2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 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.
38 #include <sys/types.h>
42 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
43 * as int64 which is wrong. MSVC doesn't define it at all, so just
48 # include <sys/param.h>
50 # define LITTLE_ENDIAN 1234
51 # define BIG_ENDIAN 4321
52 # define BYTE_ORDER LITTLE_ENDIAN
54 # define SSIZE_MAX INT_MAX
58 #define MDB_PID_T pid_t
59 #include <sys/param.h>
62 #ifdef HAVE_SYS_FILE_H
79 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
80 #include <netinet/in.h>
81 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
84 #if defined(__APPLE__) || defined (BSD)
85 # define MDB_USE_POSIX_SEM 1
86 # define MDB_FDATASYNC fsync
87 #elif defined(ANDROID)
88 # define MDB_FDATASYNC fsync
93 #ifdef MDB_USE_POSIX_SEM
94 # define MDB_USE_HASH 1
95 #include <semaphore.h>
100 #include <valgrind/memcheck.h>
101 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
102 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
103 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
104 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
105 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
107 #define VGMEMP_CREATE(h,r,z)
108 #define VGMEMP_ALLOC(h,a,s)
109 #define VGMEMP_FREE(h,a)
110 #define VGMEMP_DESTROY(h)
111 #define VGMEMP_DEFINED(a,s)
115 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
116 /* Solaris just defines one or the other */
117 # define LITTLE_ENDIAN 1234
118 # define BIG_ENDIAN 4321
119 # ifdef _LITTLE_ENDIAN
120 # define BYTE_ORDER LITTLE_ENDIAN
122 # define BYTE_ORDER BIG_ENDIAN
125 # define BYTE_ORDER __BYTE_ORDER
129 #ifndef LITTLE_ENDIAN
130 #define LITTLE_ENDIAN __LITTLE_ENDIAN
133 #define BIG_ENDIAN __BIG_ENDIAN
136 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
137 #define MISALIGNED_OK 1
143 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
144 # error "Unknown or unsupported endianness (BYTE_ORDER)"
145 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
146 # error "Two's complement, reasonably sized integer types, please"
149 /** @defgroup internal MDB Internals
152 /** @defgroup compat Windows Compatibility Macros
153 * A bunch of macros to minimize the amount of platform-specific ifdefs
154 * needed throughout the rest of the code. When the features this library
155 * needs are similar enough to POSIX to be hidden in a one-or-two line
156 * replacement, this macro approach is used.
160 #define MDB_USE_HASH 1
161 #define MDB_PIDLOCK 0
162 #define pthread_t DWORD
163 #define pthread_mutex_t HANDLE
164 #define pthread_key_t DWORD
165 #define pthread_self() GetCurrentThreadId()
166 #define pthread_key_create(x,y) \
167 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
168 #define pthread_key_delete(x) TlsFree(x)
169 #define pthread_getspecific(x) TlsGetValue(x)
170 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
171 #define pthread_mutex_unlock(x) ReleaseMutex(x)
172 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
173 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
174 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
175 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
176 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
177 #define getpid() GetCurrentProcessId()
178 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
179 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
180 #define ErrCode() GetLastError()
181 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
182 #define close(fd) (CloseHandle(fd) ? 0 : -1)
183 #define munmap(ptr,len) UnmapViewOfFile(ptr)
184 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
185 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
187 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
192 #define Z "z" /**< printf format modifier for size_t */
194 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
195 #define MDB_PIDLOCK 1
197 #ifdef MDB_USE_POSIX_SEM
199 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
200 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
201 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
202 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
205 mdb_sem_wait(sem_t *sem)
208 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
213 /** Lock the reader mutex.
215 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
216 /** Unlock the reader mutex.
218 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
220 /** Lock the writer mutex.
221 * Only a single write transaction is allowed at a time. Other writers
222 * will block waiting for this mutex.
224 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
225 /** Unlock the writer mutex.
227 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
228 #endif /* MDB_USE_POSIX_SEM */
230 /** Get the error code for the last failed system function.
232 #define ErrCode() errno
234 /** An abstraction for a file handle.
235 * On POSIX systems file handles are small integers. On Windows
236 * they're opaque pointers.
240 /** A value for an invalid file handle.
241 * Mainly used to initialize file variables and signify that they are
244 #define INVALID_HANDLE_VALUE (-1)
246 /** Get the size of a memory page for the system.
247 * This is the basic size that the platform's memory manager uses, and is
248 * fundamental to the use of memory-mapped files.
250 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
253 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
256 #define MNAME_LEN (sizeof(pthread_mutex_t))
262 /** A flag for opening a file and requesting synchronous data writes.
263 * This is only used when writing a meta page. It's not strictly needed;
264 * we could just do a normal write and then immediately perform a flush.
265 * But if this flag is available it saves us an extra system call.
267 * @note If O_DSYNC is undefined but exists in /usr/include,
268 * preferably set some compiler flag to get the definition.
269 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
272 # define MDB_DSYNC O_DSYNC
276 /** Function for flushing the data of a file. Define this to fsync
277 * if fdatasync() is not supported.
279 #ifndef MDB_FDATASYNC
280 # define MDB_FDATASYNC fdatasync
284 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
295 /** A page number in the database.
296 * Note that 64 bit page numbers are overkill, since pages themselves
297 * already represent 12-13 bits of addressable memory, and the OS will
298 * always limit applications to a maximum of 63 bits of address space.
300 * @note In the #MDB_node structure, we only store 48 bits of this value,
301 * which thus limits us to only 60 bits of addressable data.
303 typedef MDB_ID pgno_t;
305 /** A transaction ID.
306 * See struct MDB_txn.mt_txnid for details.
308 typedef MDB_ID txnid_t;
310 /** @defgroup debug Debug Macros
314 /** Enable debug output. Needs variable argument macros (a C99 feature).
315 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
316 * read from and written to the database (used for free space management).
322 static int mdb_debug;
323 static txnid_t mdb_debug_start;
325 /** Print a debug message with printf formatting.
326 * Requires double parenthesis around 2 or more args.
328 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
329 # define DPRINTF0(fmt, ...) \
330 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
332 # define DPRINTF(args) ((void) 0)
334 /** Print a debug string.
335 * The string is printed literally, with no format processing.
337 #define DPUTS(arg) DPRINTF(("%s", arg))
338 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
340 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
343 /** @brief The maximum size of a database page.
345 * This is 32k, since it must fit in #MDB_page.#mp_upper.
347 * LMDB will use database pages < OS pages if needed.
348 * That causes more I/O in write transactions: The OS must
349 * know (read) the whole page before writing a partial page.
351 * Note that we don't currently support Huge pages. On Linux,
352 * regular data files cannot use Huge pages, and in general
353 * Huge pages aren't actually pageable. We rely on the OS
354 * demand-pager to read our data and page it out when memory
355 * pressure from other processes is high. So until OSs have
356 * actual paging support for Huge pages, they're not viable.
358 #define MAX_PAGESIZE 0x8000
360 /** The minimum number of keys required in a database page.
