nsize += sizeof(indx_t);
1375 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1376 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1378 total = EVEN(total);
1380 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1381 IS_LEAF2(mp) ? PAGEHDRSZ : mp->mp_lower, total, SIZELEFT(mp));
1385 mdb_cursor_chk(MDB_cursor *mc)
1391 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1392 for (i=0; i<mc->mc_top; i++) {
1394 node = NODEPTR(mp, mc->mc_ki[i]);
1395 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1398 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1404 /** Count all the pages in each DB and in the freelist
1405 * and make sure it matches the actual number of pages
1407 * All named DBs must be open for a correct count.
1409 static void mdb_audit(MDB_txn *txn)
1413 MDB_ID freecount, count;
1418 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1419 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1420 freecount += *(MDB_ID *)data.mv_data;
1421 mdb_tassert(txn, rc == MDB_NOTFOUND);
1424 for (i = 0; i<txn->mt_numdbs; i++) {
1426 if (!(txn->mt_dbflags[i] & DB_VALID))
1428 mdb_cursor_init(&mc, txn, i, &mx);
1429 if (txn->mt_dbs[i].md_root == P_INVALID)
1431 count += txn->mt_dbs[i].md_branch_pages +
1432 txn->mt_dbs[i].md_leaf_pages +
1433 txn->mt_dbs[i].md_overflow_pages;
1434 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1435 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1436 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1439 mp = mc.mc_pg[mc.mc_top];
1440 for (j=0; j<NUMKEYS(mp); j++) {
1441 MDB_node *leaf = NODEPTR(mp, j);
1442 if (leaf->mn_flags & F_SUBDATA) {
1444 memcpy(&db, NODEDATA(leaf), sizeof(db));
1445 count += db.md_branch_pages + db.md_leaf_pages +
1446 db.md_overflow_pages;
1450 mdb_tassert(txn, rc == MDB_NOTFOUND);
1453 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1454 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1455 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1461 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1463 return txn->mt_dbxs[dbi].md_cmp(a, b);
1467 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1469 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1472 /** Allocate memory for a page.
1473 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1476 mdb_page_malloc(MDB_txn *txn, unsigned num)
1478 MDB_env *env = txn->mt_env;
1479 MDB_page *ret = env->me_dpages;
1480 size_t psize = env->me_psize, sz = psize, off;
1481 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1482 * For a single page alloc, we init everything after the page header.
1483 * For multi-page, we init the final page; if the caller needed that
1484 * many pages they will be filling in at least up to the last page.
1488 VGMEMP_ALLOC(env, ret, sz);
1489 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1490 env->me_dpages = ret->mp_next;
1493 psize -= off = PAGEHDRSZ;
1498 if ((ret = malloc(sz)) != NULL) {
1499 VGMEMP_ALLOC(env, ret, sz);
1500 if (!(env->me_flags & MDB_NOMEMINIT)) {
1501 memset((char *)ret + off, 0, psize);
1505 txn->mt_flags |= MDB_TXN_ERROR;
1509 /** Free a single page.
1510 * Saves single pages to a list, for future reuse.
1511 * (This is not used for multi-page overflow pages.)
1514 mdb_page_free(MDB_env *env, MDB_page *mp)
1516 mp->mp_next = env->me_dpages;
1517 VGMEMP_FREE(env, mp);
1518 env->me_dpages = mp;
1521 /** Free a dirty page */
1523 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1525 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1526 mdb_page_free(env, dp);
1528 /* large pages just get freed directly */
1529 VGMEMP_FREE(env, dp);
1534 /** Return all dirty pages to dpage list */
1536 mdb_dlist_free(MDB_txn *txn)
1538 MDB_env *env = txn->mt_env;
1539 MDB_ID2L dl = txn->mt_u.dirty_list;
1540 unsigned i, n = dl[0].mid;
1542 for (i = 1; i <= n; i++) {
1543 mdb_dpage_free(env, dl[i].mptr);
1548 /** Loosen or free a single page.
1549 * Saves single pages to a list for future reuse
1550 * in this same txn. It has been pulled from the freeDB
1551 * and already resides on the dirty list, but has been
1552 * deleted. Use these pages first before pulling again
1555 * If the page wasn't dirtied in this txn, just add it
1556 * to this txn's free list.
1559 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1562 pgno_t pgno = mp->mp_pgno;
1564 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1565 if (mc->mc_txn->mt_parent) {
1566 MDB_ID2 *dl = mc->mc_txn->mt_u.dirty_list;
1567 /* If txn has a parent, make sure the page is in our
1571 unsigned x = mdb_mid2l_search(dl, pgno);
1572 if (x <= dl[0].mid && dl[x].mid == pgno) {
1573 if (mp != dl[x].mptr) { /* bad cursor? */
1574 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1575 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
1576 return MDB_CORRUPTED;
1583 /* no parent txn, so it's just ours */
1588 NEXT_LOOSE_PAGE(mp) = mc->mc_txn->mt_loose_pgs;
1589 mc->mc_txn->mt_loose_pgs = mp;
1590 mp->mp_flags |= P_LOOSE;
1592 int rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, pgno);
1600 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1601 * @param[in] mc A cursor handle for the current operation.
1602 * @param[in] pflags Flags of the pages to update:
1603 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1604 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1605 * @return 0 on success, non-zero on failure.
1608 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1610 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1611 MDB_txn *txn = mc->mc_txn;
1617 int rc = MDB_SUCCESS, level;
1619 /* Mark pages seen by cursors */
1620 if (mc->mc_flags & C_UNTRACK)
1621 mc = NULL; /* will find mc in mt_cursors */
1622 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1623 for (; mc; mc=mc->mc_next) {
1624 if (!(mc->mc_flags & C_INITIALIZED))
1626 for (m3 = mc;; m3 = &mx->mx_cursor) {
1628 for (j=0; j<m3->mc_snum; j++) {
1630 if ((mp->mp_flags & Mask) == pflags)
1631 mp->mp_flags ^= P_KEEP;
1633 mx = m3->mc_xcursor;
1634 /* Proceed to mx if it is at a sub-database */
1635 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1637 if (! (mp && (mp->mp_flags & P_LEAF)))
1639 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1640 if (!(leaf->mn_flags & F_SUBDATA))
1648 /* Loose pages shouldn't be spilled */
1649 for (dp = txn->mt_loose_pgs; dp; dp = NEXT_LOOSE_PAGE(dp)) {
1650 if ((dp->mp_flags & Mask) == pflags)
1651 dp->mp_flags ^= P_KEEP;
1655 /* Mark dirty root pages */
1656 for (i=0; i<txn->mt_numdbs; i++) {
1657 if (txn->mt_dbflags[i] & DB_DIRTY) {
1658 pgno_t pgno = txn->mt_dbs[i].md_root;
1659 if (pgno == P_INVALID)
1661 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1663 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1664 dp->mp_flags ^= P_KEEP;
1672 static int mdb_page_flush(MDB_txn *txn, int keep);
1674 /** Spill pages from the dirty list back to disk.
1675 * This is intended to prevent running into #MDB_TXN_FULL situations,
1676 * but note that they may still occur in a few cases:
1677 * 1) our estimate of the txn size could be too small. Currently this
1678 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1679 * 2) child txns may run out of space if their parents dirtied a
1680 * lot of pages and never spilled them. TODO: we probably should do
1681 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1682 * the parent's dirty_room is below a given threshold.
1684 * Otherwise, if not using nested txns, it is expected that apps will
1685 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1686 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1687 * If the txn never references them again, they can be left alone.
1688 * If the txn only reads them, they can be used without any fuss.
1689 * If the txn writes them again, they can be dirtied immediately without
1690 * going thru all of the work of #mdb_page_touch(). Such references are
1691 * handled by #mdb_page_unspill().
1693 * Also note, we never spill DB root pages, nor pages of active cursors,
1694 * because we'll need these back again soon anyway. And in nested txns,
1695 * we can't spill a page in a child txn if it was already spilled in a
1696 * parent txn. That would alter the parent txns' data even though
1697 * the child hasn't committed yet, and we'd have no way to undo it if
1698 * the child aborted.
1700 * @param[in] m0 cursor A cursor handle identifying the transaction and
1701 * database for which we are checking space.
1702 * @param[in] key For a put operation, the key being stored.
1703 * @param[in] data For a put operation, the data being stored.
1704 * @return 0 on success, non-zero on failure.
1707 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1709 MDB_txn *txn = m0->mc_txn;
1711 MDB_ID2L dl = txn->mt_u.dirty_list;
1712 unsigned int i, j, need;
1715 if (m0->mc_flags & C_SUB)
1718 /* Estimate how much space this op will take */
1719 i = m0->mc_db->md_depth;
1720 /* Named DBs also dirty the main DB */
1721 if (m0->mc_dbi > MAIN_DBI)
1722 i += txn->mt_dbs[MAIN_DBI].md_depth;
1723 /* For puts, roughly factor in the key+data size */
1725 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1726 i += i; /* double it for good measure */
1729 if (txn->mt_dirty_room > i)
1732 if (!txn->mt_spill_pgs) {
1733 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1734 if (!txn->mt_spill_pgs)
1737 /* purge deleted slots */
1738 MDB_IDL sl = txn->mt_spill_pgs;
1739 unsigned int num = sl[0];
1741 for (i=1; i<=num; i++) {
1748 /* Preserve pages which may soon be dirtied again */
1749 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1752 /* Less aggressive spill - we originally spilled the entire dirty list,
1753 * with a few exceptions for cursor pages and DB root pages. But this
1754 * turns out to be a lot of wasted effort because in a large txn many
1755 * of those pages will need to be used again. So now we spill only 1/8th
1756 * of the dirty pages. Testing revealed this to be a good tradeoff,
1757 * better than 1/2, 1/4, or 1/10.
1759 if (need < MDB_IDL_UM_MAX / 8)
1760 need = MDB_IDL_UM_MAX / 8;
1762 /* Save the page IDs of all the pages we're flushing */
1763 /* flush from the tail forward, this saves a lot of shifting later on. */
1764 for (i=dl[0].mid; i && need; i--) {
1765 MDB_ID pn = dl[i].mid << 1;
1767 if (dp->mp_flags & P_KEEP)
1769 /* Can't spill twice, make sure it's not already in a parent's
1772 if (txn->mt_parent) {
1774 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1775 if (tx2->mt_spill_pgs) {
1776 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1777 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1778 dp->mp_flags |= P_KEEP;
1786 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1790 mdb_midl_sort(txn->mt_spill_pgs);
1792 /* Flush the spilled part of dirty list */
1793 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1796 /* Reset any dirty pages we kept that page_flush didn't see */
1797 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1800 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1804 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1806 mdb_find_oldest(MDB_txn *txn)
1809 txnid_t mr, oldest = txn->mt_txnid - 1;
1810 if (txn->mt_env->me_txns) {
1811 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1812 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1823 /** Add a page to the txn's dirty list */
1825 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1828 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1830 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1831 insert = mdb_mid2l_append;
1833 insert = mdb_mid2l_insert;
1835 mid.mid = mp->mp_pgno;
1837 rc = insert(txn->mt_u.dirty_list, &mid);
1838 mdb_tassert(txn, rc == 0);
1839 txn->mt_dirty_room--;
1842 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1843 * me_pghead and mt_next_pgno.
1845 * If there are free pages available from older transactions, they
1846 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1847 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1848 * and move me_pglast to say which records were consumed. Only this
1849 * function can create me_pghead and move me_pglast/mt_next_pgno.
1850 * @param[in] mc cursor A cursor handle identifying the transaction and
1851 * database for which we are allocating.
1852 * @param[in] num the number of pages to allocate.
1853 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1854 * will always be satisfied by a single contiguous chunk of memory.
1855 * @return 0 on success, non-zero on failure.
1858 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1860 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1861 /* Get at most <Max_retries> more freeDB records once me_pghead
1862 * has enough pages. If not enough, use new pages from the map.
1863 * If <Paranoid> and mc is updating the freeDB, only get new
1864 * records if me_pghead is empty. Then the freelist cannot play
1865 * catch-up with itself by growing while trying to save it.
1867 enum { Paranoid = 1, Max_retries = 500 };
1869 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1871 int rc, retry = num * 20;
1872 MDB_txn *txn = mc->mc_txn;
1873 MDB_env *env = txn->mt_env;
1874 pgno_t pgno, *mop = env->me_pghead;
1875 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1877 txnid_t oldest = 0, last;
1881 /* If there are any loose pages, just use them */
1882 if (num == 1 && txn->mt_loose_pgs) {
1883 np = txn->mt_loose_pgs;
1884 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1891 /* If our dirty list is already full, we can't do anything */
1892 if (txn->mt_dirty_room == 0) {
1897 for (op = MDB_FIRST;; op = MDB_NEXT) {
1900 pgno_t *idl, old_id, new_id;
1902 /* Seek a big enough contiguous page range. Prefer
1903 * pages at the tail, just truncating the list.
1909 if (mop[i-n2] == pgno+n2)
1916 if (op == MDB_FIRST) { /* 1st iteration */
1917 /* Prepare to fetch more and coalesce */
1918 oldest = mdb_find_oldest(txn);
1919 last = env->me_pglast;
1920 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1923 key.mv_data = &last; /* will look up last+1 */
1924 key.mv_size = sizeof(last);
1926 if (Paranoid && mc->mc_dbi == FREE_DBI)
1929 if (Paranoid && retry < 0 && mop_len)
1933 /* Do not fetch more if the record will be too recent */
1936 rc = mdb_cursor_get(&m2, &key, NULL, op);
1938 if (rc == MDB_NOTFOUND)
1942 last = *(txnid_t*)key.mv_data;
1945 np = m2.mc_pg[m2.mc_top];
1946 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1947 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1950 idl = (MDB_ID *) data.mv_data;
1953 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
1958 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1960 mop = env->me_pghead;
1962 env->me_pglast = last;
1964 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1965 last, txn->mt_dbs[FREE_DBI].md_root, i));
1967 DPRINTF(("IDL %"Z"u", idl[k]));
1969 /* Merge in descending sorted order */
1972 mop[0] = (pgno_t)-1;
1976 for (; old_id < new_id; old_id = mop[--j])
1983 /* Use new pages from the map when nothing suitable in the freeDB */
1985 pgno = txn->mt_next_pgno;
1986 if (pgno + num >= env->me_maxpg) {
1987 DPUTS("DB size maxed out");
1993 if (env->me_flags & MDB_WRITEMAP) {
1994 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1996 if (!(np = mdb_page_malloc(txn, num))) {
2002 mop[0] = mop_len -= num;
2003 /* Move any stragglers down */
2004 for (j = i-num; j < mop_len; )
2005 mop[++j] = mop[++i];
2007 txn->mt_next_pgno = pgno + num;
2010 mdb_page_dirty(txn, np);
2016 txn->mt_flags |= MDB_TXN_ERROR;
2020 /** Copy the used portions of a non-overflow page.
2021 * @param[in] dst page to copy into
2022 * @param[in] src page to copy from
2023 * @param[in] psize size of a page
2026 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2028 enum { Align = sizeof(pgno_t) };
2029 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2031 /* If page isn't full, just copy the used portion. Adjust
2032 * alignment so memcpy may copy words instead of bytes.
2034 if ((unused &= -Align) && !IS_LEAF2(src)) {
2036 memcpy(dst, src, (lower + (Align-1)) & -Align);
2037 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2040 memcpy(dst, src, psize - unused);
2044 /** Pull a page off the txn's spill list, if present.
2045 * If a page being referenced was spilled to disk in this txn, bring
2046 * it back and make it dirty/writable again.
2047 * @param[in] txn the transaction handle.
2048 * @param[in] mp the page being referenced. It must not be dirty.
2049 * @param[out] ret the writable page, if any. ret is unchanged if
2050 * mp wasn't spilled.
2053 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2055 MDB_env *env = txn->mt_env;
2058 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2060 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2061 if (!tx2->mt_spill_pgs)
2063 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2064 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2067 if (txn->mt_dirty_room == 0)
2068 return MDB_TXN_FULL;
2069 if (IS_OVERFLOW(mp))
2073 if (env->me_flags & MDB_WRITEMAP) {
2076 np = mdb_page_malloc(txn, num);
2080 memcpy(np, mp, num * env->me_psize);
2082 mdb_page_copy(np, mp, env->me_psize);
2085 /* If in current txn, this page is no longer spilled.
2086 * If it happens to be the last page, truncate the spill list.
2087 * Otherwise mark it as deleted by setting the LSB.
2089 if (x == txn->mt_spill_pgs[0])
2090 txn->mt_spill_pgs[0]--;
2092 txn->mt_spill_pgs[x] |= 1;
2093 } /* otherwise, if belonging to a parent txn, the
2094 * page remains spilled until child commits
2097 mdb_page_dirty(txn, np);
2098 np->mp_flags |= P_DIRTY;
2106 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2107 * @param[in] mc cursor pointing to the page to be touched
2108 * @return 0 on success, non-zero on failure.
2111 mdb_page_touch(MDB_cursor *mc)
2113 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2114 MDB_txn *txn = mc->mc_txn;
2115 MDB_cursor *m2, *m3;
2119 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2120 if (txn->mt_flags & MDB_TXN_SPILLS) {
2122 rc = mdb_page_unspill(txn, mp, &np);
2128 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2129 (rc = mdb_page_alloc(mc, 1, &np)))
2132 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2133 mp->mp_pgno, pgno));
2134 mdb_cassert(mc, mp->mp_pgno != pgno);
2135 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2136 /* Update the parent page, if any, to point to the new page */
2138 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2139 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2140 SETPGNO(node, pgno);
2142 mc->mc_db->md_root = pgno;
2144 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2145 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2147 /* If txn has a parent, make sure the page is in our
2151 unsigned x = mdb_mid2l_search(dl, pgno);
2152 if (x <= dl[0].mid && dl[x].mid == pgno) {
2153 if (mp != dl[x].mptr) { /* bad cursor? */
2154 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2155 txn->mt_flags |= MDB_TXN_ERROR;
2156 return MDB_CORRUPTED;
2161 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2163 np = mdb_page_malloc(txn, 1);
2168 rc = mdb_mid2l_insert(dl, &mid);
2169 mdb_cassert(mc, rc == 0);
2174 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2176 np->mp_flags |= P_DIRTY;
2179 /* Adjust cursors pointing to mp */
2180 mc->mc_pg[mc->mc_top] = np;
2181 m2 = txn->mt_cursors[mc->mc_dbi];
2182 if (mc->mc_flags & C_SUB) {
2183 for (; m2; m2=m2->mc_next) {
2184 m3 = &m2->mc_xcursor->mx_cursor;
2185 if (m3->mc_snum < mc->mc_snum) continue;
2186 if (m3->mc_pg[mc->mc_top] == mp)
2187 m3->mc_pg[mc->mc_top] = np;
2190 for (; m2; m2=m2->mc_next) {
2191 if (m2->mc_snum < mc->mc_snum) continue;
2192 if (m2->mc_pg[mc->mc_top] == mp) {
2193 m2->mc_pg[mc->mc_top] = np;
2194 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2196 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2198 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2199 if (!(leaf->mn_flags & F_SUBDATA))
2200 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2208 txn->mt_flags |= MDB_TXN_ERROR;
2213 mdb_env_sync(MDB_env *env, int force)
2216 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2217 if (env->me_flags & MDB_WRITEMAP) {
2218 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2219 ? MS_ASYNC : MS_SYNC;
2220 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2223 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2227 if (MDB_FDATASYNC(env->me_fd))
2234 /** Back up parent txn's cursors, then grab the originals for tracking */
2236 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2238 MDB_cursor *mc, *bk;
2243 for (i = src->mt_numdbs; --i >= 0; ) {
2244 if ((mc = src->mt_cursors[i]) != NULL) {
2245 size = sizeof(MDB_cursor);
2247 size += sizeof(MDB_xcursor);
2248 for (; mc; mc = bk->mc_next) {
2254 mc->mc_db = &dst->mt_dbs[i];
2255 /* Kill pointers into src - and dst to reduce abuse: The
2256 * user may not use mc until dst ends. Otherwise we'd...
2258 mc->mc_txn = NULL; /* ...set this to dst */
2259 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2260 if ((mx = mc->mc_xcursor) != NULL) {
2261 *(MDB_xcursor *)(bk+1) = *mx;
2262 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2264 mc->mc_next = dst->mt_cursors[i];
2265 dst->mt_cursors[i] = mc;
2272 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2273 * @param[in] txn the transaction handle.
2274 * @param[in] merge true to keep changes to parent cursors, false to revert.
2275 * @return 0 on success, non-zero on failure.
2278 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2280 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2284 for (i = txn->mt_numdbs; --i >= 0; ) {
2285 for (mc = cursors[i]; mc; mc = next) {
2287 if ((bk = mc->mc_backup) != NULL) {
2289 /* Commit changes to parent txn */
2290 mc->mc_next = bk->mc_next;
2291 mc->mc_backup = bk->mc_backup;
2292 mc->mc_txn = bk->mc_txn;
2293 mc->mc_db = bk->mc_db;
2294 mc->mc_dbflag = bk->mc_dbflag;
2295 if ((mx = mc->mc_xcursor) != NULL)
2296 mx->mx_cursor.mc_txn = bk->mc_txn;
2298 /* Abort nested txn */
2300 if ((mx = mc->mc_xcursor) != NULL)
2301 *mx = *(MDB_xcursor *)(bk+1);
2305 /* Only malloced cursors are permanently tracked. */
2313 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2316 mdb_txn_reset0(MDB_txn *txn, const char *act);
2318 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2324 Pidset = F_SETLK, Pidcheck = F_GETLK
2328 /** Set or check a pid lock. Set returns 0 on success.
2329 * Check returns 0 if the process is certainly dead, nonzero if it may
2330 * be alive (the lock exists or an error happened so we do not know).
2332 * On Windows Pidset is a no-op, we merely check for the existence
2333 * of the process with the given pid. On POSIX we use a single byte
2334 * lock on the lockfile, set at an offset equal to the pid.
2337 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2339 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2342 if (op == Pidcheck) {
2343 h = OpenProcess(env->me_pidquery, FALSE, pid);
2344 /* No documented "no such process" code, but other program use this: */
2346 return ErrCode() != ERROR_INVALID_PARAMETER;
2347 /* A process exists until all handles to it close. Has it exited? */
2348 ret = WaitForSingleObject(h, 0) != 0;
2355 struct flock lock_info;
2356 memset(&lock_info, 0, sizeof(lock_info));
2357 lock_info.l_type = F_WRLCK;
2358 lock_info.l_whence = SEEK_SET;
2359 lock_info.l_start = pid;
2360 lock_info.l_len = 1;
2361 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2362 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2364 } else if ((rc = ErrCode()) == EINTR) {
2372 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2373 * @param[in] txn the transaction handle to initialize
2374 * @return 0 on success, non-zero on failure.
2377 mdb_txn_renew0(MDB_txn *txn)
2379 MDB_env *env = txn->mt_env;
2380 MDB_txninfo *ti = env->me_txns;
2384 int rc, new_notls = 0;
2387 txn->mt_numdbs = env->me_numdbs;
2388 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2390 if (txn->mt_flags & MDB_TXN_RDONLY) {
2392 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2393 txn->mt_txnid = meta->mm_txnid;
2394 txn->mt_u.reader = NULL;
2396 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2397 pthread_getspecific(env->me_txkey);
2399 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2400 return MDB_BAD_RSLOT;
2402 MDB_PID_T pid = env->me_pid;
2403 MDB_THR_T tid = pthread_self();
2405 if (!env->me_live_reader) {
2406 rc = mdb_reader_pid(env, Pidset, pid);
2409 env->me_live_reader = 1;
2413 nr = ti->mti_numreaders;
2414 for (i=0; i<nr; i++)
2415 if (ti->mti_readers[i].mr_pid == 0)
2417 if (i == env->me_maxreaders) {
2418 UNLOCK_MUTEX_R(env);
2419 return MDB_READERS_FULL;
2421 ti->mti_readers[i].mr_pid = pid;
2422 ti->mti_readers[i].mr_tid = tid;
2424 ti->mti_numreaders = ++nr;
2425 /* Save numreaders for un-mutexed mdb_env_close() */
2426 env->me_numreaders = nr;
2427 UNLOCK_MUTEX_R(env);
2429 r = &ti->mti_readers[i];
2430 new_notls = (env->me_flags & MDB_NOTLS);
2431 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2436 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2437 txn->mt_u.reader = r;
2438 meta = env->me_metas[txn->mt_txnid & 1];
2444 txn->mt_txnid = ti->mti_txnid;
2445 meta = env->me_metas[txn->mt_txnid & 1];
2447 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2448 txn->mt_txnid = meta->mm_txnid;
2452 if (txn->mt_txnid == mdb_debug_start)
2455 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2456 txn->mt_u.dirty_list = env->me_dirty_list;
2457 txn->mt_u.dirty_list[0].mid = 0;
2458 txn->mt_free_pgs = env->me_free_pgs;
2459 txn->mt_free_pgs[0] = 0;
2460 txn->mt_spill_pgs = NULL;
2464 /* Copy the DB info and flags */
2465 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2467 /* Moved to here to avoid a data race in read TXNs */
2468 txn->mt_next_pgno = meta->mm_last_pg+1;
2470 for (i=2; i<txn->mt_numdbs; i++) {
2471 x = env->me_dbflags[i];
2472 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2473 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2475 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2477 if (env->me_maxpg < txn->mt_next_pgno) {
2478 mdb_txn_reset0(txn, "renew0-mapfail");
2480 txn->mt_u.reader->mr_pid = 0;
2481 txn->mt_u.reader = NULL;
2483 return MDB_MAP_RESIZED;
2490 mdb_txn_renew(MDB_txn *txn)
2494 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2497 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2498 DPUTS("environment had fatal error, must shutdown!");
2502 rc = mdb_txn_renew0(txn);
2503 if (rc == MDB_SUCCESS) {
2504 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2505 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2506 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2512 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2516 int rc, size, tsize = sizeof(MDB_txn);
2518 if (env->me_flags & MDB_FATAL_ERROR) {
2519 DPUTS("environment had fatal error, must shutdown!");
2522 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2525 /* Nested transactions: Max 1 child, write txns only, no writemap */
2526 if (parent->mt_child ||
2527 (flags & MDB_RDONLY) ||
2528 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2529 (env->me_flags & MDB_WRITEMAP))
2531 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2533 tsize = sizeof(MDB_ntxn);
2535 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2536 if (!(flags & MDB_RDONLY))
2537 size += env->me_maxdbs * sizeof(MDB_cursor *);
2539 if ((txn = calloc(1, size)) == NULL) {
2540 DPRINTF(("calloc: %s", strerror(ErrCode())));
2543 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2544 if (flags & MDB_RDONLY) {
2545 txn->mt_flags |= MDB_TXN_RDONLY;
2546 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2548 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2549 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2555 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2556 if (!txn->mt_u.dirty_list ||
2557 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2559 free(txn->mt_u.dirty_list);
2563 txn->mt_txnid = parent->mt_txnid;
2564 txn->mt_dirty_room = parent->mt_dirty_room;
2565 txn->mt_u.dirty_list[0].mid = 0;
2566 txn->mt_spill_pgs = NULL;
2567 txn->mt_next_pgno = parent->mt_next_pgno;
2568 parent->mt_child = txn;
2569 txn->mt_parent = parent;
2570 txn->mt_numdbs = parent->mt_numdbs;
2571 txn->mt_flags = parent->mt_flags;
2572 txn->mt_dbxs = parent->mt_dbxs;
2573 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2574 /* Copy parent's mt_dbflags, but clear DB_NEW */
2575 for (i=0; i<txn->mt_numdbs; i++)
2576 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2578 ntxn = (MDB_ntxn *)txn;
2579 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2580 if (env->me_pghead) {
2581 size = MDB_IDL_SIZEOF(env->me_pghead);
2582 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2584 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2589 rc = mdb_cursor_shadow(parent, txn);
2591 mdb_txn_reset0(txn, "beginchild-fail");
2593 rc = mdb_txn_renew0(txn);
2599 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2600 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2601 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2608 mdb_txn_env(MDB_txn *txn)
2610 if(!txn) return NULL;
2614 /** Export or close DBI handles opened in this txn. */
2616 mdb_dbis_update(MDB_txn *txn, int keep)
2619 MDB_dbi n = txn->mt_numdbs;
2620 MDB_env *env = txn->mt_env;
2621 unsigned char *tdbflags = txn->mt_dbflags;
2623 for (i = n; --i >= 2;) {
2624 if (tdbflags[i] & DB_NEW) {
2626 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2628 char *ptr = env->me_dbxs[i].md_name.mv_data;
2629 env->me_dbxs[i].md_name.mv_data = NULL;
2630 env->me_dbxs[i].md_name.mv_size = 0;
2631 env->me_dbflags[i] = 0;
2636 if (keep && env->me_numdbs < n)
2640 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2641 * May be called twice for readonly txns: First reset it, then abort.
2642 * @param[in] txn the transaction handle to reset
2643 * @param[in] act why the transaction is being reset
2646 mdb_txn_reset0(MDB_txn *txn, const char *act)
2648 MDB_env *env = txn->mt_env;
2650 /* Close any DBI handles opened in this txn */
2651 mdb_dbis_update(txn, 0);
2653 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2654 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2655 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2657 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2658 if (txn->mt_u.reader) {
2659 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2660 if (!(env->me_flags & MDB_NOTLS))
2661 txn->mt_u.reader = NULL; /* txn does not own reader */
2663 txn->mt_numdbs = 0; /* close nothing if called again */
2664 txn->mt_dbxs = NULL; /* mark txn as reset */
2666 mdb_cursors_close(txn, 0);
2668 if (!(env->me_flags & MDB_WRITEMAP)) {
2669 mdb_dlist_free(txn);
2671 mdb_midl_free(env->me_pghead);
2673 if (txn->mt_parent) {
2674 txn->mt_parent->mt_child = NULL;
2675 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2676 mdb_midl_free(txn->mt_free_pgs);
2677 mdb_midl_free(txn->mt_spill_pgs);
2678 free(txn->mt_u.dirty_list);
2682 if (mdb_midl_shrink(&txn->mt_free_pgs))
2683 env->me_free_pgs = txn->mt_free_pgs;
2684 env->me_pghead = NULL;
2688 /* The writer mutex was locked in mdb_txn_begin. */
2690 UNLOCK_MUTEX_W(env);
2695 mdb_txn_reset(MDB_txn *txn)
2700 /* This call is only valid for read-only txns */
2701 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2704 mdb_txn_reset0(txn, "reset");
2708 mdb_txn_abort(MDB_txn *txn)
2714 mdb_txn_abort(txn->mt_child);
2716 mdb_txn_reset0(txn, "abort");
2717 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2718 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2719 txn->mt_u.reader->mr_pid = 0;
2724 /** Save the freelist as of this transaction to the freeDB.
2725 * This changes the freelist. Keep trying until it stabilizes.
2728 mdb_freelist_save(MDB_txn *txn)
2730 /* env->me_pghead[] can grow and shrink during this call.
2731 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2732 * Page numbers cannot disappear from txn->mt_free_pgs[].
2735 MDB_env *env = txn->mt_env;
2736 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2737 txnid_t pglast = 0, head_id = 0;
2738 pgno_t freecnt = 0, *free_pgs, *mop;
2739 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2741 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2743 if (env->me_pghead) {
2744 /* Make sure first page of freeDB is touched and on freelist */
2745 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2746 if (rc && rc != MDB_NOTFOUND)
2750 /* Dispose of loose pages. Usually they will have all
2751 * been used up by the time we get here.
2753 if (txn->mt_loose_pgs) {
2754 MDB_page *mp = txn->mt_loose_pgs;
2755 /* Just return them to freeDB */
2756 if (env->me_pghead) {
2758 mop = env->me_pghead;
2759 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2760 pgno_t pg = mp->mp_pgno;
2762 for (i = mop[0]; i && mop[i] < pg; i--)
2768 /* Oh well, they were wasted. Put on freelist */
2769 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2770 mdb_midl_append(&txn->mt_free_pgs, mp->mp_pgno);
2773 txn->mt_loose_pgs = NULL;
2776 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2777 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2778 ? SSIZE_MAX : maxfree_1pg;
2781 /* Come back here after each Put() in case freelist changed */
2786 /* If using records from freeDB which we have not yet
2787 * deleted, delete them and any we reserved for me_pghead.
2789 while (pglast < env->me_pglast) {
2790 rc = mdb_cursor_first(&mc, &key, NULL);
2793 pglast = head_id = *(txnid_t *)key.mv_data;
2794 total_room = head_room = 0;
2795 mdb_tassert(txn, pglast <= env->me_pglast);
2796 rc = mdb_cursor_del(&mc, 0);
2801 /* Save the IDL of pages freed by this txn, to a single record */
2802 if (freecnt < txn->mt_free_pgs[0]) {
2804 /* Make sure last page of freeDB is touched and on freelist */
2805 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2806 if (rc && rc != MDB_NOTFOUND)
2809 free_pgs = txn->mt_free_pgs;
2810 /* Write to last page of freeDB */
2811 key.mv_size = sizeof(txn->mt_txnid);
2812 key.mv_data = &txn->mt_txnid;
2814 freecnt = free_pgs[0];
2815 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2816 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2819 /* Retry if mt_free_pgs[] grew during the Put() */
2820 free_pgs = txn->mt_free_pgs;
2821 } while (freecnt < free_pgs[0]);
2822 mdb_midl_sort(free_pgs);
2823 memcpy(data.mv_data, free_pgs, data.mv_size);
2826 unsigned int i = free_pgs[0];
2827 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2828 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2830 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2836 mop = env->me_pghead;
2837 mop_len = mop ? mop[0] : 0;
2839 /* Reserve records for me_pghead[]. Split it if multi-page,
2840 * to avoid searching freeDB for a page range. Use keys in
2841 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2843 if (total_room >= mop_len) {
2844 if (total_room == mop_len || --more < 0)
2846 } else if (head_room >= maxfree_1pg && head_id > 1) {
2847 /* Keep current record (overflow page), add a new one */
2851 /* (Re)write {key = head_id, IDL length = head_room} */
2852 total_room -= head_room;
2853 head_room = mop_len - total_room;
2854 if (head_room > maxfree_1pg && head_id > 1) {
2855 /* Overflow multi-page for part of me_pghead */
2856 head_room /= head_id; /* amortize page sizes */
2857 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2858 } else if (head_room < 0) {
2859 /* Rare case, not bothering to delete this record */
2862 key.mv_size = sizeof(head_id);
2863 key.mv_data = &head_id;
2864 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2865 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2868 /* IDL is initially empty, zero out at least the length */
2869 pgs = (pgno_t *)data.mv_data;
2870 j = head_room > clean_limit ? head_room : 0;
2874 total_room += head_room;
2877 /* Fill in the reserved me_pghead records */
2883 rc = mdb_cursor_first(&mc, &key, &data);
2884 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2885 txnid_t id = *(txnid_t *)key.mv_data;
2886 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2889 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2891 if (len > mop_len) {
2893 data.mv_size = (len + 1) * sizeof(MDB_ID);
2895 data.mv_data = mop -= len;
2898 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2900 if (rc || !(mop_len -= len))
2907 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2908 * @param[in] txn the transaction that's being committed
2909 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2910 * @return 0 on success, non-zero on failure.
2913 mdb_page_flush(MDB_txn *txn, int keep)
2915 MDB_env *env = txn->mt_env;
2916 MDB_ID2L dl = txn->mt_u.dirty_list;
2917 unsigned psize = env->me_psize, j;
2918 int i, pagecount = dl[0].mid, rc;
2919 size_t size = 0, pos = 0;
2921 MDB_page *dp = NULL;
2925 struct iovec iov[MDB_COMMIT_PAGES];
2926 ssize_t wpos = 0, wsize = 0, wres;
2927 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2933 if (env->me_flags & MDB_WRITEMAP) {
2934 /* Clear dirty flags */
2935 while (++i <= pagecount) {
2937 /* Don't flush this page yet */
2938 if (dp->mp_flags & P_KEEP) {
2939 dp->mp_flags ^= P_KEEP;
2943 dp->mp_flags &= ~P_DIRTY;
2948 /* Write the pages */
2950 if (++i <= pagecount) {
2952 /* Don't flush this page yet */
2953 if (dp->mp_flags & P_KEEP) {
2954 dp->mp_flags ^= P_KEEP;
2959 /* clear dirty flag */
2960 dp->mp_flags &= ~P_DIRTY;
2963 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2968 /* Windows actually supports scatter/gather I/O, but only on
2969 * unbuffered file handles. Since we're relying on the OS page
2970 * cache for all our data, that's self-defeating. So we just
2971 * write pages one at a time. We use the ov structure to set
2972 * the write offset, to at least save the overhead of a Seek
2975 DPRINTF(("committing page %"Z"u", pgno));
2976 memset(&ov, 0, sizeof(ov));
2977 ov.Offset = pos & 0xffffffff;
2978 ov.OffsetHigh = pos >> 16 >> 16;
2979 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2981 DPRINTF(("WriteFile: %d", rc));
2985 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2986 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2988 /* Write previous page(s) */
2989 #ifdef MDB_USE_PWRITEV
2990 wres = pwritev(env->me_fd, iov, n, wpos);
2993 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2995 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2997 DPRINTF(("lseek: %s", strerror(rc)));
3000 wres = writev(env->me_fd, iov, n);
3003 if (wres != wsize) {
3006 DPRINTF(("Write error: %s", strerror(rc)));
3008 rc = EIO; /* TODO: Use which error code? */
3009 DPUTS("short write, filesystem full?");
3020 DPRINTF(("committing page %"Z"u", pgno));
3021 next_pos = pos + size;
3022 iov[n].iov_len = size;
3023 iov[n].iov_base = (char *)dp;
3029 for (i = keep; ++i <= pagecount; ) {
3031 /* This is a page we skipped above */
3034 dl[j].mid = dp->mp_pgno;
3037 mdb_dpage_free(env, dp);
3042 txn->mt_dirty_room += i - j;
3048 mdb_txn_commit(MDB_txn *txn)
3054 if (txn == NULL || txn->mt_env == NULL)
3057 if (txn->mt_child) {
3058 rc = mdb_txn_commit(txn->mt_child);
3059 txn->mt_child = NULL;
3066 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3067 mdb_dbis_update(txn, 1);
3068 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3073 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3074 DPUTS("error flag is set, can't commit");
3076 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3081 if (txn->mt_parent) {
3082 MDB_txn *parent = txn->mt_parent;
3085 unsigned x, y, len, ps_len;
3087 /* Append our free list to parent's */
3088 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3091 mdb_midl_free(txn->mt_free_pgs);
3092 /* Failures after this must either undo the changes
3093 * to the parent or set MDB_TXN_ERROR in the parent.