361 * Setting this to a larger value will place a smaller bound on the
362 * maximum size of a data item. Data items larger than this size will
363 * be pushed into overflow pages instead of being stored directly in
364 * the B-tree node. This value used to default to 4. With a page size
365 * of 4096 bytes that meant that any item larger than 1024 bytes would
366 * go into an overflow page. That also meant that on average 2-3KB of
367 * each overflow page was wasted space. The value cannot be lower than
368 * 2 because then there would no longer be a tree structure. With this
369 * value, items larger than 2KB will go into overflow pages, and on
370 * average only 1KB will be wasted.
372 #define MDB_MINKEYS 2
374 /** A stamp that identifies a file as an MDB file.
375 * There's nothing special about this value other than that it is easily
376 * recognizable, and it will reflect any byte order mismatches.
378 #define MDB_MAGIC 0xBEEFC0DE
380 /** The version number for a database's datafile format. */
381 #define MDB_DATA_VERSION 1
382 /** The version number for a database's lockfile format. */
383 #define MDB_LOCK_VERSION 1
385 /** @brief The max size of a key we can write, or 0 for dynamic max.
387 * Define this as 0 to compute the max from the page size. 511
388 * is default for backwards compat: liblmdb <= 0.9.10 can break
389 * when modifying a DB with keys/dupsort data bigger than its max.
391 * Data items in an #MDB_DUPSORT database are also limited to
392 * this size, since they're actually keys of a sub-DB. Keys and
393 * #MDB_DUPSORT data items must fit on a node in a regular page.
395 #ifndef MDB_MAXKEYSIZE
396 #define MDB_MAXKEYSIZE 511
399 /** The maximum size of a key we can write to the environment. */
401 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
403 #define ENV_MAXKEY(env) ((env)->me_maxkey)
406 /** @brief The maximum size of a data item.
408 * We only store a 32 bit value for node sizes.
410 #define MAXDATASIZE 0xffffffffUL
413 /** Key size which fits in a #DKBUF.
416 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
419 * This is used for printing a hex dump of a key's contents.
421 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
422 /** Display a key in hex.
424 * Invoke a function to display a key in hex.
426 #define DKEY(x) mdb_dkey(x, kbuf)
432 /** An invalid page number.
433 * Mainly used to denote an empty tree.
435 #define P_INVALID (~(pgno_t)0)
437 /** Test if the flags \b f are set in a flag word \b w. */
438 #define F_ISSET(w, f) (((w) & (f)) == (f))
440 /** Round \b n up to an even number. */
441 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
443 /** Used for offsets within a single page.
444 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
447 typedef uint16_t indx_t;
449 /** Default size of memory map.
450 * This is certainly too small for any actual applications. Apps should always set
451 * the size explicitly using #mdb_env_set_mapsize().
453 #define DEFAULT_MAPSIZE 1048576
455 /** @defgroup readers Reader Lock Table
456 * Readers don't acquire any locks for their data access. Instead, they
457 * simply record their transaction ID in the reader table. The reader
458 * mutex is needed just to find an empty slot in the reader table. The
459 * slot's address is saved in thread-specific data so that subsequent read
460 * transactions started by the same thread need no further locking to proceed.
462 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
464 * No reader table is used if the database is on a read-only filesystem, or
465 * if #MDB_NOLOCK is set.
467 * Since the database uses multi-version concurrency control, readers don't
468 * actually need any locking. This table is used to keep track of which
469 * readers are using data from which old transactions, so that we'll know
470 * when a particular old transaction is no longer in use. Old transactions
471 * that have discarded any data pages can then have those pages reclaimed
472 * for use by a later write transaction.
474 * The lock table is constructed such that reader slots are aligned with the
475 * processor's cache line size. Any slot is only ever used by one thread.
476 * This alignment guarantees that there will be no contention or cache
477 * thrashing as threads update their own slot info, and also eliminates
478 * any need for locking when accessing a slot.
480 * A writer thread will scan every slot in the table to determine the oldest
481 * outstanding reader transaction. Any freed pages older than this will be
482 * reclaimed by the writer. The writer doesn't use any locks when scanning
483 * this table. This means that there's no guarantee that the writer will
484 * see the most up-to-date reader info, but that's not required for correct
485 * operation - all we need is to know the upper bound on the oldest reader,
486 * we don't care at all about the newest reader. So the only consequence of
487 * reading stale information here is that old pages might hang around a
488 * while longer before being reclaimed. That's actually good anyway, because
489 * the longer we delay reclaiming old pages, the more likely it is that a
490 * string of contiguous pages can be found after coalescing old pages from
491 * many old transactions together.
494 /** Number of slots in the reader table.
495 * This value was chosen somewhat arbitrarily. 126 readers plus a
496 * couple mutexes fit exactly into 8KB on my development machine.
497 * Applications should set the table size using #mdb_env_set_maxreaders().
499 #define DEFAULT_READERS 126
501 /** The size of a CPU cache line in bytes. We want our lock structures
502 * aligned to this size to avoid false cache line sharing in the
504 * This value works for most CPUs. For Itanium this should be 128.
510 /** The information we store in a single slot of the reader table.
511 * In addition to a transaction ID, we also record the process and
512 * thread ID that owns a slot, so that we can detect stale information,
513 * e.g. threads or processes that went away without cleaning up.
514 * @note We currently don't check for stale records. We simply re-init
515 * the table when we know that we're the only process opening the
518 typedef struct MDB_rxbody {
519 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
520 * Multiple readers that start at the same time will probably have the
521 * same ID here. Again, it's not important to exclude them from
522 * anything; all we need to know is which version of the DB they
523 * started from so we can avoid overwriting any data used in that
524 * particular version.
527 /** The process ID of the process owning this reader txn. */
529 /** The thread ID of the thread owning this txn. */
533 /** The actual reader record, with cacheline padding. */
534 typedef struct MDB_reader {
537 /** shorthand for mrb_txnid */
538 #define mr_txnid mru.mrx.mrb_txnid
539 #define mr_pid mru.mrx.mrb_pid
540 #define mr_tid mru.mrx.mrb_tid
541 /** cache line alignment */
542 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
546 /** The header for the reader table.
547 * The table resides in a memory-mapped file. (This is a different file
548 * than is used for the main database.)
550 * For POSIX the actual mutexes reside in the shared memory of this
551 * mapped file. On Windows, mutexes are named objects allocated by the
552 * kernel; we store the mutex names in this mapped file so that other
553 * processes can grab them. This same approach is also used on
554 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
555 * process-shared POSIX mutexes. For these cases where a named object
556 * is used, the object name is derived from a 64 bit FNV hash of the
557 * environment pathname. As such, naming collisions are extremely
558 * unlikely. If a collision occurs, the results are unpredictable.
560 typedef struct MDB_txbody {
561 /** Stamp identifying this as an MDB file. It must be set
564 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
566 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
567 char mtb_rmname[MNAME_LEN];
569 /** Mutex protecting access to this table.
570 * This is the reader lock that #LOCK_MUTEX_R acquires.
572 pthread_mutex_t mtb_mutex;
574 /** The ID of the last transaction committed to the database.
575 * This is recorded here only for convenience; the value can always
576 * be determined by reading the main database meta pages.
579 /** The number of slots that have been used in the reader table.
580 * This always records the maximum count, it is not decremented
581 * when readers release their slots.
583 unsigned mtb_numreaders;
586 /** The actual reader table definition. */
587 typedef struct MDB_txninfo {
590 #define mti_magic mt1.mtb.mtb_magic
591 #define mti_format mt1.mtb.mtb_format
592 #define mti_mutex mt1.mtb.mtb_mutex
593 #define mti_rmname mt1.mtb.mtb_rmname
594 #define mti_txnid mt1.mtb.mtb_txnid
595 #define mti_numreaders mt1.mtb.mtb_numreaders
596 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
599 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
600 char mt2_wmname[MNAME_LEN];
601 #define mti_wmname mt2.mt2_wmname
603 pthread_mutex_t mt2_wmutex;
604 #define mti_wmutex mt2.mt2_wmutex
606 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
608 MDB_reader mti_readers[1];
611 /** Lockfile format signature: version, features and field layout */
612 #define MDB_LOCK_FORMAT \
614 ((MDB_LOCK_VERSION) \
615 /* Flags which describe functionality */ \
616 + (((MDB_PIDLOCK) != 0) << 16)))
619 /** Common header for all page types.
620 * Overflow records occupy a number of contiguous pages with no
621 * headers on any page after the first.
623 typedef struct MDB_page {
624 #define mp_pgno mp_p.p_pgno
625 #define mp_next mp_p.p_next
627 pgno_t p_pgno; /**< page number */
628 void * p_next; /**< for in-memory list of freed structs */
631 /** @defgroup mdb_page Page Flags
633 * Flags for the page headers.
636 #define P_BRANCH 0x01 /**< branch page */
637 #define P_LEAF 0x02 /**< leaf page */
638 #define P_OVERFLOW 0x04 /**< overflow page */
639 #define P_META 0x08 /**< meta page */
640 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
641 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
642 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
643 #define P_KEEP 0x8000 /**< leave this page alone during spill */
645 uint16_t mp_flags; /**< @ref mdb_page */
646 #define mp_lower mp_pb.pb.pb_lower
647 #define mp_upper mp_pb.pb.pb_upper
648 #define mp_pages mp_pb.pb_pages
651 indx_t pb_lower; /**< lower bound of free space */
652 indx_t pb_upper; /**< upper bound of free space */
654 uint32_t pb_pages; /**< number of overflow pages */
656 indx_t mp_ptrs[1]; /**< dynamic size */
659 /** Size of the page header, excluding dynamic data at the end */
660 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
662 /** Address of first usable data byte in a page, after the header */
663 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
665 /** Number of nodes on a page */
666 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
668 /** The amount of space remaining in the page */
669 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
671 /** The percentage of space used in the page, in tenths of a percent. */
672 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
673 ((env)->me_psize - PAGEHDRSZ))
674 /** The minimum page fill factor, in tenths of a percent.
675 * Pages emptier than this are candidates for merging.
677 #define FILL_THRESHOLD 250
679 /** Test if a page is a leaf page */
680 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
681 /** Test if a page is a LEAF2 page */
682 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
683 /** Test if a page is a branch page */
684 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
685 /** Test if a page is an overflow page */
686 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
687 /** Test if a page is a sub page */
688 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
690 /** The number of overflow pages needed to store the given size. */
691 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
693 /** Header for a single key/data pair within a page.
694 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
695 * We guarantee 2-byte alignment for 'MDB_node's.
697 typedef struct MDB_node {
698 /** lo and hi are used for data size on leaf nodes and for
699 * child pgno on branch nodes. On 64 bit platforms, flags
700 * is also used for pgno. (Branch nodes have no flags).
701 * They are in host byte order in case that lets some
702 * accesses be optimized into a 32-bit word access.
704 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
705 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
706 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
707 /** @defgroup mdb_node Node Flags
709 * Flags for node headers.
712 #define F_BIGDATA 0x01 /**< data put on overflow page */
713 #define F_SUBDATA 0x02 /**< data is a sub-database */
714 #define F_DUPDATA 0x04 /**< data has duplicates */
716 /** valid flags for #mdb_node_add() */
717 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
720 unsigned short mn_flags; /**< @ref mdb_node */
721 unsigned short mn_ksize; /**< key size */
722 char mn_data[1]; /**< key and data are appended here */
725 /** Size of the node header, excluding dynamic data at the end */
726 #define NODESIZE offsetof(MDB_node, mn_data)
728 /** Bit position of top word in page number, for shifting mn_flags */
729 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
731 /** Size of a node in a branch page with a given key.
732 * This is just the node header plus the key, there is no data.
734 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
736 /** Size of a node in a leaf page with a given key and data.
737 * This is node header plus key plus data size.
739 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
741 /** Address of node \b i in page \b p */
742 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
744 /** Address of the key for the node */
745 #define NODEKEY(node) (void *)((node)->mn_data)
747 /** Address of the data for a node */
748 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
750 /** Get the page number pointed to by a branch node */
751 #define NODEPGNO(node) \
752 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
753 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
754 /** Set the page number in a branch node */
755 #define SETPGNO(node,pgno) do { \
756 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
757 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
759 /** Get the size of the data in a leaf node */
760 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
761 /** Set the size of the data for a leaf node */
762 #define SETDSZ(node,size) do { \
763 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
764 /** The size of a key in a node */
765 #define NODEKSZ(node) ((node)->mn_ksize)
767 /** Copy a page number from src to dst */
769 #define COPY_PGNO(dst,src) dst = src
771 #if SIZE_MAX > 4294967295UL
772 #define COPY_PGNO(dst,src) do { \
773 unsigned short *s, *d; \
774 s = (unsigned short *)&(src); \
775 d = (unsigned short *)&(dst); \
782 #define COPY_PGNO(dst,src) do { \
783 unsigned short *s, *d; \
784 s = (unsigned short *)&(src); \
785 d = (unsigned short *)&(dst); \
791 /** The address of a key in a LEAF2 page.
792 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
793 * There are no node headers, keys are stored contiguously.
795 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
797 /** Set the \b node's key into \b keyptr, if requested. */
798 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
799 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
801 /** Set the \b node's key into \b key. */
802 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
804 /** Information about a single database in the environment. */
805 typedef struct MDB_db {
806 uint32_t md_pad; /**< also ksize for LEAF2 pages */
807 uint16_t md_flags; /**< @ref mdb_dbi_open */
808 uint16_t md_depth; /**< depth of this tree */
809 pgno_t md_branch_pages; /**< number of internal pages */
810 pgno_t md_leaf_pages; /**< number of leaf pages */
811 pgno_t md_overflow_pages; /**< number of overflow pages */
812 size_t md_entries; /**< number of data items */
813 pgno_t md_root; /**< the root page of this tree */
816 /** mdb_dbi_open flags */
817 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
818 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
819 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
820 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
822 /** Handle for the DB used to track free pages. */
824 /** Handle for the default DB. */
827 /** Meta page content.
828 * A meta page is the start point for accessing a database snapshot.
829 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
831 typedef struct MDB_meta {
832 /** Stamp identifying this as an MDB file. It must be set
835 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
837 void *mm_address; /**< address for fixed mapping */
838 size_t mm_mapsize; /**< size of mmap region */
839 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
840 /** The size of pages used in this DB */
841 #define mm_psize mm_dbs[0].md_pad
842 /** Any persistent environment flags. @ref mdb_env */
843 #define mm_flags mm_dbs[0].md_flags
844 pgno_t mm_last_pg; /**< last used page in file */
845 txnid_t mm_txnid; /**< txnid that committed this page */
848 /** Buffer for a stack-allocated meta page.