3096 parent->mt_next_pgno = txn->mt_next_pgno;
3097 parent->mt_flags = txn->mt_flags;
3099 /* Merge our cursors into parent's and close them */
3100 mdb_cursors_close(txn, 1);
3102 /* Update parent's DB table. */
3103 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3104 parent->mt_numdbs = txn->mt_numdbs;
3105 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3106 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3107 for (i=2; i<txn->mt_numdbs; i++) {
3108 /* preserve parent's DB_NEW status */
3109 x = parent->mt_dbflags[i] & DB_NEW;
3110 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3113 dst = parent->mt_u.dirty_list;
3114 src = txn->mt_u.dirty_list;
3115 /* Remove anything in our dirty list from parent's spill list */
3116 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3118 pspill[0] = (pgno_t)-1;
3119 /* Mark our dirty pages as deleted in parent spill list */
3120 for (i=0, len=src[0].mid; ++i <= len; ) {
3121 MDB_ID pn = src[i].mid << 1;
3122 while (pn > pspill[x])
3124 if (pn == pspill[x]) {
3129 /* Squash deleted pagenums if we deleted any */
3130 for (x=y; ++x <= ps_len; )
3131 if (!(pspill[x] & 1))
3132 pspill[++y] = pspill[x];
3136 /* Find len = length of merging our dirty list with parent's */
3138 dst[0].mid = 0; /* simplify loops */
3139 if (parent->mt_parent) {
3140 len = x + src[0].mid;
3141 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3142 for (i = x; y && i; y--) {
3143 pgno_t yp = src[y].mid;
3144 while (yp < dst[i].mid)
3146 if (yp == dst[i].mid) {
3151 } else { /* Simplify the above for single-ancestor case */
3152 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3154 /* Merge our dirty list with parent's */
3156 for (i = len; y; dst[i--] = src[y--]) {
3157 pgno_t yp = src[y].mid;
3158 while (yp < dst[x].mid)
3159 dst[i--] = dst[x--];
3160 if (yp == dst[x].mid)
3161 free(dst[x--].mptr);
3163 mdb_tassert(txn, i == x);
3165 free(txn->mt_u.dirty_list);
3166 parent->mt_dirty_room = txn->mt_dirty_room;
3167 if (txn->mt_spill_pgs) {
3168 if (parent->mt_spill_pgs) {
3169 /* TODO: Prevent failure here, so parent does not fail */
3170 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3172 parent->mt_flags |= MDB_TXN_ERROR;
3173 mdb_midl_free(txn->mt_spill_pgs);
3174 mdb_midl_sort(parent->mt_spill_pgs);
3176 parent->mt_spill_pgs = txn->mt_spill_pgs;
3180 parent->mt_child = NULL;
3181 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3186 if (txn != env->me_txn) {
3187 DPUTS("attempt to commit unknown transaction");
3192 mdb_cursors_close(txn, 0);
3194 if (!txn->mt_u.dirty_list[0].mid &&
3195 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3198 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3199 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3201 /* Update DB root pointers */
3202 if (txn->mt_numdbs > 2) {
3206 data.mv_size = sizeof(MDB_db);
3208 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3209 for (i = 2; i < txn->mt_numdbs; i++) {
3210 if (txn->mt_dbflags[i] & DB_DIRTY) {
3211 data.mv_data = &txn->mt_dbs[i];
3212 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3219 rc = mdb_freelist_save(txn);
3223 mdb_midl_free(env->me_pghead);
3224 env->me_pghead = NULL;
3225 if (mdb_midl_shrink(&txn->mt_free_pgs))
3226 env->me_free_pgs = txn->mt_free_pgs;
3232 if ((rc = mdb_page_flush(txn, 0)) ||
3233 (rc = mdb_env_sync(env, 0)) ||
3234 (rc = mdb_env_write_meta(txn)))
3240 mdb_dbis_update(txn, 1);
3243 UNLOCK_MUTEX_W(env);
3253 /** Read the environment parameters of a DB environment before
3254 * mapping it into memory.
3255 * @param[in] env the environment handle
3256 * @param[out] meta address of where to store the meta information
3257 * @return 0 on success, non-zero on failure.
3260 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3266 enum { Size = sizeof(pbuf) };
3268 /* We don't know the page size yet, so use a minimum value.
3269 * Read both meta pages so we can use the latest one.
3272 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3276 memset(&ov, 0, sizeof(ov));
3278 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3279 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3282 rc = pread(env->me_fd, &pbuf, Size, off);
3285 if (rc == 0 && off == 0)
3287 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3288 DPRINTF(("read: %s", mdb_strerror(rc)));
3292 p = (MDB_page *)&pbuf;
3294 if (!F_ISSET(p->mp_flags, P_META)) {
3295 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3300 if (m->mm_magic != MDB_MAGIC) {
3301 DPUTS("meta has invalid magic");
3305 if (m->mm_version != MDB_DATA_VERSION) {
3306 DPRINTF(("database is version %u, expected version %u",
3307 m->mm_version, MDB_DATA_VERSION));
3308 return MDB_VERSION_MISMATCH;
3311 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3318 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3320 meta->mm_magic = MDB_MAGIC;
3321 meta->mm_version = MDB_DATA_VERSION;
3322 meta->mm_mapsize = env->me_mapsize;
3323 meta->mm_psize = env->me_psize;
3324 meta->mm_last_pg = 1;
3325 meta->mm_flags = env->me_flags & 0xffff;
3326 meta->mm_flags |= MDB_INTEGERKEY;
3327 meta->mm_dbs[0].md_root = P_INVALID;
3328 meta->mm_dbs[1].md_root = P_INVALID;
3331 /** Write the environment parameters of a freshly created DB environment.
3332 * @param[in] env the environment handle
3333 * @param[out] meta address of where to store the meta information
3334 * @return 0 on success, non-zero on failure.
3337 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3345 memset(&ov, 0, sizeof(ov));
3346 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3348 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3351 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3352 len = pwrite(fd, ptr, size, pos); \
3353 rc = (len >= 0); } while(0)
3356 DPUTS("writing new meta page");
3358 psize = env->me_psize;
3360 mdb_env_init_meta0(env, meta);
3362 p = calloc(2, psize);
3364 p->mp_flags = P_META;
3365 *(MDB_meta *)METADATA(p) = *meta;
3367 q = (MDB_page *)((char *)p + psize);
3369 q->mp_flags = P_META;
3370 *(MDB_meta *)METADATA(q) = *meta;
3372 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3375 else if ((unsigned) len == psize * 2)
3383 /** Update the environment info to commit a transaction.
3384 * @param[in] txn the transaction that's being committed
3385 * @return 0 on success, non-zero on failure.
3388 mdb_env_write_meta(MDB_txn *txn)
3391 MDB_meta meta, metab, *mp;
3393 int rc, len, toggle;
3402 toggle = txn->mt_txnid & 1;
3403 DPRINTF(("writing meta page %d for root page %"Z"u",
3404 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3407 mp = env->me_metas[toggle];
3409 if (env->me_flags & MDB_WRITEMAP) {
3410 /* Persist any increases of mapsize config */
3411 if (env->me_mapsize > mp->mm_mapsize)
3412 mp->mm_mapsize = env->me_mapsize;
3413 mp->mm_dbs[0] = txn->mt_dbs[0];
3414 mp->mm_dbs[1] = txn->mt_dbs[1];
3415 mp->mm_last_pg = txn->mt_next_pgno - 1;
3416 mp->mm_txnid = txn->mt_txnid;
3417 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3418 unsigned meta_size = env->me_psize;
3419 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3422 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3423 if (meta_size < env->me_os_psize)
3424 meta_size += meta_size;
3429 if (MDB_MSYNC(ptr, meta_size, rc)) {
3436 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3437 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3439 ptr = (char *)&meta;
3440 if (env->me_mapsize > mp->mm_mapsize) {
3441 /* Persist any increases of mapsize config */
3442 meta.mm_mapsize = env->me_mapsize;
3443 off = offsetof(MDB_meta, mm_mapsize);
3445 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3447 len = sizeof(MDB_meta) - off;
3450 meta.mm_dbs[0] = txn->mt_dbs[0];
3451 meta.mm_dbs[1] = txn->mt_dbs[1];
3452 meta.mm_last_pg = txn->mt_next_pgno - 1;
3453 meta.mm_txnid = txn->mt_txnid;
3456 off += env->me_psize;
3459 /* Write to the SYNC fd */
3460 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3461 env->me_fd : env->me_mfd;
3464 memset(&ov, 0, sizeof(ov));
3466 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3470 rc = pwrite(mfd, ptr, len, off);
3473 rc = rc < 0 ? ErrCode() : EIO;
3474 DPUTS("write failed, disk error?");
3475 /* On a failure, the pagecache still contains the new data.
3476 * Write some old data back, to prevent it from being used.
3477 * Use the non-SYNC fd; we know it will fail anyway.
3479 meta.mm_last_pg = metab.mm_last_pg;
3480 meta.mm_txnid = metab.mm_txnid;
3482 memset(&ov, 0, sizeof(ov));
3484 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3486 r2 = pwrite(env->me_fd, ptr, len, off);
3487 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3490 env->me_flags |= MDB_FATAL_ERROR;
3494 /* Memory ordering issues are irrelevant; since the entire writer
3495 * is wrapped by wmutex, all of these changes will become visible
3496 * after the wmutex is unlocked. Since the DB is multi-version,
3497 * readers will get consistent data regardless of how fresh or
3498 * how stale their view of these values is.
3501 env->me_txns->mti_txnid = txn->mt_txnid;
3506 /** Check both meta pages to see which one is newer.
3507 * @param[in] env the environment handle
3508 * @return meta toggle (0 or 1).
3511 mdb_env_pick_meta(const MDB_env *env)
3513 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3517 mdb_env_create(MDB_env **env)
3521 e = calloc(1, sizeof(MDB_env));
3525 e->me_maxreaders = DEFAULT_READERS;
3526 e->me_maxdbs = e->me_numdbs = 2;
3527 e->me_fd = INVALID_HANDLE_VALUE;
3528 e->me_lfd = INVALID_HANDLE_VALUE;
3529 e->me_mfd = INVALID_HANDLE_VALUE;
3530 #ifdef MDB_USE_POSIX_SEM
3531 e->me_rmutex = SEM_FAILED;
3532 e->me_wmutex = SEM_FAILED;
3534 e->me_pid = getpid();
3535 GET_PAGESIZE(e->me_os_psize);
3536 VGMEMP_CREATE(e,0,0);
3542 mdb_env_map(MDB_env *env, void *addr, int newsize)
3545 unsigned int flags = env->me_flags;
3549 LONG sizelo, sizehi;
3552 if (flags & MDB_RDONLY) {
3557 msize = env->me_mapsize;
3558 sizelo = msize & 0xffffffff;
3559 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3562 /* Windows won't create mappings for zero length files.
3563 * Just allocate the maxsize right now.
3566 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3567 || !SetEndOfFile(env->me_fd)
3568 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3571 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3572 PAGE_READWRITE : PAGE_READONLY,
3573 sizehi, sizelo, NULL);
3576 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3577 FILE_MAP_WRITE : FILE_MAP_READ,
3579 rc = env->me_map ? 0 : ErrCode();
3584 int prot = PROT_READ;
3585 if (flags & MDB_WRITEMAP) {
3587 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3590 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3592 if (env->me_map == MAP_FAILED) {
3597 if (flags & MDB_NORDAHEAD) {
3598 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3600 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3602 #ifdef POSIX_MADV_RANDOM
3603 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3604 #endif /* POSIX_MADV_RANDOM */
3605 #endif /* MADV_RANDOM */
3609 /* Can happen because the address argument to mmap() is just a
3610 * hint. mmap() can pick another, e.g. if the range is in use.
3611 * The MAP_FIXED flag would prevent that, but then mmap could
3612 * instead unmap existing pages to make room for the new map.
3614 if (addr && env->me_map != addr)
3615 return EBUSY; /* TODO: Make a new MDB_* error code? */
3617 p = (MDB_page *)env->me_map;
3618 env->me_metas[0] = METADATA(p);
3619 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3625 mdb_env_set_mapsize(MDB_env *env, size_t size)
3627 /* If env is already open, caller is responsible for making
3628 * sure there are no active txns.
3636 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3637 else if (size < env->me_mapsize) {
3638 /* If the configured size is smaller, make sure it's
3639 * still big enough. Silently round up to minimum if not.
3641 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3645 munmap(env->me_map, env->me_mapsize);
3646 env->me_mapsize = size;
3647 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3648 rc = mdb_env_map(env, old, 1);
3652 env->me_mapsize = size;
3654 env->me_maxpg = env->me_mapsize / env->me_psize;
3659 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3663 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3668 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3670 if (env->me_map || readers < 1)
3672 env->me_maxreaders = readers;
3677 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3679 if (!env || !readers)
3681 *readers = env->me_maxreaders;
3685 /** Further setup required for opening an LMDB environment
3688 mdb_env_open2(MDB_env *env)
3690 unsigned int flags = env->me_flags;
3691 int i, newenv = 0, rc;
3695 /* See if we should use QueryLimited */
3697 if ((rc & 0xff) > 5)
3698 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3700 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3703 memset(&meta, 0, sizeof(meta));
3705 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3708 DPUTS("new mdbenv");
3710 env->me_psize = env->me_os_psize;
3711 if (env->me_psize > MAX_PAGESIZE)
3712 env->me_psize = MAX_PAGESIZE;
3714 env->me_psize = meta.mm_psize;
3717 /* Was a mapsize configured? */
3718 if (!env->me_mapsize) {
3719 /* If this is a new environment, take the default,
3720 * else use the size recorded in the existing env.
3722 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3723 } else if (env->me_mapsize < meta.mm_mapsize) {
3724 /* If the configured size is smaller, make sure it's
3725 * still big enough. Silently round up to minimum if not.
3727 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3728 if (env->me_mapsize < minsize)
3729 env->me_mapsize = minsize;
3732 rc = mdb_env_map(env, meta.mm_address, newenv || env->me_mapsize != meta.mm_mapsize);
3737 if (flags & MDB_FIXEDMAP)
3738 meta.mm_address = env->me_map;
3739 i = mdb_env_init_meta(env, &meta);
3740 if (i != MDB_SUCCESS) {
3745 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3746 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3748 #if !(MDB_MAXKEYSIZE)
3749 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3751 env->me_maxpg = env->me_mapsize / env->me_psize;
3755 int toggle = mdb_env_pick_meta(env);
3756 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3758 DPRINTF(("opened database version %u, pagesize %u",
3759 env->me_metas[0]->mm_version, env->me_psize));
3760 DPRINTF(("using meta page %d", toggle));
3761 DPRINTF(("depth: %u", db->md_depth));
3762 DPRINTF(("entries: %"Z"u", db->md_entries));
3763 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3764 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3765 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3766 DPRINTF(("root: %"Z"u", db->md_root));
3774 /** Release a reader thread's slot in the reader lock table.
3775 * This function is called automatically when a thread exits.
3776 * @param[in] ptr This points to the slot in the reader lock table.
3779 mdb_env_reader_dest(void *ptr)
3781 MDB_reader *reader = ptr;
3787 /** Junk for arranging thread-specific callbacks on Windows. This is
3788 * necessarily platform and compiler-specific. Windows supports up
3789 * to 1088 keys. Let's assume nobody opens more than 64 environments
3790 * in a single process, for now. They can override this if needed.
3792 #ifndef MAX_TLS_KEYS
3793 #define MAX_TLS_KEYS 64
3795 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3796 static int mdb_tls_nkeys;
3798 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3802 case DLL_PROCESS_ATTACH: break;
3803 case DLL_THREAD_ATTACH: break;
3804 case DLL_THREAD_DETACH:
3805 for (i=0; i<mdb_tls_nkeys; i++) {
3806 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3808 mdb_env_reader_dest(r);
3812 case DLL_PROCESS_DETACH: break;
3817 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3819 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3823 /* Force some symbol references.
3824 * _tls_used forces the linker to create the TLS directory if not already done
3825 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3827 #pragma comment(linker, "/INCLUDE:_tls_used")
3828 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3829 #pragma const_seg(".CRT$XLB")
3830 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3831 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3834 #pragma comment(linker, "/INCLUDE:__tls_used")
3835 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3836 #pragma data_seg(".CRT$XLB")
3837 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3839 #endif /* WIN 32/64 */
3840 #endif /* !__GNUC__ */
3843 /** Downgrade the exclusive lock on the region back to shared */
3845 mdb_env_share_locks(MDB_env *env, int *excl)
3847 int rc = 0, toggle = mdb_env_pick_meta(env);
3849 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3854 /* First acquire a shared lock. The Unlock will
3855 * then release the existing exclusive lock.
3857 memset(&ov, 0, sizeof(ov));
3858 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3861 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3867 struct flock lock_info;
3868 /* The shared lock replaces the existing lock */
3869 memset((void *)&lock_info, 0, sizeof(lock_info));
3870 lock_info.l_type = F_RDLCK;
3871 lock_info.l_whence = SEEK_SET;
3872 lock_info.l_start = 0;
3873 lock_info.l_len = 1;
3874 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3875 (rc = ErrCode()) == EINTR) ;
3876 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3883 /** Try to get exlusive lock, otherwise shared.
3884 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3887 mdb_env_excl_lock(MDB_env *env, int *excl)
3891 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3895 memset(&ov, 0, sizeof(ov));
3896 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3903 struct flock lock_info;
3904 memset((void *)&lock_info, 0, sizeof(lock_info));
3905 lock_info.l_type = F_WRLCK;
3906 lock_info.l_whence = SEEK_SET;
3907 lock_info.l_start = 0;
3908 lock_info.l_len = 1;
3909 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3910 (rc = ErrCode()) == EINTR) ;
3914 # ifdef MDB_USE_POSIX_SEM
3915 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3918 lock_info.l_type = F_RDLCK;
3919 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3920 (rc = ErrCode()) == EINTR) ;
3930 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3932 * @(#) $Revision: 5.1 $
3933 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3934 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3936 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3940 * Please do not copyright this code. This code is in the public domain.
3942 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3943 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3944 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3945 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3946 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3947 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3948 * PERFORMANCE OF THIS SOFTWARE.
3951 * chongo <Landon Curt Noll> /\oo/\
3952 * http://www.isthe.com/chongo/
3954 * Share and Enjoy! :-)
3957 typedef unsigned long long mdb_hash_t;
3958 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3960 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3961 * @param[in] val value to hash
3962 * @param[in] hval initial value for hash
3963 * @return 64 bit hash
3965 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3966 * hval arg on the first call.
3969 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3971 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3972 unsigned char *end = s + val->mv_size;
3974 * FNV-1a hash each octet of the string
3977 /* xor the bottom with the current octet */
3978 hval ^= (mdb_hash_t)*s++;
3980 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3981 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3982 (hval << 7) + (hval << 8) + (hval << 40);
3984 /* return our new hash value */
3988 /** Hash the string and output the encoded hash.
3989 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3990 * very short name limits. We don't care about the encoding being reversible,
3991 * we just want to preserve as many bits of the input as possible in a
3992 * small printable string.
3993 * @param[in] str string to hash
3994 * @param[out] encbuf an array of 11 chars to hold the hash
3996 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3999 mdb_pack85(unsigned long l, char *out)
4003 for (i=0; i<5; i++) {
4004 *out++ = mdb_a85[l % 85];
4010 mdb_hash_enc(MDB_val *val, char *encbuf)
4012 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4014 mdb_pack85(h, encbuf);
4015 mdb_pack85(h>>32, encbuf+5);
4020 /** Open and/or initialize the lock region for the environment.
4021 * @param[in] env The LMDB environment.
4022 * @param[in] lpath The pathname of the file used for the lock region.
4023 * @param[in] mode The Unix permissions for the file, if we create it.
4024 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4025 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4026 * @return 0 on success, non-zero on failure.
4029 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4032 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4034 # define MDB_ERRCODE_ROFS EROFS
4035 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4036 # define MDB_CLOEXEC O_CLOEXEC
4039 # define MDB_CLOEXEC 0
4046 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4047 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4048 FILE_ATTRIBUTE_NORMAL, NULL);
4050 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4052 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4054 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4059 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4060 /* Lose record locks when exec*() */
4061 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4062 fcntl(env->me_lfd, F_SETFD, fdflags);
4065 if (!(env->me_flags & MDB_NOTLS)) {
4066 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4069 env->me_flags |= MDB_ENV_TXKEY;
4071 /* Windows TLS callbacks need help finding their TLS info. */
4072 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4076 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4080 /* Try to get exclusive lock. If we succeed, then
4081 * nobody is using the lock region and we should initialize it.
4083 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4086 size = GetFileSize(env->me_lfd, NULL);
4088 size = lseek(env->me_lfd, 0, SEEK_END);
4089 if (size == -1) goto fail_errno;
4091 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4092 if (size < rsize && *excl > 0) {
4094 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4095 || !SetEndOfFile(env->me_lfd))
4098 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4102 size = rsize - sizeof(MDB_txninfo);
4103 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4108 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4110 if (!mh) goto fail_errno;
4111 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4113 if (!env->me_txns) goto fail_errno;
4115 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4117 if (m == MAP_FAILED) goto fail_errno;
4123 BY_HANDLE_FILE_INFORMATION stbuf;
4132 if (!mdb_sec_inited) {
4133 InitializeSecurityDescriptor(&mdb_null_sd,
4134 SECURITY_DESCRIPTOR_REVISION);
4135 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4136 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4137 mdb_all_sa.bInheritHandle = FALSE;
4138 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4141 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4142 idbuf.volume = stbuf.dwVolumeSerialNumber;
4143 idbuf.nhigh = stbuf.nFileIndexHigh;
4144 idbuf.nlow = stbuf.nFileIndexLow;
4145 val.mv_data = &idbuf;
4146 val.mv_size = sizeof(idbuf);
4147 mdb_hash_enc(&val, encbuf);
4148 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4149 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4150 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4151 if (!env->me_rmutex) goto fail_errno;
4152 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4153 if (!env->me_wmutex) goto fail_errno;
4154 #elif defined(MDB_USE_POSIX_SEM)
4163 #if defined(__NetBSD__)
4164 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4166 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4167 idbuf.dev = stbuf.st_dev;
4168 idbuf.ino = stbuf.st_ino;
4169 val.mv_data = &idbuf;
4170 val.mv_size = sizeof(idbuf);
4171 mdb_hash_enc(&val, encbuf);
4172 #ifdef MDB_SHORT_SEMNAMES
4173 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4175 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4176 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4177 /* Clean up after a previous run, if needed: Try to
4178 * remove both semaphores before doing anything else.
4180 sem_unlink(env->me_txns->mti_rmname);
4181 sem_unlink(env->me_txns->mti_wmname);
4182 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4183 O_CREAT|O_EXCL, mode, 1);
4184 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4185 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4186 O_CREAT|O_EXCL, mode, 1);
4187 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4188 #else /* MDB_USE_POSIX_SEM */
4189 pthread_mutexattr_t mattr;
4191 if ((rc = pthread_mutexattr_init(&mattr))
4192 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4193 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4194 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4196 pthread_mutexattr_destroy(&mattr);
4197 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4199 env->me_txns->mti_magic = MDB_MAGIC;
4200 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4201 env->me_txns->mti_txnid = 0;
4202 env->me_txns->mti_numreaders = 0;
4205 if (env->me_txns->mti_magic != MDB_MAGIC) {
4206 DPUTS("lock region has invalid magic");
4210 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4211 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4212 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4213 rc = MDB_VERSION_MISMATCH;
4217 if (rc && rc != EACCES && rc != EAGAIN) {
4221 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4222 if (!env->me_rmutex) goto fail_errno;
4223 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4224 if (!env->me_wmutex) goto fail_errno;
4225 #elif defined(MDB_USE_POSIX_SEM)
4226 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4227 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4228 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4229 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4240 /** The name of the lock file in the DB environment */
4241 #define LOCKNAME "/lock.mdb"
4242 /** The name of the data file in the DB environment */
4243 #define DATANAME "/data.mdb"
4244 /** The suffix of the lock file when no subdir is used */
4245 #define LOCKSUFF "-lock"
4246 /** Only a subset of the @ref mdb_env flags can be changed
4247 * at runtime. Changing other flags requires closing the
4248 * environment and re-opening it with the new flags.
4250 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4251 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4252 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4254 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4255 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4259 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4261 int oflags, rc, len, excl = -1;
4262 char *lpath, *dpath;
4264 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4268 if (flags & MDB_NOSUBDIR) {
4269 rc = len + sizeof(LOCKSUFF) + len + 1;
4271 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4276 if (flags & MDB_NOSUBDIR) {
4277 dpath = lpath + len + sizeof(LOCKSUFF);
4278 sprintf(lpath, "%s" LOCKSUFF, path);
4279 strcpy(dpath, path);
4281 dpath = lpath + len + sizeof(LOCKNAME);
4282 sprintf(lpath, "%s" LOCKNAME, path);
4283 sprintf(dpath, "%s" DATANAME, path);
4287 flags |= env->me_flags;
4288 if (flags & MDB_RDONLY) {
4289 /* silently ignore WRITEMAP when we're only getting read access */
4290 flags &= ~MDB_WRITEMAP;
4292 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4293 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4296 env->me_flags = flags |= MDB_ENV_ACTIVE;
4300 env->me_path = strdup(path);
4301 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4302 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4303 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4308 /* For RDONLY, get lockfile after we know datafile exists */
4309 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4310 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4316 if (F_ISSET(flags, MDB_RDONLY)) {
4317 oflags = GENERIC_READ;
4318 len = OPEN_EXISTING;
4320 oflags = GENERIC_READ|GENERIC_WRITE;
4323 mode = FILE_ATTRIBUTE_NORMAL;
4324 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4325 NULL, len, mode, NULL);
4327 if (F_ISSET(flags, MDB_RDONLY))
4330 oflags = O_RDWR | O_CREAT;
4332 env->me_fd = open(dpath, oflags, mode);
4334 if (env->me_fd == INVALID_HANDLE_VALUE) {
4339 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4340 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4345 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4346 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4347 env->me_mfd = env->me_fd;
4349 /* Synchronous fd for meta writes. Needed even with
4350 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4353 len = OPEN_EXISTING;
4354 env->me_mfd = CreateFile(dpath, oflags,
4355 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4356 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4359 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4361 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4366 DPRINTF(("opened dbenv %p", (void *) env));
4368 rc = mdb_env_share_locks(env, &excl);
4372 if (!((flags & MDB_RDONLY) ||
4373 (env->me_pbuf = calloc(1, env->me_psize))))
4379 mdb_env_close0(env, excl);
4385 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4387 mdb_env_close0(MDB_env *env, int excl)
4391 if (!(env->me_flags & MDB_ENV_ACTIVE))
4394 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4395 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4396 free(env->me_dbxs[i].md_name.mv_data);
4399 free(env->me_dbflags);
4402 free(env->me_dirty_list);
4403 mdb_midl_free(env->me_free_pgs);
4405 if (env->me_flags & MDB_ENV_TXKEY) {
4406 pthread_key_delete(env->me_txkey);
4408 /* Delete our key from the global list */
4409 for (i=0; i<mdb_tls_nkeys; i++)
4410 if (mdb_tls_keys[i] == env->me_txkey) {
4411 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4419 munmap(env->me_map, env->me_mapsize);
4421 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4422 (void) close(env->me_mfd);
4423 if (env->me_fd != INVALID_HANDLE_VALUE)
4424 (void) close(env->me_fd);
4426 MDB_PID_T pid = env->me_pid;
4427 /* Clearing readers is done in this function because
4428 * me_txkey with its destructor must be disabled first.
4430 for (i = env->me_numreaders; --i >= 0; )
4431 if (env->me_txns->mti_readers[i].mr_pid == pid)
4432 env->me_txns->mti_readers[i].mr_pid = 0;
4434 if (env->me_rmutex) {
4435 CloseHandle(env->me_rmutex);
4436 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4438 /* Windows automatically destroys the mutexes when
4439 * the last handle closes.
4441 #elif defined(MDB_USE_POSIX_SEM)
4442 if (env->me_rmutex != SEM_FAILED) {
4443 sem_close(env->me_rmutex);
4444 if (env->me_wmutex != SEM_FAILED)
4445 sem_close(env->me_wmutex);
4446 /* If we have the filelock: If we are the
4447 * only remaining user, clean up semaphores.
4450 mdb_env_excl_lock(env, &excl);
4452 sem_unlink(env->me_txns->mti_rmname);
4453 sem_unlink(env->me_txns->mti_wmname);
4457 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4459 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4462 /* Unlock the lockfile. Windows would have unlocked it
4463 * after closing anyway, but not necessarily at once.
4465 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4468 (void) close(env->me_lfd);
4471 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4476 mdb_env_close(MDB_env *env)
4483 VGMEMP_DESTROY(env);
4484 while ((dp = env->me_dpages) != NULL) {
4485 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4486 env->me_dpages = dp->mp_next;
4490 mdb_env_close0(env, 0);
4494 /** Compare two items pointing at aligned size_t's */
4496 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4498 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4499 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4502 /** Compare two items pointing at aligned unsigned int's */
4504 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4506 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4507 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4510 /** Compare two items pointing at unsigned ints of unknown alignment.
4511 * Nodes and keys are guaranteed to be 2-byte aligned.
4514 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4516 #if BYTE_ORDER == LITTLE_ENDIAN
4517 unsigned short *u, *c;
4520 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4521 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4524 } while(!x && u > (unsigned short *)a->mv_data);
4527 unsigned short *u, *c, *end;
4530 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4531 u = (unsigned short *)a->mv_data;
4532 c = (unsigned short *)b->mv_data;
4535 } while(!x && u < end);
4540 /** Compare two items pointing at size_t's of unknown alignment. */
4541 #ifdef MISALIGNED_OK
4542 # define mdb_cmp_clong mdb_cmp_long
4544 # define mdb_cmp_clong mdb_cmp_cint
4547 /** Compare two items lexically */
4549 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4556 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4562 diff = memcmp(a->mv_data, b->mv_data, len);
4563 return diff ? diff : len_diff<0 ? -1 : len_diff;
4566 /** Compare two items in reverse byte order */
4568 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4570 const unsigned char *p1, *p2, *p1_lim;
4574 p1_lim = (const unsigned char *)a->mv_data;
4575 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4576 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4578 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4584 while (p1 > p1_lim) {
4585 diff = *--p1 - *--p2;
4589 return len_diff<0 ? -1 : len_diff;
4592 /** Search for key within a page, using binary search.
4593 * Returns the smallest entry larger or equal to the key.
4594 * If exactp is non-null, stores whether the found entry was an exact match
4595 * in *exactp (1 or 0).
4596 * Updates the cursor index with the index of the found entry.
4597 * If no entry larger or equal to the key is found, returns NULL.
4600 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4602 unsigned int i = 0, nkeys;
4605 MDB_page *mp = mc->mc_pg[mc->mc_top];
4606 MDB_node *node = NULL;
4611 nkeys = NUMKEYS(mp);
4613 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4614 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4617 low = IS_LEAF(mp) ? 0 : 1;
4619 cmp = mc->mc_dbx->md_cmp;
4621 /* Branch pages have no data, so if using integer keys,
4622 * alignment is guaranteed. Use faster mdb_cmp_int.
4624 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4625 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4632 nodekey.mv_size = mc->mc_db->md_pad;
4633 node = NODEPTR(mp, 0); /* fake */
4634 while (low <= high) {
4635 i = (low + high) >> 1;
4636 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4637 rc = cmp(key, &nodekey);
4638 DPRINTF(("found leaf index %u [%s], rc = %i",
4639 i, DKEY(&nodekey), rc));
4648 while (low <= high) {
4649 i = (low + high) >> 1;
4651 node = NODEPTR(mp, i);
4652 nodekey.mv_size = NODEKSZ(node);
4653 nodekey.mv_data = NODEKEY(node);
4655 rc = cmp(key, &nodekey);
4658 DPRINTF(("found leaf index %u [%s], rc = %i",
4659 i, DKEY(&nodekey), rc));
4661 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4662 i, DKEY(&nodekey), NODEPGNO(node), rc));
4673 if (rc > 0) { /* Found entry is less than the key. */
4674 i++; /* Skip to get the smallest entry larger than key. */
4676 node = NODEPTR(mp, i);
4679 *exactp = (rc == 0 && nkeys > 0);
4680 /* store the key index */
4681 mc->mc_ki[mc->mc_top] = i;
4683 /* There is no entry larger or equal to the key. */
4686 /* nodeptr is fake for LEAF2 */
4692 mdb_cursor_adjust(MDB_cursor *mc, func)
4696 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4697 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4704 /** Pop a page off the top of the cursor's stack. */
4706 mdb_cursor_pop(MDB_cursor *mc)
4710 MDB_page *top = mc->mc_pg[mc->mc_top];
4716 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4717 DDBI(mc), (void *) mc));
4721 /** Push a page onto the top of the cursor's stack. */
4723 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4725 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4726 DDBI(mc), (void *) mc));
4728 if (mc->mc_snum >= CURSOR_STACK) {
4729 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4730 return MDB_CURSOR_FULL;
4733 mc->mc_top = mc->mc_snum++;
4734 mc->mc_pg[mc->mc_top] = mp;
4735 mc->mc_ki[mc->mc_top] = 0;
4740 /** Find the address of the page corresponding to a given page number.
4741 * @param[in] txn the transaction for this access.
4742 * @param[in] pgno the page number for the page to retrieve.
4743 * @param[out] ret address of a pointer where the page's address will be stored.
4744 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4745 * @return 0 on success, non-zero on failure.
4748 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4750 MDB_env *env = txn->mt_env;
4754 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4758 MDB_ID2L dl = tx2->mt_u.dirty_list;
4760 /* Spilled pages were dirtied in this txn and flushed
4761 * because the dirty list got full. Bring this page
4762 * back in from the map (but don't unspill it here,
4763 * leave that unless page_touch happens again).
4765 if (tx2->mt_spill_pgs) {
4766 MDB_ID pn = pgno << 1;
4767 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4768 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4769 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4774 unsigned x = mdb_mid2l_search(dl, pgno);
4775 if (x <= dl[0].mid && dl[x].mid == pgno) {
4781 } while ((tx2 = tx2->mt_parent) != NULL);
4784 if (pgno < txn->mt_next_pgno) {
4786 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4788 DPRINTF(("page %"Z"u not found", pgno));
4789 txn->mt_flags |= MDB_TXN_ERROR;
4790 return MDB_PAGE_NOTFOUND;
4800 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4801 * The cursor is at the root page, set up the rest of it.
4804 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4806 MDB_page *mp = mc->mc_pg[mc->mc_top];
4810 while (IS_BRANCH(mp)) {
4814 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4815 mdb_cassert(mc, NUMKEYS(mp) > 1);
4816 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4818 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4820 if (flags & MDB_PS_LAST)
4821 i = NUMKEYS(mp) - 1;
4824 node = mdb_node_search(mc, key, &exact);
4826 i = NUMKEYS(mp) - 1;
4828 i = mc->mc_ki[mc->mc_top];
4830 mdb_cassert(mc, i > 0);
4834 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4837 mdb_cassert(mc, i < NUMKEYS(mp));
4838 node = NODEPTR(mp, i);
4840 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4843 mc->mc_ki[mc->mc_top] = i;
4844 if ((rc = mdb_cursor_push(mc, mp)))
4847 if (flags & MDB_PS_MODIFY) {
4848 if ((rc = mdb_page_touch(mc)) != 0)
4850 mp = mc->mc_pg[mc->mc_top];
4855 DPRINTF(("internal error, index points to a %02X page!?",
4857 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4858 return MDB_CORRUPTED;
4861 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4862 key ? DKEY(key) : "null"));
4863 mc->mc_flags |= C_INITIALIZED;
4864 mc->mc_flags &= ~C_EOF;
4869 /** Search for the lowest key under the current branch page.
4870 * This just bypasses a NUMKEYS check in the current page
4871 * before calling mdb_page_search_root(), because the callers
4872 * are all in situations where the current page is known to
4876 mdb_page_search_lowest(MDB_cursor *mc)
4878 MDB_page *mp = mc->mc_pg[mc->mc_top];
4879 MDB_node *node = NODEPTR(mp, 0);
4882 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4885 mc->mc_ki[mc->mc_top] = 0;
4886 if ((rc = mdb_cursor_push(mc, mp)))
4888 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4891 /** Search for the page a given key should be in.
4892 * Push it and its parent pages on the cursor stack.
4893 * @param[in,out] mc the cursor for this operation.
4894 * @param[in] key the key to search for, or NULL for first/last page.
4895 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4896 * are touched (updated with new page numbers).
4897 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4898 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4899 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4900 * @return 0 on success, non-zero on failure.
4903 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4908 /* Make sure the txn is still viable, then find the root from
4909 * the txn's db table and set it as the root of the cursor's stack.
4911 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4912 DPUTS("transaction has failed, must abort");
4915 /* Make sure we're using an up-to-date root */
4916 if (*mc->mc_dbflag & DB_STALE) {
4918 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4919 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4926 MDB_node *leaf = mdb_node_search(&mc2,
4927 &mc->mc_dbx->md_name, &exact);
4929 return MDB_NOTFOUND;
4930 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4933 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4935 /* The txn may not know this DBI, or another process may
4936 * have dropped and recreated the DB with other flags.
4938 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4939 return MDB_INCOMPATIBLE;
4940 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4942 *mc->mc_dbflag &= ~DB_STALE;
4944 root = mc->mc_db->md_root;
4946 if (root == P_INVALID) { /* Tree is empty. */
4947 DPUTS("tree is empty");
4948 return MDB_NOTFOUND;
4952 mdb_cassert(mc, root > 1);
4953 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4954 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4960 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4961 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4963 if (flags & MDB_PS_MODIFY) {
4964 if ((rc = mdb_page_touch(mc)))
4968 if (flags & MDB_PS_ROOTONLY)
4971 return mdb_page_search_root(mc, key, flags);
4975 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4977 MDB_txn *txn = mc->mc_txn;
4978 pgno_t pg = mp->mp_pgno;
4979 unsigned x = 0, ovpages = mp->mp_pages;
4980 MDB_env *env = txn->mt_env;
4981 MDB_IDL sl = txn->mt_spill_pgs;
4982 MDB_ID pn = pg << 1;
4985 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4986 /* If the page is dirty or on the spill list we just acquired it,
4987 * so we should give it back to our current free list, if any.
4988 * Otherwise put it onto the list of pages we freed in this txn.
4990 * Won't create me_pghead: me_pglast must be inited along with it.
4991 * Unsupported in nested txns: They would need to hide the page
4992 * range in ancestor txns' dirty and spilled lists.
4994 if (env->me_pghead &&
4996 ((mp->mp_flags & P_DIRTY) ||
4997 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5001 MDB_ID2 *dl, ix, iy;
5002 rc = mdb_midl_need(&env->me_pghead, ovpages);
5005 if (!(mp->mp_flags & P_DIRTY)) {
5006 /* This page is no longer spilled */
5013 /* Remove from dirty list */
5014 dl = txn->mt_u.dirty_list;
5016 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5022 mdb_cassert(mc, x > 1);
5024 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5025 txn->mt_flags |= MDB_TXN_ERROR;
5026 return MDB_CORRUPTED;
5029 if (!(env->me_flags & MDB_WRITEMAP))
5030 mdb_dpage_free(env, mp);
5032 /* Insert in me_pghead */
5033 mop = env->me_pghead;
5034 j = mop[0] + ovpages;
5035 for (i = mop[0]; i && mop[i] < pg; i--)
5041 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5045 mc->mc_db->md_overflow_pages -= ovpages;
5049 /** Return the data associated with a given node.
5050 * @param[in] txn The transaction for this operation.
5051 * @param[in] leaf The node being read.
5052 * @param[out] data Updated to point to the node's data.
5053 * @return 0 on success, non-zero on failure.
5056 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5058 MDB_page *omp; /* overflow page */
5062 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5063 data->mv_size = NODEDSZ(leaf);
5064 data->mv_data = NODEDATA(leaf);
5068 /* Read overflow data.