849 * The members define size and alignment, and silence type
850 * aliasing warnings. They are not used directly; that could
851 * mean incorrectly using several union members in parallel.
853 typedef union MDB_metabuf {
856 char mm_pad[PAGEHDRSZ];
861 /** Auxiliary DB info.
862 * The information here is mostly static/read-only. There is
863 * only a single copy of this record in the environment.
865 typedef struct MDB_dbx {
866 MDB_val md_name; /**< name of the database */
867 MDB_cmp_func *md_cmp; /**< function for comparing keys */
868 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
869 MDB_rel_func *md_rel; /**< user relocate function */
870 void *md_relctx; /**< user-provided context for md_rel */
873 /** A database transaction.
874 * Every operation requires a transaction handle.
877 MDB_txn *mt_parent; /**< parent of a nested txn */
878 MDB_txn *mt_child; /**< nested txn under this txn */
879 pgno_t mt_next_pgno; /**< next unallocated page */
880 /** The ID of this transaction. IDs are integers incrementing from 1.
881 * Only committed write transactions increment the ID. If a transaction
882 * aborts, the ID may be re-used by the next writer.
885 MDB_env *mt_env; /**< the DB environment */
886 /** The list of pages that became unused during this transaction.
889 /** The sorted list of dirty pages we temporarily wrote to disk
890 * because the dirty list was full. page numbers in here are
891 * shifted left by 1, deleted slots have the LSB set.
893 MDB_IDL mt_spill_pgs;
895 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
897 /** For read txns: This thread/txn's reader table slot, or NULL. */
900 /** Array of records for each DB known in the environment. */
902 /** Array of MDB_db records for each known DB */
904 /** @defgroup mt_dbflag Transaction DB Flags
908 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
909 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
910 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
911 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
913 /** In write txns, array of cursors for each DB */
914 MDB_cursor **mt_cursors;
915 /** Array of flags for each DB */
916 unsigned char *mt_dbflags;
917 /** Number of DB records in use. This number only ever increments;
918 * we don't decrement it when individual DB handles are closed.
922 /** @defgroup mdb_txn Transaction Flags
926 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
927 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
928 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
929 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
931 unsigned int mt_flags; /**< @ref mdb_txn */
932 /** dirty_list room: Array size - #dirty pages visible to this txn.
933 * Includes ancestor txns' dirty pages not hidden by other txns'
934 * dirty/spilled pages. Thus commit(nested txn) has room to merge
935 * dirty_list into mt_parent after freeing hidden mt_parent pages.
937 unsigned int mt_dirty_room;
940 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
941 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
942 * raise this on a 64 bit machine.
944 #define CURSOR_STACK 32
948 /** Cursors are used for all DB operations.
949 * A cursor holds a path of (page pointer, key index) from the DB
950 * root to a position in the DB, plus other state. #MDB_DUPSORT
951 * cursors include an xcursor to the current data item. Write txns
952 * track their cursors and keep them up to date when data moves.
953 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
954 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
957 /** Next cursor on this DB in this txn */
959 /** Backup of the original cursor if this cursor is a shadow */
960 MDB_cursor *mc_backup;
961 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
962 struct MDB_xcursor *mc_xcursor;
963 /** The transaction that owns this cursor */
965 /** The database handle this cursor operates on */
967 /** The database record for this cursor */
969 /** The database auxiliary record for this cursor */
971 /** The @ref mt_dbflag for this database */
972 unsigned char *mc_dbflag;
973 unsigned short mc_snum; /**< number of pushed pages */
974 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
975 /** @defgroup mdb_cursor Cursor Flags
977 * Cursor state flags.
980 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
981 #define C_EOF 0x02 /**< No more data */
982 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
983 #define C_DEL 0x08 /**< last op was a cursor_del */
984 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
985 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
987 unsigned int mc_flags; /**< @ref mdb_cursor */
988 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
989 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
992 /** Context for sorted-dup records.
993 * We could have gone to a fully recursive design, with arbitrarily
994 * deep nesting of sub-databases. But for now we only handle these
995 * levels - main DB, optional sub-DB, sorted-duplicate DB.
997 typedef struct MDB_xcursor {
998 /** A sub-cursor for traversing the Dup DB */
999 MDB_cursor mx_cursor;
1000 /** The database record for this Dup DB */
1002 /** The auxiliary DB record for this Dup DB */
1004 /** The @ref mt_dbflag for this Dup DB */
1005 unsigned char mx_dbflag;
1008 /** State of FreeDB old pages, stored in the MDB_env */
1009 typedef struct MDB_pgstate {
1010 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1011 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1014 /** The database environment. */
1016 HANDLE me_fd; /**< The main data file */
1017 HANDLE me_lfd; /**< The lock file */
1018 HANDLE me_mfd; /**< just for writing the meta pages */
1019 /** Failed to update the meta page. Probably an I/O error. */
1020 #define MDB_FATAL_ERROR 0x80000000U
1021 /** Some fields are initialized. */
1022 #define MDB_ENV_ACTIVE 0x20000000U
1023 /** me_txkey is set */
1024 #define MDB_ENV_TXKEY 0x10000000U
1025 /** Have liveness lock in reader table */
1026 #define MDB_LIVE_READER 0x08000000U
1027 uint32_t me_flags; /**< @ref mdb_env */
1028 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1029 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1030 unsigned int me_maxreaders; /**< size of the reader table */
1031 unsigned int me_numreaders; /**< max numreaders set by this env */
1032 MDB_dbi me_numdbs; /**< number of DBs opened */
1033 MDB_dbi me_maxdbs; /**< size of the DB table */
1034 MDB_PID_T me_pid; /**< process ID of this env */
1035 char *me_path; /**< path to the DB files */
1036 char *me_map; /**< the memory map of the data file */
1037 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1038 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1039 void *me_pbuf; /**< scratch area for DUPSORT put() */
1040 MDB_txn *me_txn; /**< current write transaction */
1041 size_t me_mapsize; /**< size of the data memory map */
1042 off_t me_size; /**< current file size */
1043 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1044 MDB_dbx *me_dbxs; /**< array of static DB info */
1045 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1046 pthread_key_t me_txkey; /**< thread-key for readers */
1047 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1048 # define me_pglast me_pgstate.mf_pglast
1049 # define me_pghead me_pgstate.mf_pghead
1050 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1051 /** IDL of pages that became unused in a write txn */
1052 MDB_IDL me_free_pgs;
1053 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1054 MDB_ID2L me_dirty_list;
1055 /** Max number of freelist items that can fit in a single overflow page */
1057 /** Max size of a node on a page */
1058 unsigned int me_nodemax;
1059 #if !(MDB_MAXKEYSIZE)
1060 unsigned int me_maxkey; /**< max size of a key */
1063 int me_pidquery; /**< Used in OpenProcess */
1064 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1066 #elif defined(MDB_USE_POSIX_SEM)
1067 sem_t *me_rmutex; /* Shared mutexes are not supported */
1072 /** Nested transaction */
1073 typedef struct MDB_ntxn {
1074 MDB_txn mnt_txn; /**< the transaction */
1075 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1078 /** max number of pages to commit in one writev() call */
1079 #define MDB_COMMIT_PAGES 64
1080 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1081 #undef MDB_COMMIT_PAGES