5070 data->mv_size = NODEDSZ(leaf);
5071 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5072 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5073 DPRINTF(("read overflow page %"Z"u failed", pgno));
5076 data->mv_data = METADATA(omp);
5082 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5083 MDB_val *key, MDB_val *data)
5090 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5092 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5095 if (txn->mt_flags & MDB_TXN_ERROR)
5098 mdb_cursor_init(&mc, txn, dbi, &mx);
5099 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5102 /** Find a sibling for a page.
5103 * Replaces the page at the top of the cursor's stack with the
5104 * specified sibling, if one exists.
5105 * @param[in] mc The cursor for this operation.
5106 * @param[in] move_right Non-zero if the right sibling is requested,
5107 * otherwise the left sibling.
5108 * @return 0 on success, non-zero on failure.
5111 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5117 if (mc->mc_snum < 2) {
5118 return MDB_NOTFOUND; /* root has no siblings */
5122 DPRINTF(("parent page is page %"Z"u, index %u",
5123 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5125 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5126 : (mc->mc_ki[mc->mc_top] == 0)) {
5127 DPRINTF(("no more keys left, moving to %s sibling",
5128 move_right ? "right" : "left"));
5129 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5130 /* undo cursor_pop before returning */
5137 mc->mc_ki[mc->mc_top]++;
5139 mc->mc_ki[mc->mc_top]--;
5140 DPRINTF(("just moving to %s index key %u",
5141 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5143 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5145 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5146 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5147 /* mc will be inconsistent if caller does mc_snum++ as above */
5148 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5152 mdb_cursor_push(mc, mp);
5154 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5159 /** Move the cursor to the next data item. */
5161 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5167 if (mc->mc_flags & C_EOF) {
5168 return MDB_NOTFOUND;
5171 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5173 mp = mc->mc_pg[mc->mc_top];
5175 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5176 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5177 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5178 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5179 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5180 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5181 if (rc == MDB_SUCCESS)
5182 MDB_GET_KEY(leaf, key);
5187 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5188 if (op == MDB_NEXT_DUP)
5189 return MDB_NOTFOUND;
5193 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5194 mdb_dbg_pgno(mp), (void *) mc));
5195 if (mc->mc_flags & C_DEL)
5198 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5199 DPUTS("=====> move to next sibling page");
5200 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5201 mc->mc_flags |= C_EOF;
5204 mp = mc->mc_pg[mc->mc_top];
5205 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5207 mc->mc_ki[mc->mc_top]++;
5210 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5211 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5214 key->mv_size = mc->mc_db->md_pad;
5215 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5219 mdb_cassert(mc, IS_LEAF(mp));
5220 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5222 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5223 mdb_xcursor_init1(mc, leaf);
5226 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5229 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5230 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5231 if (rc != MDB_SUCCESS)
5236 MDB_GET_KEY(leaf, key);
5240 /** Move the cursor to the previous data item. */
5242 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5248 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5250 mp = mc->mc_pg[mc->mc_top];
5252 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5253 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5254 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5255 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5256 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5257 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5258 if (rc == MDB_SUCCESS)
5259 MDB_GET_KEY(leaf, key);
5263 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5264 if (op == MDB_PREV_DUP)
5265 return MDB_NOTFOUND;
5270 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5271 mdb_dbg_pgno(mp), (void *) mc));
5273 if (mc->mc_ki[mc->mc_top] == 0) {
5274 DPUTS("=====> move to prev sibling page");
5275 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5278 mp = mc->mc_pg[mc->mc_top];
5279 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5280 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5282 mc->mc_ki[mc->mc_top]--;
5284 mc->mc_flags &= ~C_EOF;
5286 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5287 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5290 key->mv_size = mc->mc_db->md_pad;
5291 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5295 mdb_cassert(mc, IS_LEAF(mp));
5296 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5298 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5299 mdb_xcursor_init1(mc, leaf);
5302 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5305 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5306 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5307 if (rc != MDB_SUCCESS)
5312 MDB_GET_KEY(leaf, key);
5316 /** Set the cursor on a specific data item. */
5318 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5319 MDB_cursor_op op, int *exactp)
5323 MDB_node *leaf = NULL;
5326 if (key->mv_size == 0)
5327 return MDB_BAD_VALSIZE;
5330 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5332 /* See if we're already on the right page */
5333 if (mc->mc_flags & C_INITIALIZED) {
5336 mp = mc->mc_pg[mc->mc_top];
5338 mc->mc_ki[mc->mc_top] = 0;
5339 return MDB_NOTFOUND;
5341 if (mp->mp_flags & P_LEAF2) {
5342 nodekey.mv_size = mc->mc_db->md_pad;
5343 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5345 leaf = NODEPTR(mp, 0);
5346 MDB_GET_KEY2(leaf, nodekey);
5348 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5350 /* Probably happens rarely, but first node on the page
5351 * was the one we wanted.
5353 mc->mc_ki[mc->mc_top] = 0;
5360 unsigned int nkeys = NUMKEYS(mp);
5362 if (mp->mp_flags & P_LEAF2) {
5363 nodekey.mv_data = LEAF2KEY(mp,
5364 nkeys-1, nodekey.mv_size);
5366 leaf = NODEPTR(mp, nkeys-1);
5367 MDB_GET_KEY2(leaf, nodekey);
5369 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5371 /* last node was the one we wanted */
5372 mc->mc_ki[mc->mc_top] = nkeys-1;
5378 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5379 /* This is definitely the right page, skip search_page */
5380 if (mp->mp_flags & P_LEAF2) {
5381 nodekey.mv_data = LEAF2KEY(mp,
5382 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5384 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5385 MDB_GET_KEY2(leaf, nodekey);
5387 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5389 /* current node was the one we wanted */
5399 /* If any parents have right-sibs, search.
5400 * Otherwise, there's nothing further.
5402 for (i=0; i<mc->mc_top; i++)
5404 NUMKEYS(mc->mc_pg[i])-1)
5406 if (i == mc->mc_top) {
5407 /* There are no other pages */
5408 mc->mc_ki[mc->mc_top] = nkeys;
5409 return MDB_NOTFOUND;
5413 /* There are no other pages */
5414 mc->mc_ki[mc->mc_top] = 0;
5415 if (op == MDB_SET_RANGE && !exactp) {
5419 return MDB_NOTFOUND;
5423 rc = mdb_page_search(mc, key, 0);
5424 if (rc != MDB_SUCCESS)
5427 mp = mc->mc_pg[mc->mc_top];
5428 mdb_cassert(mc, IS_LEAF(mp));
5431 leaf = mdb_node_search(mc, key, exactp);
5432 if (exactp != NULL && !*exactp) {
5433 /* MDB_SET specified and not an exact match. */
5434 return MDB_NOTFOUND;
5438 DPUTS("===> inexact leaf not found, goto sibling");
5439 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5440 return rc; /* no entries matched */
5441 mp = mc->mc_pg[mc->mc_top];
5442 mdb_cassert(mc, IS_LEAF(mp));
5443 leaf = NODEPTR(mp, 0);
5447 mc->mc_flags |= C_INITIALIZED;
5448 mc->mc_flags &= ~C_EOF;
5451 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5452 key->mv_size = mc->mc_db->md_pad;
5453 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5458 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5459 mdb_xcursor_init1(mc, leaf);
5462 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5463 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5464 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5467 if (op == MDB_GET_BOTH) {
5473 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5474 if (rc != MDB_SUCCESS)
5477 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5479 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5481 rc = mc->mc_dbx->md_dcmp(data, &d2);
5483 if (op == MDB_GET_BOTH || rc > 0)
5484 return MDB_NOTFOUND;
5491 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5492 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5497 /* The key already matches in all other cases */
5498 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5499 MDB_GET_KEY(leaf, key);
5500 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5505 /** Move the cursor to the first item in the database. */
5507 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5513 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5515 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5516 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5517 if (rc != MDB_SUCCESS)
5520 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5522 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5523 mc->mc_flags |= C_INITIALIZED;
5524 mc->mc_flags &= ~C_EOF;
5526 mc->mc_ki[mc->mc_top] = 0;
5528 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5529 key->mv_size = mc->mc_db->md_pad;
5530 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5535 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5536 mdb_xcursor_init1(mc, leaf);
5537 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5541 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5545 MDB_GET_KEY(leaf, key);
5549 /** Move the cursor to the last item in the database. */
5551 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5557 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5559 if (!(mc->mc_flags & C_EOF)) {
5561 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5562 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5563 if (rc != MDB_SUCCESS)
5566 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5569 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5570 mc->mc_flags |= C_INITIALIZED|C_EOF;
5571 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5573 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5574 key->mv_size = mc->mc_db->md_pad;
5575 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5580 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5581 mdb_xcursor_init1(mc, leaf);
5582 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5586 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5591 MDB_GET_KEY(leaf, key);
5596 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5601 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5606 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5610 case MDB_GET_CURRENT:
5611 if (!(mc->mc_flags & C_INITIALIZED)) {
5614 MDB_page *mp = mc->mc_pg[mc->mc_top];
5615 int nkeys = NUMKEYS(mp);
5616 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5617 mc->mc_ki[mc->mc_top] = nkeys;
5623 key->mv_size = mc->mc_db->md_pad;
5624 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5626 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5627 MDB_GET_KEY(leaf, key);
5629 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5630 if (mc->mc_flags & C_DEL)
5631 mdb_xcursor_init1(mc, leaf);
5632 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5634 rc = mdb_node_read(mc->mc_txn, leaf, data);
5641 case MDB_GET_BOTH_RANGE:
5646 if (mc->mc_xcursor == NULL) {
5647 rc = MDB_INCOMPATIBLE;
5657 rc = mdb_cursor_set(mc, key, data, op,
5658 op == MDB_SET_RANGE ? NULL : &exact);
5661 case MDB_GET_MULTIPLE:
5662 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5666 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5667 rc = MDB_INCOMPATIBLE;
5671 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5672 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5675 case MDB_NEXT_MULTIPLE:
5680 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5681 rc = MDB_INCOMPATIBLE;
5684 if (!(mc->mc_flags & C_INITIALIZED))
5685 rc = mdb_cursor_first(mc, key, data);
5687 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5688 if (rc == MDB_SUCCESS) {
5689 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5692 mx = &mc->mc_xcursor->mx_cursor;
5693 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5695 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5696 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5704 case MDB_NEXT_NODUP:
5705 if (!(mc->mc_flags & C_INITIALIZED))
5706 rc = mdb_cursor_first(mc, key, data);
5708 rc = mdb_cursor_next(mc, key, data, op);
5712 case MDB_PREV_NODUP:
5713 if (!(mc->mc_flags & C_INITIALIZED)) {
5714 rc = mdb_cursor_last(mc, key, data);
5717 mc->mc_flags |= C_INITIALIZED;
5718 mc->mc_ki[mc->mc_top]++;
5720 rc = mdb_cursor_prev(mc, key, data, op);
5723 rc = mdb_cursor_first(mc, key, data);
5726 mfunc = mdb_cursor_first;
5728 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5732 if (mc->mc_xcursor == NULL) {
5733 rc = MDB_INCOMPATIBLE;
5736 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5740 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5743 rc = mdb_cursor_last(mc, key, data);
5746 mfunc = mdb_cursor_last;
5749 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5754 if (mc->mc_flags & C_DEL)
5755 mc->mc_flags ^= C_DEL;
5760 /** Touch all the pages in the cursor stack. Set mc_top.
5761 * Makes sure all the pages are writable, before attempting a write operation.
5762 * @param[in] mc The cursor to operate on.
5765 mdb_cursor_touch(MDB_cursor *mc)
5767 int rc = MDB_SUCCESS;
5769 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5772 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5773 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5776 *mc->mc_dbflag |= DB_DIRTY;
5781 rc = mdb_page_touch(mc);
5782 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5783 mc->mc_top = mc->mc_snum-1;
5788 /** Do not spill pages to disk if txn is getting full, may fail instead */
5789 #define MDB_NOSPILL 0x8000
5792 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5795 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5797 MDB_node *leaf = NULL;
5800 MDB_val xdata, *rdata, dkey, olddata;
5802 int do_sub = 0, insert_key, insert_data;
5803 unsigned int mcount = 0, dcount = 0, nospill;
5806 unsigned int nflags;
5809 if (mc == NULL || key == NULL)
5812 env = mc->mc_txn->mt_env;
5814 /* Check this first so counter will always be zero on any
5817 if (flags & MDB_MULTIPLE) {
5818 dcount = data[1].mv_size;
5819 data[1].mv_size = 0;
5820 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5821 return MDB_INCOMPATIBLE;
5824 nospill = flags & MDB_NOSPILL;
5825 flags &= ~MDB_NOSPILL;
5827 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5828 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5830 if (key->mv_size-1 >= ENV_MAXKEY(env))
5831 return MDB_BAD_VALSIZE;
5833 #if SIZE_MAX > MAXDATASIZE
5834 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5835 return MDB_BAD_VALSIZE;
5837 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5838 return MDB_BAD_VALSIZE;
5841 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5842 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5846 if (flags == MDB_CURRENT) {
5847 if (!(mc->mc_flags & C_INITIALIZED))
5850 } else if (mc->mc_db->md_root == P_INVALID) {
5851 /* new database, cursor has nothing to point to */
5854 mc->mc_flags &= ~C_INITIALIZED;
5859 if (flags & MDB_APPEND) {
5861 rc = mdb_cursor_last(mc, &k2, &d2);
5863 rc = mc->mc_dbx->md_cmp(key, &k2);
5866 mc->mc_ki[mc->mc_top]++;
5868 /* new key is <= last key */
5873 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5875 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5876 DPRINTF(("duplicate key [%s]", DKEY(key)));
5878 return MDB_KEYEXIST;
5880 if (rc && rc != MDB_NOTFOUND)
5884 if (mc->mc_flags & C_DEL)
5885 mc->mc_flags ^= C_DEL;
5887 /* Cursor is positioned, check for room in the dirty list */
5889 if (flags & MDB_MULTIPLE) {
5891 xdata.mv_size = data->mv_size * dcount;
5895 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5899 if (rc == MDB_NO_ROOT) {
5901 /* new database, write a root leaf page */
5902 DPUTS("allocating new root leaf page");
5903 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5906 mdb_cursor_push(mc, np);
5907 mc->mc_db->md_root = np->mp_pgno;
5908 mc->mc_db->md_depth++;
5909 *mc->mc_dbflag |= DB_DIRTY;
5910 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5912 np->mp_flags |= P_LEAF2;
5913 mc->mc_flags |= C_INITIALIZED;
5915 /* make sure all cursor pages are writable */
5916 rc2 = mdb_cursor_touch(mc);
5921 insert_key = insert_data = rc;
5923 /* The key does not exist */
5924 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5925 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5926 LEAFSIZE(key, data) > env->me_nodemax)
5928 /* Too big for a node, insert in sub-DB */
5929 fp_flags = P_LEAF|P_DIRTY;
5931 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5932 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
5936 /* there's only a key anyway, so this is a no-op */
5937 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5939 unsigned int ksize = mc->mc_db->md_pad;
5940 if (key->mv_size != ksize)
5941 return MDB_BAD_VALSIZE;
5942 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5943 memcpy(ptr, key->mv_data, ksize);
5948 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5949 olddata.mv_size = NODEDSZ(leaf);
5950 olddata.mv_data = NODEDATA(leaf);
5953 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5954 /* Prepare (sub-)page/sub-DB to accept the new item,
5955 * if needed. fp: old sub-page or a header faking
5956 * it. mp: new (sub-)page. offset: growth in page
5957 * size. xdata: node data with new page or DB.
5959 unsigned i, offset = 0;
5960 mp = fp = xdata.mv_data = env->me_pbuf;
5961 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5963 /* Was a single item before, must convert now */
5964 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5965 /* Just overwrite the current item */
5966 if (flags == MDB_CURRENT)
5969 #if UINT_MAX < SIZE_MAX
5970 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5971 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
5973 /* does data match? */
5974 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5975 if (flags & MDB_NODUPDATA)
5976 return MDB_KEYEXIST;
5981 /* Back up original data item */
5982 dkey.mv_size = olddata.mv_size;
5983 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5985 /* Make sub-page header for the dup items, with dummy body */
5986 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5987 fp->mp_lower = PAGEHDRSZ;
5988 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5989 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5990 fp->mp_flags |= P_LEAF2;
5991 fp->mp_pad = data->mv_size;
5992 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
5994 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
5995 (dkey.mv_size & 1) + (data->mv_size & 1);
5997 fp->mp_upper = xdata.mv_size;
5998 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
5999 } else if (leaf->mn_flags & F_SUBDATA) {
6000 /* Data is on sub-DB, just store it */
6001 flags |= F_DUPDATA|F_SUBDATA;
6004 /* Data is on sub-page */
6005 fp = olddata.mv_data;
6008 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6009 offset = EVEN(NODESIZE + sizeof(indx_t) +
6013 offset = fp->mp_pad;
6014 if (SIZELEFT(fp) < offset) {
6015 offset *= 4; /* space for 4 more */
6018 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6020 fp->mp_flags |= P_DIRTY;
6021 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6022 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6026 xdata.mv_size = olddata.mv_size + offset;
6029 fp_flags = fp->mp_flags;
6030 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6031 /* Too big for a sub-page, convert to sub-DB */
6032 fp_flags &= ~P_SUBP;
6034 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6035 fp_flags |= P_LEAF2;
6036 dummy.md_pad = fp->mp_pad;
6037 dummy.md_flags = MDB_DUPFIXED;
6038 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6039 dummy.md_flags |= MDB_INTEGERKEY;
6045 dummy.md_branch_pages = 0;
6046 dummy.md_leaf_pages = 1;
6047 dummy.md_overflow_pages = 0;
6048 dummy.md_entries = NUMKEYS(fp);
6049 xdata.mv_size = sizeof(MDB_db);
6050 xdata.mv_data = &dummy;
6051 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6053 offset = env->me_psize - olddata.mv_size;
6054 flags |= F_DUPDATA|F_SUBDATA;
6055 dummy.md_root = mp->mp_pgno;
6058 mp->mp_flags = fp_flags | P_DIRTY;
6059 mp->mp_pad = fp->mp_pad;
6060 mp->mp_lower = fp->mp_lower;
6061 mp->mp_upper = fp->mp_upper + offset;
6062 if (fp_flags & P_LEAF2) {
6063 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6065 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
6066 olddata.mv_size - fp->mp_upper);
6067 for (i=0; i<NUMKEYS(fp); i++)
6068 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6076 mdb_node_del(mc, 0);
6080 /* overflow page overwrites need special handling */
6081 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6084 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6086 memcpy(&pg, olddata.mv_data, sizeof(pg));
6087 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6089 ovpages = omp->mp_pages;
6091 /* Is the ov page large enough? */
6092 if (ovpages >= dpages) {
6093 if (!(omp->mp_flags & P_DIRTY) &&
6094 (level || (env->me_flags & MDB_WRITEMAP)))
6096 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6099 level = 0; /* dirty in this txn or clean */
6102 if (omp->mp_flags & P_DIRTY) {
6103 /* yes, overwrite it. Note in this case we don't
6104 * bother to try shrinking the page if the new data
6105 * is smaller than the overflow threshold.
6108 /* It is writable only in a parent txn */
6109 size_t sz = (size_t) env->me_psize * ovpages, off;
6110 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6116 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6117 mdb_cassert(mc, rc2 == 0);
6118 if (!(flags & MDB_RESERVE)) {
6119 /* Copy end of page, adjusting alignment so
6120 * compiler may copy words instead of bytes.