1082 #define MDB_COMMIT_PAGES IOV_MAX
1085 /* max bytes to write in one call */
1086 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1088 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1089 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1090 static int mdb_page_touch(MDB_cursor *mc);
1092 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1093 static int mdb_page_search_root(MDB_cursor *mc,
1094 MDB_val *key, int modify);
1095 #define MDB_PS_MODIFY 1
1096 #define MDB_PS_ROOTONLY 2
1097 #define MDB_PS_FIRST 4
1098 #define MDB_PS_LAST 8
1099 static int mdb_page_search(MDB_cursor *mc,
1100 MDB_val *key, int flags);
1101 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1103 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1104 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1105 pgno_t newpgno, unsigned int nflags);
1107 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1108 static int mdb_env_pick_meta(const MDB_env *env);
1109 static int mdb_env_write_meta(MDB_txn *txn);
1110 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1111 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1113 static void mdb_env_close0(MDB_env *env, int excl);
1115 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1116 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1117 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1118 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1119 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1120 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1121 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1122 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1123 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1125 static int mdb_rebalance(MDB_cursor *mc);
1126 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1128 static void mdb_cursor_pop(MDB_cursor *mc);
1129 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1131 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1132 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1133 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1134 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1135 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1137 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1138 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1140 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1141 static void mdb_xcursor_init0(MDB_cursor *mc);
1142 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1144 static int mdb_drop0(MDB_cursor *mc, int subs);
1145 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1148 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1152 static SECURITY_DESCRIPTOR mdb_null_sd;
1153 static SECURITY_ATTRIBUTES mdb_all_sa;
1154 static int mdb_sec_inited;
1157 /** Return the library version info. */
1159 mdb_version(int *major, int *minor, int *patch)
1161 if (major) *major = MDB_VERSION_MAJOR;
1162 if (minor) *minor = MDB_VERSION_MINOR;
1163 if (patch) *patch = MDB_VERSION_PATCH;
1164 return MDB_VERSION_STRING;
1167 /** Table of descriptions for MDB @ref errors */
1168 static char *const mdb_errstr[] = {
1169 "MDB_KEYEXIST: Key/data pair already exists",
1170 "MDB_NOTFOUND: No matching key/data pair found",
1171 "MDB_PAGE_NOTFOUND: Requested page not found",
1172 "MDB_CORRUPTED: Located page was wrong type",
1173 "MDB_PANIC: Update of meta page failed",
1174 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1175 "MDB_INVALID: File is not an MDB file",
1176 "MDB_MAP_FULL: Environment mapsize limit reached",
1177 "MDB_DBS_FULL: Environment maxdbs limit reached",
1178 "MDB_READERS_FULL: Environment maxreaders limit reached",
1179 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1180 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1181 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1182 "MDB_PAGE_FULL: Internal error - page has no more space",
1183 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1184 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1185 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1186 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1187 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1191 mdb_strerror(int err)
1195 return ("Successful return: 0");
1197 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1198 i = err - MDB_KEYEXIST;
1199 return mdb_errstr[i];
1202 return strerror(err);
1206 /** Display a key in hexadecimal and return the address of the result.
1207 * @param[in] key the key to display
1208 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1209 * @return The key in hexadecimal form.
1212 mdb_dkey(MDB_val *key, char *buf)
1215 unsigned char *c = key->mv_data;
1221 if (key->mv_size > DKBUF_MAXKEYSIZE)
1222 return "MDB_MAXKEYSIZE";
1223 /* may want to make this a dynamic check: if the key is mostly
1224 * printable characters, print it as-is instead of converting to hex.
1228 for (i=0; i<key->mv_size; i++)
1229 ptr += sprintf(ptr, "%02x", *c++);
1231 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1236 /** Display all the keys in the page. */
1238 mdb_page_list(MDB_page *mp)
1241 unsigned int i, nkeys, nsize, total = 0;
1245 nkeys = NUMKEYS(mp);
1246 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1247 for (i=0; i<nkeys; i++) {
1248 node = NODEPTR(mp, i);
1249 key.mv_size = node->mn_ksize;
1250 key.mv_data = node->mn_data;
1251 nsize = NODESIZE + key.mv_size;
1252 if (IS_BRANCH(mp)) {
1253 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1257 if (F_ISSET(node->mn_flags, F_BIGDATA))
1258 nsize += sizeof(pgno_t);
1260 nsize += NODEDSZ(node);
1262 nsize += sizeof(indx_t);
1263 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1265 total = EVEN(total);
1267 fprintf(stderr, "Total: %d\n", total);
1271 mdb_cursor_chk(MDB_cursor *mc)
1277 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1278 for (i=0; i<mc->mc_top; i++) {
1280 node = NODEPTR(mp, mc->mc_ki[i]);
1281 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1284 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1290 /** Count all the pages in each DB and in the freelist
1291 * and make sure it matches the actual number of pages
1294 static void mdb_audit(MDB_txn *txn)
1298 MDB_ID freecount, count;
1303 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1304 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1305 freecount += *(MDB_ID *)data.mv_data;
1308 for (i = 0; i<txn->mt_numdbs; i++) {
1310 mdb_cursor_init(&mc, txn, i, &mx);
1311 if (txn->mt_dbs[i].md_root == P_INVALID)
1313 count += txn->mt_dbs[i].md_branch_pages +
1314 txn->mt_dbs[i].md_leaf_pages +
1315 txn->mt_dbs[i].md_overflow_pages;
1316 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1317 mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1321 mp = mc.mc_pg[mc.mc_top];
1322 for (j=0; j<NUMKEYS(mp); j++) {
1323 MDB_node *leaf = NODEPTR(mp, j);
1324 if (leaf->mn_flags & F_SUBDATA) {
1326 memcpy(&db, NODEDATA(leaf), sizeof(db));
1327 count += db.md_branch_pages + db.md_leaf_pages +
1328 db.md_overflow_pages;
1332 while (mdb_cursor_sibling(&mc, 1) == 0);
1335 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1336 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1337 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1343 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1345 return txn->mt_dbxs[dbi].md_cmp(a, b);
1349 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1351 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1354 /** Allocate memory for a page.
1355 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1358 mdb_page_malloc(MDB_txn *txn, unsigned num)
1360 MDB_env *env = txn->mt_env;
1361 MDB_page *ret = env->me_dpages;
1362 size_t psize = env->me_psize, sz = psize, off;
1363 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1364 * For a single page alloc, we init everything after the page header.
1365 * For multi-page, we init the final page; if the caller needed that
1366 * many pages they will be filling in at least up to the last page.
1370 VGMEMP_ALLOC(env, ret, sz);
1371 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1372 env->me_dpages = ret->mp_next;
1375 psize -= off = PAGEHDRSZ;
1380 if ((ret = malloc(sz)) != NULL) {
1381 if (!(env->me_flags & MDB_NOMEMINIT)) {
1382 memset((char *)ret + off, 0, psize);
1385 VGMEMP_ALLOC(env, ret, sz);
1390 /** Free a single page.
1391 * Saves single pages to a list, for future reuse.
1392 * (This is not used for multi-page overflow pages.)
1395 mdb_page_free(MDB_env *env, MDB_page *mp)
1397 mp->mp_next = env->me_dpages;
1398 VGMEMP_FREE(env, mp);
1399 env->me_dpages = mp;
1402 /** Free a dirty page */
1404 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1406 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1407 mdb_page_free(env, dp);
1409 /* large pages just get freed directly */
1410 VGMEMP_FREE(env, dp);
1415 /** Return all dirty pages to dpage list */
1417 mdb_dlist_free(MDB_txn *txn)
1419 MDB_env *env = txn->mt_env;
1420 MDB_ID2L dl = txn->mt_u.dirty_list;
1421 unsigned i, n = dl[0].mid;
1423 for (i = 1; i <= n; i++) {
1424 mdb_dpage_free(env, dl[i].mptr);
1429 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1430 * @param[in] mc A cursor handle for the current operation.
1431 * @param[in] pflags Flags of the pages to update:
1432 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1433 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1434 * @return 0 on success, non-zero on failure.
1437 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1439 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1440 MDB_txn *txn = mc->mc_txn;
1446 int rc = MDB_SUCCESS, level;
1448 /* Mark pages seen by cursors */
1449 if (mc->mc_flags & C_UNTRACK)
1450 mc = NULL; /* will find mc in mt_cursors */
1451 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1452 for (; mc; mc=mc->mc_next) {
1453 if (!(mc->mc_flags & C_INITIALIZED))
1455 for (m3 = mc;; m3 = &mx->mx_cursor) {
1457 for (j=0; j<m3->mc_snum; j++) {
1459 if ((mp->mp_flags & Mask) == pflags)
1460 mp->mp_flags ^= P_KEEP;
1462 mx = m3->mc_xcursor;
1463 /* Proceed to mx if it is at a sub-database */
1464 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1466 if (! (mp && (mp->mp_flags & P_LEAF)))
1468 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1469 if (!(leaf->mn_flags & F_SUBDATA))
1478 /* Mark dirty root pages */
1479 for (i=0; i<txn->mt_numdbs; i++) {
1480 if (txn->mt_dbflags[i] & DB_DIRTY) {
1481 pgno_t pgno = txn->mt_dbs[i].md_root;
1482 if (pgno == P_INVALID)
1484 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1486 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1487 dp->mp_flags ^= P_KEEP;
1495 static int mdb_page_flush(MDB_txn *txn, int keep);
1497 /** Spill pages from the dirty list back to disk.
1498 * This is intended to prevent running into #MDB_TXN_FULL situations,
1499 * but note that they may still occur in a few cases:
1500 * 1) our estimate of the txn size could be too small. Currently this
1501 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1502 * 2) child txns may run out of space if their parents dirtied a
1503 * lot of pages and never spilled them. TODO: we probably should do
1504 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1505 * the parent's dirty_room is below a given threshold.
1507 * Otherwise, if not using nested txns, it is expected that apps will
1508 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1509 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1510 * If the txn never references them again, they can be left alone.
1511 * If the txn only reads them, they can be used without any fuss.
1512 * If the txn writes them again, they can be dirtied immediately without
1513 * going thru all of the work of #mdb_page_touch(). Such references are
1514 * handled by #mdb_page_unspill().
1516 * Also note, we never spill DB root pages, nor pages of active cursors,
1517 * because we'll need these back again soon anyway. And in nested txns,
1518 * we can't spill a page in a child txn if it was already spilled in a
1519 * parent txn. That would alter the parent txns' data even though
1520 * the child hasn't committed yet, and we'd have no way to undo it if
1521 * the child aborted.
1523 * @param[in] m0 cursor A cursor handle identifying the transaction and
1524 * database for which we are checking space.
1525 * @param[in] key For a put operation, the key being stored.
1526 * @param[in] data For a put operation, the data being stored.
1527 * @return 0 on success, non-zero on failure.
1530 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1532 MDB_txn *txn = m0->mc_txn;
1534 MDB_ID2L dl = txn->mt_u.dirty_list;
1535 unsigned int i, j, need;
1538 if (m0->mc_flags & C_SUB)
1541 /* Estimate how much space this op will take */
1542 i = m0->mc_db->md_depth;
1543 /* Named DBs also dirty the main DB */
1544 if (m0->mc_dbi > MAIN_DBI)
1545 i += txn->mt_dbs[MAIN_DBI].md_depth;
1546 /* For puts, roughly factor in the key+data size */
1548 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1549 i += i; /* double it for good measure */
1552 if (txn->mt_dirty_room > i)
1555 if (!txn->mt_spill_pgs) {
1556 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1557 if (!txn->mt_spill_pgs)
1560 /* purge deleted slots */
1561 MDB_IDL sl = txn->mt_spill_pgs;
1562 unsigned int num = sl[0];
1564 for (i=1; i<=num; i++) {
1571 /* Preserve pages which may soon be dirtied again */
1572 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1575 /* Less aggressive spill - we originally spilled the entire dirty list,
1576 * with a few exceptions for cursor pages and DB root pages. But this
1577 * turns out to be a lot of wasted effort because in a large txn many
1578 * of those pages will need to be used again. So now we spill only 1/8th
1579 * of the dirty pages. Testing revealed this to be a good tradeoff,
1580 * better than 1/2, 1/4, or 1/10.
1582 if (need < MDB_IDL_UM_MAX / 8)
1583 need = MDB_IDL_UM_MAX / 8;
1585 /* Save the page IDs of all the pages we're flushing */
1586 /* flush from the tail forward, this saves a lot of shifting later on. */
1587 for (i=dl[0].mid; i && need; i--) {
1588 MDB_ID pn = dl[i].mid << 1;
1590 if (dp->mp_flags & P_KEEP)
1592 /* Can't spill twice, make sure it's not already in a parent's
1595 if (txn->mt_parent) {
1597 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1598 if (tx2->mt_spill_pgs) {
1599 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1600 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1601 dp->mp_flags |= P_KEEP;
1609 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1613 mdb_midl_sort(txn->mt_spill_pgs);
1615 /* Flush the spilled part of dirty list */
1616 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1619 /* Reset any dirty pages we kept that page_flush didn't see */
1620 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1623 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1627 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1629 mdb_find_oldest(MDB_txn *txn)
1632 txnid_t mr, oldest = txn->mt_txnid - 1;
1633 if (txn->mt_env->me_txns) {
1634 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1635 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1646 /** Add a page to the txn's dirty list */
1648 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1651 int (*insert)(MDB_ID2L, MDB_ID2 *);
1653 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1654 insert = mdb_mid2l_append;
1656 insert = mdb_mid2l_insert;
1658 mid.mid = mp->mp_pgno;
1660 insert(txn->mt_u.dirty_list, &mid);
1661 txn->mt_dirty_room--;
1664 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1665 * me_pghead and mt_next_pgno.
1667 * If there are free pages available from older transactions, they
1668 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1669 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1670 * and move me_pglast to say which records were consumed. Only this
1671 * function can create me_pghead and move me_pglast/mt_next_pgno.
1672 * @param[in] mc cursor A cursor handle identifying the transaction and
1673 * database for which we are allocating.