6122 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6123 memcpy((size_t *)((char *)np + off),
6124 (size_t *)((char *)omp + off), sz - off);
6127 memcpy(np, omp, sz); /* Copy beginning of page */
6130 SETDSZ(leaf, data->mv_size);
6131 if (F_ISSET(flags, MDB_RESERVE))
6132 data->mv_data = METADATA(omp);
6134 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6138 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6140 } else if (data->mv_size == olddata.mv_size) {
6141 /* same size, just replace it. Note that we could
6142 * also reuse this node if the new data is smaller,
6143 * but instead we opt to shrink the node in that case.
6145 if (F_ISSET(flags, MDB_RESERVE))
6146 data->mv_data = olddata.mv_data;
6147 else if (!(mc->mc_flags & C_SUB))
6148 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6150 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6153 mdb_node_del(mc, 0);
6159 nflags = flags & NODE_ADD_FLAGS;
6160 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6161 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6162 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6163 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6165 nflags |= MDB_SPLIT_REPLACE;
6166 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6168 /* There is room already in this leaf page. */
6169 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6170 if (rc == 0 && insert_key) {
6171 /* Adjust other cursors pointing to mp */
6172 MDB_cursor *m2, *m3;
6173 MDB_dbi dbi = mc->mc_dbi;
6174 unsigned i = mc->mc_top;
6175 MDB_page *mp = mc->mc_pg[i];
6177 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6178 if (mc->mc_flags & C_SUB)
6179 m3 = &m2->mc_xcursor->mx_cursor;
6182 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6183 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6190 if (rc == MDB_SUCCESS) {
6191 /* Now store the actual data in the child DB. Note that we're
6192 * storing the user data in the keys field, so there are strict
6193 * size limits on dupdata. The actual data fields of the child
6194 * DB are all zero size.
6202 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6203 if (flags & MDB_CURRENT) {
6204 xflags = MDB_CURRENT|MDB_NOSPILL;
6206 mdb_xcursor_init1(mc, leaf);
6207 xflags = (flags & MDB_NODUPDATA) ?
6208 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6210 /* converted, write the original data first */
6212 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6216 /* Adjust other cursors pointing to mp */
6218 unsigned i = mc->mc_top;
6219 MDB_page *mp = mc->mc_pg[i];
6221 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6222 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6223 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6224 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6225 mdb_xcursor_init1(m2, leaf);
6229 /* we've done our job */
6232 ecount = mc->mc_xcursor->mx_db.md_entries;
6233 if (flags & MDB_APPENDDUP)
6234 xflags |= MDB_APPEND;
6235 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6236 if (flags & F_SUBDATA) {
6237 void *db = NODEDATA(leaf);
6238 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6240 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6242 /* Increment count unless we just replaced an existing item. */
6244 mc->mc_db->md_entries++;
6246 /* Invalidate txn if we created an empty sub-DB */
6249 /* If we succeeded and the key didn't exist before,
6250 * make sure the cursor is marked valid.
6252 mc->mc_flags |= C_INITIALIZED;
6254 if (flags & MDB_MULTIPLE) {
6257 /* let caller know how many succeeded, if any */
6258 data[1].mv_size = mcount;
6259 if (mcount < dcount) {
6260 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6261 insert_key = insert_data = 0;
6268 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6271 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6276 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6282 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6283 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6285 if (!(mc->mc_flags & C_INITIALIZED))
6288 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6289 return MDB_NOTFOUND;
6291 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6294 rc = mdb_cursor_touch(mc);
6298 mp = mc->mc_pg[mc->mc_top];
6301 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6303 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6304 if (flags & MDB_NODUPDATA) {
6305 /* mdb_cursor_del0() will subtract the final entry */
6306 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6308 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6309 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6311 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6314 /* If sub-DB still has entries, we're done */
6315 if (mc->mc_xcursor->mx_db.md_entries) {
6316 if (leaf->mn_flags & F_SUBDATA) {
6317 /* update subDB info */
6318 void *db = NODEDATA(leaf);
6319 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6322 /* shrink fake page */
6323 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6324 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6325 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6326 /* fix other sub-DB cursors pointed at this fake page */
6327 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6328 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6329 if (m2->mc_pg[mc->mc_top] == mp &&
6330 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6331 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6334 mc->mc_db->md_entries--;
6335 mc->mc_flags |= C_DEL;
6338 /* otherwise fall thru and delete the sub-DB */
6341 if (leaf->mn_flags & F_SUBDATA) {
6342 /* add all the child DB's pages to the free list */
6343 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6349 /* add overflow pages to free list */
6350 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6354 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6355 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6356 (rc = mdb_ovpage_free(mc, omp)))
6361 return mdb_cursor_del0(mc);
6364 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6368 /** Allocate and initialize new pages for a database.
6369 * @param[in] mc a cursor on the database being added to.
6370 * @param[in] flags flags defining what type of page is being allocated.
6371 * @param[in] num the number of pages to allocate. This is usually 1,
6372 * unless allocating overflow pages for a large record.
6373 * @param[out] mp Address of a page, or NULL on failure.
6374 * @return 0 on success, non-zero on failure.
6377 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6382 if ((rc = mdb_page_alloc(mc, num, &np)))
6384 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6385 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6386 np->mp_flags = flags | P_DIRTY;
6387 np->mp_lower = PAGEHDRSZ;
6388 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6391 mc->mc_db->md_branch_pages++;
6392 else if (IS_LEAF(np))
6393 mc->mc_db->md_leaf_pages++;
6394 else if (IS_OVERFLOW(np)) {
6395 mc->mc_db->md_overflow_pages += num;
6403 /** Calculate the size of a leaf node.
6404 * The size depends on the environment's page size; if a data item
6405 * is too large it will be put onto an overflow page and the node
6406 * size will only include the key and not the data. Sizes are always
6407 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6408 * of the #MDB_node headers.
6409 * @param[in] env The environment handle.
6410 * @param[in] key The key for the node.
6411 * @param[in] data The data for the node.
6412 * @return The number of bytes needed to store the node.
6415 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6419 sz = LEAFSIZE(key, data);
6420 if (sz > env->me_nodemax) {
6421 /* put on overflow page */
6422 sz -= data->mv_size - sizeof(pgno_t);
6425 return EVEN(sz + sizeof(indx_t));
6428 /** Calculate the size of a branch node.
6429 * The size should depend on the environment's page size but since
6430 * we currently don't support spilling large keys onto overflow
6431 * pages, it's simply the size of the #MDB_node header plus the
6432 * size of the key. Sizes are always rounded up to an even number
6433 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6434 * @param[in] env The environment handle.
6435 * @param[in] key The key for the node.
6436 * @return The number of bytes needed to store the node.
6439 mdb_branch_size(MDB_env *env, MDB_val *key)
6444 if (sz > env->me_nodemax) {
6445 /* put on overflow page */
6446 /* not implemented */
6447 /* sz -= key->size - sizeof(pgno_t); */
6450 return sz + sizeof(indx_t);
6453 /** Add a node to the page pointed to by the cursor.
6454 * @param[in] mc The cursor for this operation.
6455 * @param[in] indx The index on the page where the new node should be added.
6456 * @param[in] key The key for the new node.
6457 * @param[in] data The data for the new node, if any.
6458 * @param[in] pgno The page number, if adding a branch node.
6459 * @param[in] flags Flags for the node.
6460 * @return 0 on success, non-zero on failure. Possible errors are:
6462 * <li>ENOMEM - failed to allocate overflow pages for the node.
6463 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6464 * should never happen since all callers already calculate the
6465 * page's free space before calling this function.
6469 mdb_node_add(MDB_cursor *mc, indx_t indx,
6470 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6473 size_t node_size = NODESIZE;
6477 MDB_page *mp = mc->mc_pg[mc->mc_top];
6478 MDB_page *ofp = NULL; /* overflow page */
6481 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6483 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6484 IS_LEAF(mp) ? "leaf" : "branch",
6485 IS_SUBP(mp) ? "sub-" : "",
6486 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6487 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6490 /* Move higher keys up one slot. */
6491 int ksize = mc->mc_db->md_pad, dif;
6492 char *ptr = LEAF2KEY(mp, indx, ksize);
6493 dif = NUMKEYS(mp) - indx;
6495 memmove(ptr+ksize, ptr, dif*ksize);
6496 /* insert new key */
6497 memcpy(ptr, key->mv_data, ksize);
6499 /* Just using these for counting */
6500 mp->mp_lower += sizeof(indx_t);
6501 mp->mp_upper -= ksize - sizeof(indx_t);
6505 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6507 node_size += key->mv_size;
6509 mdb_cassert(mc, data);
6510 if (F_ISSET(flags, F_BIGDATA)) {
6511 /* Data already on overflow page. */
6512 node_size += sizeof(pgno_t);
6513 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6514 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6516 /* Put data on overflow page. */
6517 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6518 data->mv_size, node_size+data->mv_size));
6519 node_size = EVEN(node_size + sizeof(pgno_t));
6520 if ((ssize_t)node_size > room)
6522 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6524 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6528 node_size += data->mv_size;
6531 node_size = EVEN(node_size);
6532 if ((ssize_t)node_size > room)
6536 /* Move higher pointers up one slot. */
6537 for (i = NUMKEYS(mp); i > indx; i--)
6538 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6540 /* Adjust free space offsets. */
6541 ofs = mp->mp_upper - node_size;
6542 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6543 mp->mp_ptrs[indx] = ofs;
6545 mp->mp_lower += sizeof(indx_t);
6547 /* Write the node data. */
6548 node = NODEPTR(mp, indx);
6549 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6550 node->mn_flags = flags;
6552 SETDSZ(node,data->mv_size);
6557 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6560 mdb_cassert(mc, key);
6562 if (F_ISSET(flags, F_BIGDATA))
6563 memcpy(node->mn_data + key->mv_size, data->mv_data,
6565 else if (F_ISSET(flags, MDB_RESERVE))
6566 data->mv_data = node->mn_data + key->mv_size;
6568 memcpy(node->mn_data + key->mv_size, data->mv_data,
6571 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6573 if (F_ISSET(flags, MDB_RESERVE))
6574 data->mv_data = METADATA(ofp);
6576 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6583 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6584 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6585 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6586 DPRINTF(("node size = %"Z"u", node_size));
6587 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6588 return MDB_PAGE_FULL;
6591 /** Delete the specified node from a page.
6592 * @param[in] mc Cursor pointing to the node to delete.
6593 * @param[in] ksize The size of a node. Only used if the page is
6594 * part of a #MDB_DUPFIXED database.
6597 mdb_node_del(MDB_cursor *mc, int ksize)
6599 MDB_page *mp = mc->mc_pg[mc->mc_top];
6600 indx_t indx = mc->mc_ki[mc->mc_top];
6602 indx_t i, j, numkeys, ptr;
6606 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6607 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6608 numkeys = NUMKEYS(mp);
6609 mdb_cassert(mc, indx < numkeys);
6612 int x = numkeys - 1 - indx;
6613 base = LEAF2KEY(mp, indx, ksize);
6615 memmove(base, base + ksize, x * ksize);
6616 mp->mp_lower -= sizeof(indx_t);
6617 mp->mp_upper += ksize - sizeof(indx_t);
6621 node = NODEPTR(mp, indx);
6622 sz = NODESIZE + node->mn_ksize;
6624 if (F_ISSET(node->mn_flags, F_BIGDATA))
6625 sz += sizeof(pgno_t);
6627 sz += NODEDSZ(node);
6631 ptr = mp->mp_ptrs[indx];
6632 for (i = j = 0; i < numkeys; i++) {
6634 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6635 if (mp->mp_ptrs[i] < ptr)
6636 mp->mp_ptrs[j] += sz;
6641 base = (char *)mp + mp->mp_upper;
6642 memmove(base + sz, base, ptr - mp->mp_upper);
6644 mp->mp_lower -= sizeof(indx_t);
6648 /** Compact the main page after deleting a node on a subpage.
6649 * @param[in] mp The main page to operate on.
6650 * @param[in] indx The index of the subpage on the main page.
6653 mdb_node_shrink(MDB_page *mp, indx_t indx)
6659 indx_t i, numkeys, ptr;
6661 node = NODEPTR(mp, indx);
6662 sp = (MDB_page *)NODEDATA(node);
6663 delta = SIZELEFT(sp);
6664 xp = (MDB_page *)((char *)sp + delta);
6666 /* shift subpage upward */
6668 nsize = NUMKEYS(sp) * sp->mp_pad;
6670 return; /* do not make the node uneven-sized */
6671 memmove(METADATA(xp), METADATA(sp), nsize);
6674 numkeys = NUMKEYS(sp);
6675 for (i=numkeys-1; i>=0; i--)
6676 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6678 xp->mp_upper = sp->mp_lower;
6679 xp->mp_lower = sp->mp_lower;
6680 xp->mp_flags = sp->mp_flags;
6681 xp->mp_pad = sp->mp_pad;
6682 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6684 nsize = NODEDSZ(node) - delta;
6685 SETDSZ(node, nsize);
6687 /* shift lower nodes upward */
6688 ptr = mp->mp_ptrs[indx];
6689 numkeys = NUMKEYS(mp);
6690 for (i = 0; i < numkeys; i++) {
6691 if (mp->mp_ptrs[i] <= ptr)
6692 mp->mp_ptrs[i] += delta;
6695 base = (char *)mp + mp->mp_upper;
6696 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6697 mp->mp_upper += delta;
6700 /** Initial setup of a sorted-dups cursor.
6701 * Sorted duplicates are implemented as a sub-database for the given key.
6702 * The duplicate data items are actually keys of the sub-database.
6703 * Operations on the duplicate data items are performed using a sub-cursor
6704 * initialized when the sub-database is first accessed. This function does
6705 * the preliminary setup of the sub-cursor, filling in the fields that
6706 * depend only on the parent DB.
6707 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6710 mdb_xcursor_init0(MDB_cursor *mc)
6712 MDB_xcursor *mx = mc->mc_xcursor;
6714 mx->mx_cursor.mc_xcursor = NULL;
6715 mx->mx_cursor.mc_txn = mc->mc_txn;
6716 mx->mx_cursor.mc_db = &mx->mx_db;
6717 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6718 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6719 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6720 mx->mx_cursor.mc_snum = 0;
6721 mx->mx_cursor.mc_top = 0;
6722 mx->mx_cursor.mc_flags = C_SUB;
6723 mx->mx_dbx.md_name.mv_size = 0;
6724 mx->mx_dbx.md_name.mv_data = NULL;
6725 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6726 mx->mx_dbx.md_dcmp = NULL;
6727 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6730 /** Final setup of a sorted-dups cursor.
6731 * Sets up the fields that depend on the data from the main cursor.
6732 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6733 * @param[in] node The data containing the #MDB_db record for the
6734 * sorted-dup database.
6737 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6739 MDB_xcursor *mx = mc->mc_xcursor;
6741 if (node->mn_flags & F_SUBDATA) {
6742 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6743 mx->mx_cursor.mc_pg[0] = 0;
6744 mx->mx_cursor.mc_snum = 0;
6745 mx->mx_cursor.mc_top = 0;
6746 mx->mx_cursor.mc_flags = C_SUB;
6748 MDB_page *fp = NODEDATA(node);
6749 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6750 mx->mx_db.md_flags = 0;
6751 mx->mx_db.md_depth = 1;
6752 mx->mx_db.md_branch_pages = 0;
6753 mx->mx_db.md_leaf_pages = 1;
6754 mx->mx_db.md_overflow_pages = 0;
6755 mx->mx_db.md_entries = NUMKEYS(fp);
6756 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6757 mx->mx_cursor.mc_snum = 1;
6758 mx->mx_cursor.mc_top = 0;
6759 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6760 mx->mx_cursor.mc_pg[0] = fp;
6761 mx->mx_cursor.mc_ki[0] = 0;
6762 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6763 mx->mx_db.md_flags = MDB_DUPFIXED;
6764 mx->mx_db.md_pad = fp->mp_pad;
6765 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6766 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6769 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6770 mx->mx_db.md_root));
6771 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6772 #if UINT_MAX < SIZE_MAX
6773 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6774 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6778 /** Initialize a cursor for a given transaction and database. */
6780 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6783 mc->mc_backup = NULL;
6786 mc->mc_db = &txn->mt_dbs[dbi];
6787 mc->mc_dbx = &txn->mt_dbxs[dbi];
6788 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6793 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6794 mdb_tassert(txn, mx != NULL);
6795 mc->mc_xcursor = mx;
6796 mdb_xcursor_init0(mc);
6798 mc->mc_xcursor = NULL;
6800 if (*mc->mc_dbflag & DB_STALE) {
6801 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6806 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6809 size_t size = sizeof(MDB_cursor);
6811 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6814 if (txn->mt_flags & MDB_TXN_ERROR)
6817 /* Allow read access to the freelist */
6818 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6821 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6822 size += sizeof(MDB_xcursor);
6824 if ((mc = malloc(size)) != NULL) {
6825 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6826 if (txn->mt_cursors) {
6827 mc->mc_next = txn->mt_cursors[dbi];
6828 txn->mt_cursors[dbi] = mc;
6829 mc->mc_flags |= C_UNTRACK;
6841 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6843 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
6846 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6849 if (txn->mt_flags & MDB_TXN_ERROR)
6852 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6856 /* Return the count of duplicate data items for the current key */
6858 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6862 if (mc == NULL || countp == NULL)
6865 if (mc->mc_xcursor == NULL)
6866 return MDB_INCOMPATIBLE;
6868 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6871 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6872 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6875 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6878 *countp = mc->mc_xcursor->mx_db.md_entries;
6884 mdb_cursor_close(MDB_cursor *mc)
6886 if (mc && !mc->mc_backup) {
6887 /* remove from txn, if tracked */
6888 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6889 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6890 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6892 *prev = mc->mc_next;
6899 mdb_cursor_txn(MDB_cursor *mc)
6901 if (!mc) return NULL;
6906 mdb_cursor_dbi(MDB_cursor *mc)
6911 /** Replace the key for a branch node with a new key.
6912 * @param[in] mc Cursor pointing to the node to operate on.
6913 * @param[in] key The new key to use.
6914 * @return 0 on success, non-zero on failure.
6917 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6923 int delta, ksize, oksize;
6924 indx_t ptr, i, numkeys, indx;
6927 indx = mc->mc_ki[mc->mc_top];
6928 mp = mc->mc_pg[mc->mc_top];
6929 node = NODEPTR(mp, indx);
6930 ptr = mp->mp_ptrs[indx];
6934 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6935 k2.mv_data = NODEKEY(node);
6936 k2.mv_size = node->mn_ksize;
6937 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6939 mdb_dkey(&k2, kbuf2),
6945 /* Sizes must be 2-byte aligned. */
6946 ksize = EVEN(key->mv_size);
6947 oksize = EVEN(node->mn_ksize);
6948 delta = ksize - oksize;
6950 /* Shift node contents if EVEN(key length) changed. */
6952 if (delta > 0 && SIZELEFT(mp) < delta) {
6954 /* not enough space left, do a delete and split */
6955 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6956 pgno = NODEPGNO(node);
6957 mdb_node_del(mc, 0);
6958 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6961 numkeys = NUMKEYS(mp);
6962 for (i = 0; i < numkeys; i++) {
6963 if (mp->mp_ptrs[i] <= ptr)
6964 mp->mp_ptrs[i] -= delta;
6967 base = (char *)mp + mp->mp_upper;
6968 len = ptr - mp->mp_upper + NODESIZE;
6969 memmove(base - delta, base, len);
6970 mp->mp_upper -= delta;
6972 node = NODEPTR(mp, indx);
6975 /* But even if no shift was needed, update ksize */
6976 if (node->mn_ksize != key->mv_size)
6977 node->mn_ksize = key->mv_size;
6980 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6986 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6988 /** Move a node from csrc to cdst.