1674 * @param[in] num the number of pages to allocate.
1675 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1676 * will always be satisfied by a single contiguous chunk of memory.
1677 * @return 0 on success, non-zero on failure.
1680 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1682 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1683 /* Get at most <Max_retries> more freeDB records once me_pghead
1684 * has enough pages. If not enough, use new pages from the map.
1685 * If <Paranoid> and mc is updating the freeDB, only get new
1686 * records if me_pghead is empty. Then the freelist cannot play
1687 * catch-up with itself by growing while trying to save it.
1689 enum { Paranoid = 1, Max_retries = 500 };
1691 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1693 int rc, retry = Max_retries;
1694 MDB_txn *txn = mc->mc_txn;
1695 MDB_env *env = txn->mt_env;
1696 pgno_t pgno, *mop = env->me_pghead;
1697 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1699 txnid_t oldest = 0, last;
1705 /* If our dirty list is already full, we can't do anything */
1706 if (txn->mt_dirty_room == 0)
1707 return MDB_TXN_FULL;
1709 for (op = MDB_FIRST;; op = MDB_NEXT) {
1712 pgno_t *idl, old_id, new_id;
1714 /* Seek a big enough contiguous page range. Prefer
1715 * pages at the tail, just truncating the list.
1721 if (mop[i-n2] == pgno+n2)
1724 if (Max_retries < INT_MAX && --retry < 0)
1728 if (op == MDB_FIRST) { /* 1st iteration */
1729 /* Prepare to fetch more and coalesce */
1730 oldest = mdb_find_oldest(txn);
1731 last = env->me_pglast;
1732 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1735 key.mv_data = &last; /* will look up last+1 */
1736 key.mv_size = sizeof(last);
1738 if (Paranoid && mc->mc_dbi == FREE_DBI)
1741 if (Paranoid && retry < 0 && mop_len)
1745 /* Do not fetch more if the record will be too recent */
1748 rc = mdb_cursor_get(&m2, &key, NULL, op);
1750 if (rc == MDB_NOTFOUND)
1754 last = *(txnid_t*)key.mv_data;
1757 np = m2.mc_pg[m2.mc_top];
1758 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1759 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1762 idl = (MDB_ID *) data.mv_data;
1765 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1768 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1770 mop = env->me_pghead;
1772 env->me_pglast = last;
1774 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1775 last, txn->mt_dbs[FREE_DBI].md_root, i));
1777 DPRINTF(("IDL %"Z"u", idl[k]));
1779 /* Merge in descending sorted order */
1782 mop[0] = (pgno_t)-1;
1786 for (; old_id < new_id; old_id = mop[--j])
1793 /* Use new pages from the map when nothing suitable in the freeDB */
1795 pgno = txn->mt_next_pgno;
1796 if (pgno + num >= env->me_maxpg) {
1797 DPUTS("DB size maxed out");
1798 return MDB_MAP_FULL;
1802 if (env->me_flags & MDB_WRITEMAP) {
1803 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1805 if (!(np = mdb_page_malloc(txn, num)))
1809 mop[0] = mop_len -= num;
1810 /* Move any stragglers down */
1811 for (j = i-num; j < mop_len; )
1812 mop[++j] = mop[++i];
1814 txn->mt_next_pgno = pgno + num;
1817 mdb_page_dirty(txn, np);
1823 /** Copy the used portions of a non-overflow page.
1824 * @param[in] dst page to copy into
1825 * @param[in] src page to copy from
1826 * @param[in] psize size of a page
1829 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1831 enum { Align = sizeof(pgno_t) };
1832 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1834 /* If page isn't full, just copy the used portion. Adjust
1835 * alignment so memcpy may copy words instead of bytes.
1837 if ((unused &= -Align) && !IS_LEAF2(src)) {
1839 memcpy(dst, src, (lower + (Align-1)) & -Align);
1840 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1843 memcpy(dst, src, psize - unused);
1847 /** Pull a page off the txn's spill list, if present.
1848 * If a page being referenced was spilled to disk in this txn, bring
1849 * it back and make it dirty/writable again.
1850 * @param[in] txn the transaction handle.
1851 * @param[in] mp the page being referenced. It must not be dirty.
1852 * @param[out] ret the writable page, if any. ret is unchanged if
1853 * mp wasn't spilled.
1856 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1858 MDB_env *env = txn->mt_env;
1861 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1863 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1864 if (!tx2->mt_spill_pgs)
1866 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1867 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1870 if (txn->mt_dirty_room == 0)
1871 return MDB_TXN_FULL;
1872 if (IS_OVERFLOW(mp))
1876 if (env->me_flags & MDB_WRITEMAP) {
1879 np = mdb_page_malloc(txn, num);
1883 memcpy(np, mp, num * env->me_psize);
1885 mdb_page_copy(np, mp, env->me_psize);
1888 /* If in current txn, this page is no longer spilled.
1889 * If it happens to be the last page, truncate the spill list.
1890 * Otherwise mark it as deleted by setting the LSB.
1892 if (x == txn->mt_spill_pgs[0])
1893 txn->mt_spill_pgs[0]--;
1895 txn->mt_spill_pgs[x] |= 1;
1896 } /* otherwise, if belonging to a parent txn, the
1897 * page remains spilled until child commits
1900 mdb_page_dirty(txn, np);
1901 np->mp_flags |= P_DIRTY;
1909 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1910 * @param[in] mc cursor pointing to the page to be touched
1911 * @return 0 on success, non-zero on failure.
1914 mdb_page_touch(MDB_cursor *mc)
1916 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1917 MDB_txn *txn = mc->mc_txn;
1918 MDB_cursor *m2, *m3;
1922 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1923 if (txn->mt_flags & MDB_TXN_SPILLS) {
1925 rc = mdb_page_unspill(txn, mp, &np);
1931 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1932 (rc = mdb_page_alloc(mc, 1, &np)))
1935 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
1936 mp->mp_pgno, pgno));
1937 assert(mp->mp_pgno != pgno);
1938 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1939 /* Update the parent page, if any, to point to the new page */
1941 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1942 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1943 SETPGNO(node, pgno);
1945 mc->mc_db->md_root = pgno;
1947 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1948 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1950 /* If txn has a parent, make sure the page is in our
1954 unsigned x = mdb_mid2l_search(dl, pgno);
1955 if (x <= dl[0].mid && dl[x].mid == pgno) {
1956 if (mp != dl[x].mptr) { /* bad cursor? */
1957 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1958 return MDB_CORRUPTED;
1963 assert(dl[0].mid < MDB_IDL_UM_MAX);
1965 np = mdb_page_malloc(txn, 1);
1970 mdb_mid2l_insert(dl, &mid);
1975 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1977 np->mp_flags |= P_DIRTY;
1980 /* Adjust cursors pointing to mp */
1981 mc->mc_pg[mc->mc_top] = np;
1982 m2 = txn->mt_cursors[mc->mc_dbi];
1983 if (mc->mc_flags & C_SUB) {
1984 for (; m2; m2=m2->mc_next) {
1985 m3 = &m2->mc_xcursor->mx_cursor;
1986 if (m3->mc_snum < mc->mc_snum) continue;
1987 if (m3->mc_pg[mc->mc_top] == mp)
1988 m3->mc_pg[mc->mc_top] = np;
1991 for (; m2; m2=m2->mc_next) {
1992 if (m2->mc_snum < mc->mc_snum) continue;
1993 if (m2->mc_pg[mc->mc_top] == mp) {
1994 m2->mc_pg[mc->mc_top] = np;
1995 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1996 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1998 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1999 if (!(leaf->mn_flags & F_SUBDATA))
2000 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2009 mdb_env_sync(MDB_env *env, int force)
2012 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2013 if (env->me_flags & MDB_WRITEMAP) {
2014 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2015 ? MS_ASYNC : MS_SYNC;
2016 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2019 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2023 if (MDB_FDATASYNC(env->me_fd))
2030 /** Back up parent txn's cursors, then grab the originals for tracking */
2032 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2034 MDB_cursor *mc, *bk;
2039 for (i = src->mt_numdbs; --i >= 0; ) {
2040 if ((mc = src->mt_cursors[i]) != NULL) {
2041 size = sizeof(MDB_cursor);
2043 size += sizeof(MDB_xcursor);
2044 for (; mc; mc = bk->mc_next) {
2050 mc->mc_db = &dst->mt_dbs[i];
2051 /* Kill pointers into src - and dst to reduce abuse: The
2052 * user may not use mc until dst ends. Otherwise we'd...