6991 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6998 unsigned short flags;
7002 /* Mark src and dst as dirty. */
7003 if ((rc = mdb_page_touch(csrc)) ||
7004 (rc = mdb_page_touch(cdst)))
7007 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7008 key.mv_size = csrc->mc_db->md_pad;
7009 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7011 data.mv_data = NULL;
7015 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7016 mdb_cassert(csrc, !((size_t)srcnode & 1));
7017 srcpg = NODEPGNO(srcnode);
7018 flags = srcnode->mn_flags;
7019 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7020 unsigned int snum = csrc->mc_snum;
7022 /* must find the lowest key below src */
7023 rc = mdb_page_search_lowest(csrc);
7026 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7027 key.mv_size = csrc->mc_db->md_pad;
7028 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7030 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7031 key.mv_size = NODEKSZ(s2);
7032 key.mv_data = NODEKEY(s2);
7034 csrc->mc_snum = snum--;
7035 csrc->mc_top = snum;
7037 key.mv_size = NODEKSZ(srcnode);
7038 key.mv_data = NODEKEY(srcnode);
7040 data.mv_size = NODEDSZ(srcnode);
7041 data.mv_data = NODEDATA(srcnode);
7043 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7044 unsigned int snum = cdst->mc_snum;
7047 /* must find the lowest key below dst */
7048 mdb_cursor_copy(cdst, &mn);
7049 rc = mdb_page_search_lowest(&mn);
7052 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7053 bkey.mv_size = mn.mc_db->md_pad;
7054 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7056 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7057 bkey.mv_size = NODEKSZ(s2);
7058 bkey.mv_data = NODEKEY(s2);
7060 mn.mc_snum = snum--;
7063 rc = mdb_update_key(&mn, &bkey);
7068 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7069 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7070 csrc->mc_ki[csrc->mc_top],
7072 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7073 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7075 /* Add the node to the destination page.
7077 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7078 if (rc != MDB_SUCCESS)
7081 /* Delete the node from the source page.
7083 mdb_node_del(csrc, key.mv_size);
7086 /* Adjust other cursors pointing to mp */
7087 MDB_cursor *m2, *m3;
7088 MDB_dbi dbi = csrc->mc_dbi;
7089 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7091 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7092 if (csrc->mc_flags & C_SUB)
7093 m3 = &m2->mc_xcursor->mx_cursor;
7096 if (m3 == csrc) continue;
7097 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7098 csrc->mc_ki[csrc->mc_top]) {
7099 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7100 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7105 /* Update the parent separators.
7107 if (csrc->mc_ki[csrc->mc_top] == 0) {
7108 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7109 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7110 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7112 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7113 key.mv_size = NODEKSZ(srcnode);
7114 key.mv_data = NODEKEY(srcnode);
7116 DPRINTF(("update separator for source page %"Z"u to [%s]",
7117 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7118 mdb_cursor_copy(csrc, &mn);
7121 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7124 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7126 indx_t ix = csrc->mc_ki[csrc->mc_top];
7127 nullkey.mv_size = 0;
7128 csrc->mc_ki[csrc->mc_top] = 0;
7129 rc = mdb_update_key(csrc, &nullkey);
7130 csrc->mc_ki[csrc->mc_top] = ix;
7131 mdb_cassert(csrc, rc == MDB_SUCCESS);
7135 if (cdst->mc_ki[cdst->mc_top] == 0) {
7136 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7137 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7138 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7140 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7141 key.mv_size = NODEKSZ(srcnode);
7142 key.mv_data = NODEKEY(srcnode);
7144 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7145 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7146 mdb_cursor_copy(cdst, &mn);
7149 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7152 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7154 indx_t ix = cdst->mc_ki[cdst->mc_top];
7155 nullkey.mv_size = 0;
7156 cdst->mc_ki[cdst->mc_top] = 0;
7157 rc = mdb_update_key(cdst, &nullkey);
7158 cdst->mc_ki[cdst->mc_top] = ix;
7159 mdb_cassert(csrc, rc == MDB_SUCCESS);
7166 /** Merge one page into another.
7167 * The nodes from the page pointed to by \b csrc will
7168 * be copied to the page pointed to by \b cdst and then
7169 * the \b csrc page will be freed.
7170 * @param[in] csrc Cursor pointing to the source page.
7171 * @param[in] cdst Cursor pointing to the destination page.
7172 * @return 0 on success, non-zero on failure.
7175 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7177 MDB_page *psrc, *pdst;
7184 psrc = csrc->mc_pg[csrc->mc_top];
7185 pdst = cdst->mc_pg[cdst->mc_top];
7187 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7189 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7190 mdb_cassert(csrc, cdst->mc_snum > 1);
7192 /* Mark dst as dirty. */
7193 if ((rc = mdb_page_touch(cdst)))
7196 /* Move all nodes from src to dst.
7198 j = nkeys = NUMKEYS(pdst);
7199 if (IS_LEAF2(psrc)) {
7200 key.mv_size = csrc->mc_db->md_pad;
7201 key.mv_data = METADATA(psrc);
7202 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7203 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7204 if (rc != MDB_SUCCESS)
7206 key.mv_data = (char *)key.mv_data + key.mv_size;
7209 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7210 srcnode = NODEPTR(psrc, i);
7211 if (i == 0 && IS_BRANCH(psrc)) {
7214 mdb_cursor_copy(csrc, &mn);
7215 /* must find the lowest key below src */
7216 rc = mdb_page_search_lowest(&mn);
7219 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7220 key.mv_size = mn.mc_db->md_pad;
7221 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7223 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7224 key.mv_size = NODEKSZ(s2);
7225 key.mv_data = NODEKEY(s2);
7228 key.mv_size = srcnode->mn_ksize;
7229 key.mv_data = NODEKEY(srcnode);
7232 data.mv_size = NODEDSZ(srcnode);
7233 data.mv_data = NODEDATA(srcnode);
7234 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7235 if (rc != MDB_SUCCESS)
7240 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7241 pdst->mp_pgno, NUMKEYS(pdst),
7242 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7244 /* Unlink the src page from parent and add to free list.
7247 mdb_node_del(csrc, 0);
7248 if (csrc->mc_ki[csrc->mc_top] == 0) {
7250 rc = mdb_update_key(csrc, &key);
7258 psrc = csrc->mc_pg[csrc->mc_top];
7259 /* If not operating on FreeDB, allow this page to be reused
7260 * in this txn. Otherwise just add to free list.
7262 rc = mdb_page_loose(csrc, psrc);
7266 csrc->mc_db->md_leaf_pages--;
7268 csrc->mc_db->md_branch_pages--;
7270 /* Adjust other cursors pointing to mp */
7271 MDB_cursor *m2, *m3;
7272 MDB_dbi dbi = csrc->mc_dbi;
7274 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7275 if (csrc->mc_flags & C_SUB)
7276 m3 = &m2->mc_xcursor->mx_cursor;
7279 if (m3 == csrc) continue;
7280 if (m3->mc_snum < csrc->mc_snum) continue;
7281 if (m3->mc_pg[csrc->mc_top] == psrc) {
7282 m3->mc_pg[csrc->mc_top] = pdst;
7283 m3->mc_ki[csrc->mc_top] += nkeys;
7288 unsigned int snum = cdst->mc_snum;
7289 uint16_t depth = cdst->mc_db->md_depth;
7290 mdb_cursor_pop(cdst);
7291 rc = mdb_rebalance(cdst);
7292 /* Did the tree shrink? */
7293 if (depth > cdst->mc_db->md_depth)
7295 cdst->mc_snum = snum;
7296 cdst->mc_top = snum-1;
7301 /** Copy the contents of a cursor.
7302 * @param[in] csrc The cursor to copy from.
7303 * @param[out] cdst The cursor to copy to.
7306 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7310 cdst->mc_txn = csrc->mc_txn;
7311 cdst->mc_dbi = csrc->mc_dbi;
7312 cdst->mc_db = csrc->mc_db;
7313 cdst->mc_dbx = csrc->mc_dbx;
7314 cdst->mc_snum = csrc->mc_snum;
7315 cdst->mc_top = csrc->mc_top;
7316 cdst->mc_flags = csrc->mc_flags;
7318 for (i=0; i<csrc->mc_snum; i++) {
7319 cdst->mc_pg[i] = csrc->mc_pg[i];
7320 cdst->mc_ki[i] = csrc->mc_ki[i];
7324 /** Rebalance the tree after a delete operation.
7325 * @param[in] mc Cursor pointing to the page where rebalancing
7327 * @return 0 on success, non-zero on failure.
7330 mdb_rebalance(MDB_cursor *mc)
7334 unsigned int ptop, minkeys;
7338 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7339 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7340 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7341 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7342 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7344 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7345 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7346 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7347 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7351 if (mc->mc_snum < 2) {
7352 MDB_page *mp = mc->mc_pg[0];
7354 DPUTS("Can't rebalance a subpage, ignoring");
7357 if (NUMKEYS(mp) == 0) {
7358 DPUTS("tree is completely empty");
7359 mc->mc_db->md_root = P_INVALID;
7360 mc->mc_db->md_depth = 0;
7361 mc->mc_db->md_leaf_pages = 0;
7362 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7365 /* Adjust cursors pointing to mp */
7368 mc->mc_flags &= ~C_INITIALIZED;
7370 MDB_cursor *m2, *m3;
7371 MDB_dbi dbi = mc->mc_dbi;
7373 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7374 if (mc->mc_flags & C_SUB)
7375 m3 = &m2->mc_xcursor->mx_cursor;
7378 if (m3->mc_snum < mc->mc_snum) continue;
7379 if (m3->mc_pg[0] == mp) {
7382 m3->mc_flags &= ~C_INITIALIZED;
7386 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7388 DPUTS("collapsing root page!");
7389 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7392 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7393 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7396 mc->mc_db->md_depth--;
7397 mc->mc_db->md_branch_pages--;
7398 mc->mc_ki[0] = mc->mc_ki[1];
7399 for (i = 1; i<mc->mc_db->md_depth; i++) {
7400 mc->mc_pg[i] = mc->mc_pg[i+1];
7401 mc->mc_ki[i] = mc->mc_ki[i+1];
7404 /* Adjust other cursors pointing to mp */
7405 MDB_cursor *m2, *m3;
7406 MDB_dbi dbi = mc->mc_dbi;
7408 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7409 if (mc->mc_flags & C_SUB)
7410 m3 = &m2->mc_xcursor->mx_cursor;
7413 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7414 if (m3->mc_pg[0] == mp) {
7417 for (i=0; i<m3->mc_snum; i++) {
7418 m3->mc_pg[i] = m3->mc_pg[i+1];
7419 m3->mc_ki[i] = m3->mc_ki[i+1];
7425 DPUTS("root page doesn't need rebalancing");
7429 /* The parent (branch page) must have at least 2 pointers,
7430 * otherwise the tree is invalid.
7432 ptop = mc->mc_top-1;
7433 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7435 /* Leaf page fill factor is below the threshold.
7436 * Try to move keys from left or right neighbor, or
7437 * merge with a neighbor page.
7442 mdb_cursor_copy(mc, &mn);
7443 mn.mc_xcursor = NULL;
7445 oldki = mc->mc_ki[mc->mc_top];
7446 if (mc->mc_ki[ptop] == 0) {
7447 /* We're the leftmost leaf in our parent.
7449 DPUTS("reading right neighbor");
7451 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7452 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7455 mn.mc_ki[mn.mc_top] = 0;
7456 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7458 /* There is at least one neighbor to the left.
7460 DPUTS("reading left neighbor");
7462 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7463 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7466 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7467 mc->mc_ki[mc->mc_top] = 0;
7470 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7471 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7472 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7474 /* If the neighbor page is above threshold and has enough keys,
7475 * move one key from it. Otherwise we should try to merge them.
7476 * (A branch page must never have less than 2 keys.)
7478 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7479 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7480 rc = mdb_node_move(&mn, mc);
7481 if (mc->mc_ki[ptop]) {
7485 if (mc->mc_ki[ptop] == 0) {
7486 rc = mdb_page_merge(&mn, mc);
7488 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7489 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7490 rc = mdb_page_merge(mc, &mn);
7491 mdb_cursor_copy(&mn, mc);
7493 mc->mc_flags &= ~C_EOF;
7495 mc->mc_ki[mc->mc_top] = oldki;
7499 /** Complete a delete operation started by #mdb_cursor_del(). */
7501 mdb_cursor_del0(MDB_cursor *mc)
7508 ki = mc->mc_ki[mc->mc_top];
7509 mdb_node_del(mc, mc->mc_db->md_pad);
7510 mc->mc_db->md_entries--;
7511 rc = mdb_rebalance(mc);
7513 if (rc == MDB_SUCCESS) {
7514 MDB_cursor *m2, *m3;
7515 MDB_dbi dbi = mc->mc_dbi;
7517 mp = mc->mc_pg[mc->mc_top];
7518 nkeys = NUMKEYS(mp);
7520 /* if mc points past last node in page, find next sibling */
7521 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7522 rc = mdb_cursor_sibling(mc, 1);
7523 if (rc == MDB_NOTFOUND)
7527 /* Adjust other cursors pointing to mp */
7528 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7529 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7530 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7532 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7534 if (m3->mc_pg[mc->mc_top] == mp) {
7535 if (m3->mc_ki[mc->mc_top] >= ki) {
7536 m3->mc_flags |= C_DEL;
7537 if (m3->mc_ki[mc->mc_top] > ki)
7538 m3->mc_ki[mc->mc_top]--;
7540 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7541 rc = mdb_cursor_sibling(m3, 1);
7542 if (rc == MDB_NOTFOUND)
7547 mc->mc_flags |= C_DEL;
7551 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7556 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7557 MDB_val *key, MDB_val *data)
7559 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7562 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7563 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7565 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7566 /* must ignore any data */
7570 return mdb_del0(txn, dbi, key, data, 0);
7574 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7575 MDB_val *key, MDB_val *data, unsigned flags)
7580 MDB_val rdata, *xdata;
7584 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7586 mdb_cursor_init(&mc, txn, dbi, &mx);
7595 flags |= MDB_NODUPDATA;
7597 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7599 /* let mdb_page_split know about this cursor if needed:
7600 * delete will trigger a rebalance; if it needs to move
7601 * a node from one page to another, it will have to
7602 * update the parent's separator key(s). If the new sepkey
7603 * is larger than the current one, the parent page may
7604 * run out of space, triggering a split. We need this
7605 * cursor to be consistent until the end of the rebalance.
7607 mc.mc_flags |= C_UNTRACK;
7608 mc.mc_next = txn->mt_cursors[dbi];
7609 txn->mt_cursors[dbi] = &mc;
7610 rc = mdb_cursor_del(&mc, flags);
7611 txn->mt_cursors[dbi] = mc.mc_next;
7616 /** Split a page and insert a new node.
7617 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7618 * The cursor will be updated to point to the actual page and index where
7619 * the node got inserted after the split.
7620 * @param[in] newkey The key for the newly inserted node.
7621 * @param[in] newdata The data for the newly inserted node.
7622 * @param[in] newpgno The page number, if the new node is a branch node.
7623 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7624 * @return 0 on success, non-zero on failure.
7627 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7628 unsigned int nflags)
7631 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7634 int i, j, split_indx, nkeys, pmax;
7635 MDB_env *env = mc->mc_txn->mt_env;
7637 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7638 MDB_page *copy = NULL;
7639 MDB_page *mp, *rp, *pp;
7644 mp = mc->mc_pg[mc->mc_top];
7645 newindx = mc->mc_ki[mc->mc_top];
7646 nkeys = NUMKEYS(mp);
7648 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7649 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7650 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7652 /* Create a right sibling. */
7653 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7655 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7657 if (mc->mc_snum < 2) {
7658 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7660 /* shift current top to make room for new parent */
7661 mc->mc_pg[1] = mc->mc_pg[0];
7662 mc->mc_ki[1] = mc->mc_ki[0];
7665 mc->mc_db->md_root = pp->mp_pgno;
7666 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7667 mc->mc_db->md_depth++;
7670 /* Add left (implicit) pointer. */
7671 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7672 /* undo the pre-push */
7673 mc->mc_pg[0] = mc->mc_pg[1];
7674 mc->mc_ki[0] = mc->mc_ki[1];
7675 mc->mc_db->md_root = mp->mp_pgno;
7676 mc->mc_db->md_depth--;
7683 ptop = mc->mc_top-1;
7684 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7687 mc->mc_flags |= C_SPLITTING;
7688 mdb_cursor_copy(mc, &mn);
7689 mn.mc_pg[mn.mc_top] = rp;
7690 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7692 if (nflags & MDB_APPEND) {
7693 mn.mc_ki[mn.mc_top] = 0;
7695 split_indx = newindx;
7699 split_indx = (nkeys+1) / 2;
7704 unsigned int lsize, rsize, ksize;
7705 /* Move half of the keys to the right sibling */
7706 x = mc->mc_ki[mc->mc_top] - split_indx;
7707 ksize = mc->mc_db->md_pad;
7708 split = LEAF2KEY(mp, split_indx, ksize);
7709 rsize = (nkeys - split_indx) * ksize;
7710 lsize = (nkeys - split_indx) * sizeof(indx_t);
7711 mp->mp_lower -= lsize;
7712 rp->mp_lower += lsize;
7713 mp->mp_upper += rsize - lsize;
7714 rp->mp_upper -= rsize - lsize;
7715 sepkey.mv_size = ksize;
7716 if (newindx == split_indx) {
7717 sepkey.mv_data = newkey->mv_data;
7719 sepkey.mv_data = split;
7722 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7723 memcpy(rp->mp_ptrs, split, rsize);
7724 sepkey.mv_data = rp->mp_ptrs;
7725 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7726 memcpy(ins, newkey->mv_data, ksize);
7727 mp->mp_lower += sizeof(indx_t);
7728 mp->mp_upper -= ksize - sizeof(indx_t);
7731 memcpy(rp->mp_ptrs, split, x * ksize);
7732 ins = LEAF2KEY(rp, x, ksize);
7733 memcpy(ins, newkey->mv_data, ksize);
7734 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7735 rp->mp_lower += sizeof(indx_t);
7736 rp->mp_upper -= ksize - sizeof(indx_t);
7737 mc->mc_ki[mc->mc_top] = x;
7738 mc->mc_pg[mc->mc_top] = rp;
7741 int psize, nsize, k;
7742 /* Maximum free space in an empty page */
7743 pmax = env->me_psize - PAGEHDRSZ;
7745 nsize = mdb_leaf_size(env, newkey, newdata);
7747 nsize = mdb_branch_size(env, newkey);
7748 nsize = EVEN(nsize);
7750 /* grab a page to hold a temporary copy */
7751 copy = mdb_page_malloc(mc->mc_txn, 1);
7756 copy->mp_pgno = mp->mp_pgno;
7757 copy->mp_flags = mp->mp_flags;
7758 copy->mp_lower = PAGEHDRSZ;
7759 copy->mp_upper = env->me_psize;
7761 /* prepare to insert */
7762 for (i=0, j=0; i<nkeys; i++) {
7764 copy->mp_ptrs[j++] = 0;
7766 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7769 /* When items are relatively large the split point needs
7770 * to be checked, because being off-by-one will make the
7771 * difference between success or failure in mdb_node_add.
7773 * It's also relevant if a page happens to be laid out
7774 * such that one half of its nodes are all "small" and
7775 * the other half of its nodes are "large." If the new
7776 * item is also "large" and falls on the half with
7777 * "large" nodes, it also may not fit.
7779 * As a final tweak, if the new item goes on the last
7780 * spot on the page (and thus, onto the new page), bias
7781 * the split so the new page is emptier than the old page.
7782 * This yields better packing during sequential inserts.
7784 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7785 /* Find split point */
7787 if (newindx <= split_indx || newindx >= nkeys) {
7789 k = newindx >= nkeys ? nkeys : split_indx+2;
7794 for (; i!=k; i+=j) {
7799 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7800 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7802 if (F_ISSET(node->mn_flags, F_BIGDATA))
7803 psize += sizeof(pgno_t);
7805 psize += NODEDSZ(node);
7807 psize = EVEN(psize);
7809 if (psize > pmax || i == k-j) {
7810 split_indx = i + (j<0);
7815 if (split_indx == newindx) {
7816 sepkey.mv_size = newkey->mv_size;
7817 sepkey.mv_data = newkey->mv_data;
7819 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7820 sepkey.mv_size = node->mn_ksize;
7821 sepkey.mv_data = NODEKEY(node);
7826 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7828 /* Copy separator key to the parent.
7830 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7834 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7839 if (mn.mc_snum == mc->mc_snum) {
7840 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7841 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7842 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7843 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7848 /* Right page might now have changed parent.
7849 * Check if left page also changed parent.
7851 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7852 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7853 for (i=0; i<ptop; i++) {
7854 mc->mc_pg[i] = mn.mc_pg[i];
7855 mc->mc_ki[i] = mn.mc_ki[i];
7857 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7858 if (mn.mc_ki[ptop]) {
7859 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7861 /* find right page's left sibling */
7862 mc->mc_ki[ptop] = mn.mc_ki[ptop];
7863 mdb_cursor_sibling(mc, 0);
7868 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7871 mc->mc_flags ^= C_SPLITTING;
7872 if (rc != MDB_SUCCESS) {
7875 if (nflags & MDB_APPEND) {
7876 mc->mc_pg[mc->mc_top] = rp;
7877 mc->mc_ki[mc->mc_top] = 0;
7878 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7881 for (i=0; i<mc->mc_top; i++)
7882 mc->mc_ki[i] = mn.mc_ki[i];
7883 } else if (!IS_LEAF2(mp)) {
7885 mc->mc_pg[mc->mc_top] = rp;
7890 rkey.mv_data = newkey->mv_data;
7891 rkey.mv_size = newkey->mv_size;
7897 /* Update index for the new key. */
7898 mc->mc_ki[mc->mc_top] = j;
7900 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7901 rkey.mv_data = NODEKEY(node);
7902 rkey.mv_size = node->mn_ksize;
7904 xdata.mv_data = NODEDATA(node);
7905 xdata.mv_size = NODEDSZ(node);
7908 pgno = NODEPGNO(node);
7909 flags = node->mn_flags;
7912 if (!IS_LEAF(mp) && j == 0) {
7913 /* First branch index doesn't need key data. */
7917 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7923 mc->mc_pg[mc->mc_top] = copy;
7928 } while (i != split_indx);
7930 nkeys = NUMKEYS(copy);
7931 for (i=0; i<nkeys; i++)
7932 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7933 mp->mp_lower = copy->mp_lower;
7934 mp->mp_upper = copy->mp_upper;
7935 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7936 env->me_psize - copy->mp_upper);
7938 /* reset back to original page */
7939 if (newindx < split_indx) {
7940 mc->mc_pg[mc->mc_top] = mp;
7941 if (nflags & MDB_RESERVE) {
7942 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7943 if (!(node->mn_flags & F_BIGDATA))
7944 newdata->mv_data = NODEDATA(node);
7947 mc->mc_pg[mc->mc_top] = rp;
7949 /* Make sure mc_ki is still valid.
7951 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7952 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7953 for (i=0; i<=ptop; i++) {
7954 mc->mc_pg[i] = mn.mc_pg[i];
7955 mc->mc_ki[i] = mn.mc_ki[i];
7962 /* Adjust other cursors pointing to mp */
7963 MDB_cursor *m2, *m3;
7964 MDB_dbi dbi = mc->mc_dbi;
7965 int fixup = NUMKEYS(mp);
7967 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7968 if (mc->mc_flags & C_SUB)
7969 m3 = &m2->mc_xcursor->mx_cursor;
7974 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7976 if (m3->mc_flags & C_SPLITTING)
7981 for (k=m3->mc_top; k>=0; k--) {
7982 m3->mc_ki[k+1] = m3->mc_ki[k];
7983 m3->mc_pg[k+1] = m3->mc_pg[k];
7985 if (m3->mc_ki[0] >= split_indx) {
7990 m3->mc_pg[0] = mc->mc_pg[0];
7994 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7995 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7996 m3->mc_ki[mc->mc_top]++;
7997 if (m3->mc_ki[mc->mc_top] >= fixup) {
7998 m3->mc_pg[mc->mc_top] = rp;
7999 m3->mc_ki[mc->mc_top] -= fixup;
8000 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8002 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8003 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8008 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8011 if (copy) /* tmp page */
8012 mdb_page_free(env, copy);
8014 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8019 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8020 MDB_val *key, MDB_val *data, unsigned int flags)
8025 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8028 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8031 mdb_cursor_init(&mc, txn, dbi, &mx);
8032 return mdb_cursor_put(&mc, key, data, flags);
8036 #define MDB_WBUF (1024*1024)
8039 /** State needed for a compacting copy. */
8040 typedef struct mdb_copy {
8041 pthread_mutex_t mc_mutex;
8042 pthread_cond_t mc_cond;
8049 pgno_t mc_next_pgno;
8052 volatile int mc_new;
8057 /** Dedicated writer thread for compacting copy. */
8059 mdb_env_copythr(void *arg)
8063 int toggle = 0, wsize, rc;
8066 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8069 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8072 pthread_mutex_lock(&my->mc_mutex);
8074 pthread_cond_signal(&my->mc_cond);
8077 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8078 if (my->mc_new < 0) {
8083 wsize = my->mc_wlen[toggle];
8084 ptr = my->mc_wbuf[toggle];
8087 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8091 } else if (len > 0) {
8105 /* If there's an overflow page tail, write it too */
8106 if (my->mc_olen[toggle]) {
8107 wsize = my->mc_olen[toggle];
8108 ptr = my->mc_over[toggle];
8109 my->mc_olen[toggle] = 0;
8112 my->mc_wlen[toggle] = 0;
8114 pthread_cond_signal(&my->mc_cond);
8116 pthread_cond_signal(&my->mc_cond);
8117 pthread_mutex_unlock(&my->mc_mutex);
8118 return (THREAD_RET)0;
8122 /** Tell the writer thread there's a buffer ready to write */
8124 mdb_env_cthr_toggle(mdb_copy *my, int st)
8126 int toggle = my->mc_toggle ^ 1;
8127 pthread_mutex_lock(&my->mc_mutex);
8128 if (my->mc_status) {
8129 pthread_mutex_unlock(&my->mc_mutex);
8130 return my->mc_status;
8132 while (my->mc_new == 1)
8133 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8135 my->mc_toggle = toggle;
8136 pthread_cond_signal(&my->mc_cond);
8137 pthread_mutex_unlock(&my->mc_mutex);
8141 /** Depth-first tree traversal for compacting copy. */
8143 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8146 MDB_txn *txn = my->mc_txn;
8148 MDB_page *mo, *mp, *leaf;
8153 /* Empty DB, nothing to do */
8154 if (*pg == P_INVALID)
8161 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8164 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8168 /* Make cursor pages writable */
8169 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8173 for (i=0; i<mc.mc_top; i++) {
8174 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8175 mc.mc_pg[i] = (MDB_page *)ptr;
8176 ptr += my->mc_env->me_psize;
8179 /* This is writable space for a leaf page. Usually not needed. */
8180 leaf = (MDB_page *)ptr;
8182 toggle = my->mc_toggle;
8183 while (mc.mc_snum > 0) {
8185 mp = mc.mc_pg[mc.mc_top];
8189 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8190 for (i=0; i<n; i++) {
8191 ni = NODEPTR(mp, i);
8192 if (ni->mn_flags & F_BIGDATA) {
8196 /* Need writable leaf */
8198 mc.mc_pg[mc.mc_top] = leaf;
8199 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8201 ni = NODEPTR(mp, i);
8204 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8205 rc = mdb_page_get(txn, pg, &omp, NULL);
8208 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8209 rc = mdb_env_cthr_toggle(my, 1);
8212 toggle = my->mc_toggle;
8214 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8215 memcpy(mo, omp, my->mc_env->me_psize);
8216 mo->mp_pgno = my->mc_next_pgno;
8217 my->mc_next_pgno += omp->mp_pages;
8218 my->mc_wlen[toggle] += my->mc_env->me_psize;
8219 if (omp->mp_pages > 1) {
8220 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8221 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8222 rc = mdb_env_cthr_toggle(my, 1);
8225 toggle = my->mc_toggle;
8227 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8228 } else if (ni->mn_flags & F_SUBDATA) {
8231 /* Need writable leaf */
8233 mc.mc_pg[mc.mc_top] = leaf;
8234 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8236 ni = NODEPTR(mp, i);
8239 memcpy(&db, NODEDATA(ni), sizeof(db));
8240 my->mc_toggle = toggle;
8241 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8244 toggle = my->mc_toggle;
8245 memcpy(NODEDATA(ni), &db, sizeof(db));
8250 mc.mc_ki[mc.mc_top]++;
8251 if (mc.mc_ki[mc.mc_top] < n) {
8254 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8256 rc = mdb_page_get(txn, pg, &mp, NULL);
8261 mc.mc_ki[mc.mc_top] = 0;
8262 if (IS_BRANCH(mp)) {
8263 /* Whenever we advance to a sibling branch page,
8264 * we must proceed all the way down to its first leaf.
8266 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8269 mc.mc_pg[mc.mc_top] = mp;
8273 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8274 rc = mdb_env_cthr_toggle(my, 1);
8277 toggle = my->mc_toggle;
8279 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8280 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8281 mo->mp_pgno = my->mc_next_pgno++;
8282 my->mc_wlen[toggle] += my->mc_env->me_psize;
8284 /* Update parent if there is one */
8285 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8286 SETPGNO(ni, mo->mp_pgno);
8287 mdb_cursor_pop(&mc);
8289 /* Otherwise we're done */
8299 /** Copy environment with compaction. */
8301 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8306 MDB_txn *txn = NULL;
8311 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8312 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8313 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_psize);
8314 if (my.mc_wbuf[0] == NULL)
8317 pthread_mutex_init(&my.mc_mutex, NULL);
8318 pthread_cond_init(&my.mc_cond, NULL);
8319 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_psize, MDB_WBUF*2);
8323 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8328 my.mc_next_pgno = 2;
8335 /* Do the lock/unlock of the reader mutex before starting the
8336 * write txn. Otherwise other read txns could block writers.
8338 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8343 /* We must start the actual read txn after blocking writers */
8344 mdb_txn_reset0(txn, "reset-stage1");
8346 /* Temporarily block writers until we snapshot the meta pages */
8349 rc = mdb_txn_renew0(txn);
8351 UNLOCK_MUTEX_W(env);
8356 THREAD_CREATE(thr, mdb_env_copythr, &my);
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;
8691 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8695 if (env == NULL || arg == NULL)
8698 toggle = mdb_env_pick_meta(env);
8700 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8704 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8708 if (env == NULL || arg == NULL)
8711 toggle = mdb_env_pick_meta(env);
8712 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8713 arg->me_mapsize = env->me_mapsize;
8714 arg->me_maxreaders = env->me_maxreaders;
8716 /* me_numreaders may be zero if this process never used any readers. Use
8717 * the shared numreader count if it exists.
8719 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8721 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8722 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8726 /** Set the default comparison functions for a database.
8727 * Called immediately after a database is opened to set the defaults.
8728 * The user can then override them with #mdb_set_compare() or
8729 * #mdb_set_dupsort().
8730 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8731 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8734 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8736 uint16_t f = txn->mt_dbs[dbi].md_flags;
8738 txn->mt_dbxs[dbi].md_cmp =
8739 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8740 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8742 txn->mt_dbxs[dbi].md_dcmp =
8743 !(f & MDB_DUPSORT) ? 0 :
8744 ((f & MDB_INTEGERDUP)
8745 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8746 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8749 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8754 int rc, dbflag, exact;
8755 unsigned int unused = 0;
8758 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8759 mdb_default_cmp(txn, FREE_DBI);
8762 if ((flags & VALID_FLAGS) != flags)
8764 if (txn->mt_flags & MDB_TXN_ERROR)
8770 if (flags & PERSISTENT_FLAGS) {
8771 uint16_t f2 = flags & PERSISTENT_FLAGS;
8772 /* make sure flag changes get committed */
8773 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8774 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8775 txn->mt_flags |= MDB_TXN_DIRTY;
8778 mdb_default_cmp(txn, MAIN_DBI);
8782 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8783 mdb_default_cmp(txn, MAIN_DBI);
8786 /* Is the DB already open? */
8788 for (i=2; i<txn->mt_numdbs; i++) {
8789 if (!txn->mt_dbxs[i].md_name.mv_size) {
8790 /* Remember this free slot */
8791 if (!unused) unused = i;
8794 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8795 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8801 /* If no free slot and max hit, fail */
8802 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8803 return MDB_DBS_FULL;
8805 /* Cannot mix named databases with some mainDB flags */
8806 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8807 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8809 /* Find the DB info */
8810 dbflag = DB_NEW|DB_VALID;
8813 key.mv_data = (void *)name;
8814 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8815 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8816 if (rc == MDB_SUCCESS) {
8817 /* make sure this is actually a DB */
8818 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8819 if (!(node->mn_flags & F_SUBDATA))
8820 return MDB_INCOMPATIBLE;
8821 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8822 /* Create if requested */
8824 data.mv_size = sizeof(MDB_db);
8825 data.mv_data = &dummy;
8826 memset(&dummy, 0, sizeof(dummy));
8827 dummy.md_root = P_INVALID;
8828 dummy.md_flags = flags & PERSISTENT_FLAGS;
8829 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8833 /* OK, got info, add to table */
8834 if (rc == MDB_SUCCESS) {
8835 unsigned int slot = unused ? unused : txn->mt_numdbs;
8836 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8837 txn->mt_dbxs[slot].md_name.mv_size = len;
8838 txn->mt_dbxs[slot].md_rel = NULL;
8839 txn->mt_dbflags[slot] = dbflag;
8840 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8842 mdb_default_cmp(txn, slot);
8851 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8853 if (!arg || !TXN_DBI_EXIST(txn, dbi))
8856 if (txn->mt_flags & MDB_TXN_ERROR)
8859 if (txn->mt_dbflags[dbi] & DB_STALE) {
8862 /* Stale, must read the DB's root. cursor_init does it for us. */
8863 mdb_cursor_init(&mc, txn, dbi, &mx);
8865 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8868 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8871 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8873 ptr = env->me_dbxs[dbi].md_name.mv_data;
8874 env->me_dbxs[dbi].md_name.mv_data = NULL;
8875 env->me_dbxs[dbi].md_name.mv_size = 0;
8876 env->me_dbflags[dbi] = 0;
8880 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8882 /* We could return the flags for the FREE_DBI too but what's the point? */
8883 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8885 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8889 /** Add all the DB's pages to the free list.
8890 * @param[in] mc Cursor on the DB to free.
8891 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8892 * @return 0 on success, non-zero on failure.
8895 mdb_drop0(MDB_cursor *mc, int subs)
8899 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8900 if (rc == MDB_SUCCESS) {
8901 MDB_txn *txn = mc->mc_txn;
8906 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8907 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8910 mdb_cursor_copy(mc, &mx);
8911 while (mc->mc_snum > 0) {
8912 MDB_page *mp = mc->mc_pg[mc->mc_top];
8913 unsigned n = NUMKEYS(mp);
8915 for (i=0; i<n; i++) {
8916 ni = NODEPTR(mp, i);
8917 if (ni->mn_flags & F_BIGDATA) {
8920 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8921 rc = mdb_page_get(txn, pg, &omp, NULL);
8924 mdb_cassert(mc, IS_OVERFLOW(omp));
8925 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8929 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8930 mdb_xcursor_init1(mc, ni);
8931 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8937 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8939 for (i=0; i<n; i++) {
8941 ni = NODEPTR(mp, i);
8944 mdb_midl_xappend(txn->mt_free_pgs, pg);
8949 mc->mc_ki[mc->mc_top] = i;
8950 rc = mdb_cursor_sibling(mc, 1);
8952 if (rc != MDB_NOTFOUND)
8954 /* no more siblings, go back to beginning
8955 * of previous level.
8959 for (i=1; i<mc->mc_snum; i++) {
8961 mc->mc_pg[i] = mx.mc_pg[i];
8966 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8969 txn->mt_flags |= MDB_TXN_ERROR;
8970 } else if (rc == MDB_NOTFOUND) {
8976 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8978 MDB_cursor *mc, *m2;
8981 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8984 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8987 rc = mdb_cursor_open(txn, dbi, &mc);
8991 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8992 /* Invalidate the dropped DB's cursors */
8993 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8994 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8998 /* Can't delete the main DB */
8999 if (del && dbi > MAIN_DBI) {
9000 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9002 txn->mt_dbflags[dbi] = DB_STALE;
9003 mdb_dbi_close(txn->mt_env, dbi);
9005 txn->mt_flags |= MDB_TXN_ERROR;
9008 /* reset the DB record, mark it dirty */
9009 txn->mt_dbflags[dbi] |= DB_DIRTY;
9010 txn->mt_dbs[dbi].md_depth = 0;
9011 txn->mt_dbs[dbi].md_branch_pages = 0;
9012 txn->mt_dbs[dbi].md_leaf_pages = 0;
9013 txn->mt_dbs[dbi].md_overflow_pages = 0;
9014 txn->mt_dbs[dbi].md_entries = 0;
9015 txn->mt_dbs[dbi].md_root = P_INVALID;
9017 txn->mt_flags |= MDB_TXN_DIRTY;
9020 mdb_cursor_close(mc);
9024 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9026 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9029 txn->mt_dbxs[dbi].md_cmp = cmp;
9033 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9035 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9038 txn->mt_dbxs[dbi].md_dcmp = cmp;
9042 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9044 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9047 txn->mt_dbxs[dbi].md_rel = rel;
9051 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9053 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9056 txn->mt_dbxs[dbi].md_relctx = ctx;
9060 int mdb_env_get_maxkeysize(MDB_env *env)
9062 return ENV_MAXKEY(env);
9065 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9067 unsigned int i, rdrs;
9070 int rc = 0, first = 1;
9074 if (!env->me_txns) {
9075 return func("(no reader locks)\n", ctx);
9077 rdrs = env->me_txns->mti_numreaders;
9078 mr = env->me_txns->mti_readers;
9079 for (i=0; i<rdrs; i++) {
9081 txnid_t txnid = mr[i].mr_txnid;
9082 sprintf(buf, txnid == (txnid_t)-1 ?
9083 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9084 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9087 rc = func(" pid thread txnid\n", ctx);
9091 rc = func(buf, ctx);
9097 rc = func("(no active readers)\n", ctx);
9102 /** Insert pid into list if not already present.
9103 * return -1 if already present.
9105 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9107 /* binary search of pid in list */
9109 unsigned cursor = 1;
9111 unsigned n = ids[0];
9114 unsigned pivot = n >> 1;
9115 cursor = base + pivot + 1;
9116 val = pid - ids[cursor];
9121 } else if ( val > 0 ) {
9126 /* found, so it's a duplicate */
9135 for (n = ids[0]; n > cursor; n--)
9141 int mdb_reader_check(MDB_env *env, int *dead)
9143 unsigned int i, j, rdrs;
9145 MDB_PID_T *pids, pid;
9154 rdrs = env->me_txns->mti_numreaders;
9155 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9159 mr = env->me_txns->mti_readers;
9160 for (i=0; i<rdrs; i++) {
9161 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9163 if (mdb_pid_insert(pids, pid) == 0) {
9164 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9166 /* Recheck, a new process may have reused pid */
9167 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9168 for (j=i; j<rdrs; j++)
9169 if (mr[j].mr_pid == pid) {
9170 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9171 (unsigned) pid, mr[j].mr_txnid));
9176 UNLOCK_MUTEX_R(env);
projects / openldap / blob