2054 mc->mc_txn = NULL; /* ...set this to dst */
2055 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2056 if ((mx = mc->mc_xcursor) != NULL) {
2057 *(MDB_xcursor *)(bk+1) = *mx;
2058 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2060 mc->mc_next = dst->mt_cursors[i];
2061 dst->mt_cursors[i] = mc;
2068 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2069 * @param[in] txn the transaction handle.
2070 * @param[in] merge true to keep changes to parent cursors, false to revert.
2071 * @return 0 on success, non-zero on failure.
2074 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2076 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2080 for (i = txn->mt_numdbs; --i >= 0; ) {
2081 for (mc = cursors[i]; mc; mc = next) {
2083 if ((bk = mc->mc_backup) != NULL) {
2085 /* Commit changes to parent txn */
2086 mc->mc_next = bk->mc_next;
2087 mc->mc_backup = bk->mc_backup;
2088 mc->mc_txn = bk->mc_txn;
2089 mc->mc_db = bk->mc_db;
2090 mc->mc_dbflag = bk->mc_dbflag;
2091 if ((mx = mc->mc_xcursor) != NULL)
2092 mx->mx_cursor.mc_txn = bk->mc_txn;
2094 /* Abort nested txn */
2096 if ((mx = mc->mc_xcursor) != NULL)
2097 *mx = *(MDB_xcursor *)(bk+1);
2101 /* Only malloced cursors are permanently tracked. */
2109 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2112 mdb_txn_reset0(MDB_txn *txn, const char *act);
2114 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2120 Pidset = F_SETLK, Pidcheck = F_GETLK
2124 /** Set or check a pid lock. Set returns 0 on success.
2125 * Check returns 0 if the process is certainly dead, nonzero if it may
2126 * be alive (the lock exists or an error happened so we do not know).
2128 * On Windows Pidset is a no-op, we merely check for the existence
2129 * of the process with the given pid. On POSIX we use a single byte
2130 * lock on the lockfile, set at an offset equal to the pid.
2133 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2135 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2138 if (op == Pidcheck) {
2139 h = OpenProcess(env->me_pidquery, FALSE, pid);
2140 /* No documented "no such process" code, but other program use this: */
2142 return ErrCode() != ERROR_INVALID_PARAMETER;
2143 /* A process exists until all handles to it close. Has it exited? */
2144 ret = WaitForSingleObject(h, 0) != 0;
2151 struct flock lock_info;
2152 memset(&lock_info, 0, sizeof(lock_info));
2153 lock_info.l_type = F_WRLCK;
2154 lock_info.l_whence = SEEK_SET;
2155 lock_info.l_start = pid;
2156 lock_info.l_len = 1;
2157 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2158 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2160 } else if ((rc = ErrCode()) == EINTR) {
2168 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2169 * @param[in] txn the transaction handle to initialize
2170 * @return 0 on success, non-zero on failure.
2173 mdb_txn_renew0(MDB_txn *txn)
2175 MDB_env *env = txn->mt_env;
2176 MDB_txninfo *ti = env->me_txns;
2180 int rc, new_notls = 0;
2183 txn->mt_numdbs = env->me_numdbs;
2184 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2186 if (txn->mt_flags & MDB_TXN_RDONLY) {
2188 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2189 txn->mt_txnid = meta->mm_txnid;
2190 txn->mt_u.reader = NULL;
2192 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2193 pthread_getspecific(env->me_txkey);
2195 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2196 return MDB_BAD_RSLOT;
2198 MDB_PID_T pid = env->me_pid;
2199 pthread_t tid = pthread_self();
2201 if (!(env->me_flags & MDB_LIVE_READER)) {
2202 rc = mdb_reader_pid(env, Pidset, pid);
2204 UNLOCK_MUTEX_R(env);
2207 env->me_flags |= MDB_LIVE_READER;
2211 nr = ti->mti_numreaders;
2212 for (i=0; i<nr; i++)
2213 if (ti->mti_readers[i].mr_pid == 0)
2215 if (i == env->me_maxreaders) {
2216 UNLOCK_MUTEX_R(env);
2217 return MDB_READERS_FULL;
2219 ti->mti_readers[i].mr_pid = pid;
2220 ti->mti_readers[i].mr_tid = tid;
2222 ti->mti_numreaders = ++nr;
2223 /* Save numreaders for un-mutexed mdb_env_close() */
2224 env->me_numreaders = nr;
2225 UNLOCK_MUTEX_R(env);
2227 r = &ti->mti_readers[i];
2228 new_notls = (env->me_flags & MDB_NOTLS);
2229 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2234 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2235 txn->mt_u.reader = r;
2236 meta = env->me_metas[txn->mt_txnid & 1];
2242 txn->mt_txnid = ti->mti_txnid;
2243 meta = env->me_metas[txn->mt_txnid & 1];
2245 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2246 txn->mt_txnid = meta->mm_txnid;
2250 if (txn->mt_txnid == mdb_debug_start)
2253 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2254 txn->mt_u.dirty_list = env->me_dirty_list;
2255 txn->mt_u.dirty_list[0].mid = 0;
2256 txn->mt_free_pgs = env->me_free_pgs;
2257 txn->mt_free_pgs[0] = 0;
2258 txn->mt_spill_pgs = NULL;
2262 /* Copy the DB info and flags */
2263 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2265 /* Moved to here to avoid a data race in read TXNs */
2266 txn->mt_next_pgno = meta->mm_last_pg+1;
2268 for (i=2; i<txn->mt_numdbs; i++) {
2269 x = env->me_dbflags[i];
2270 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2271 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2273 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2275 if (env->me_maxpg < txn->mt_next_pgno) {
2276 mdb_txn_reset0(txn, "renew0-mapfail");
2278 txn->mt_u.reader->mr_pid = 0;
2279 txn->mt_u.reader = NULL;
2281 return MDB_MAP_RESIZED;
2288 mdb_txn_renew(MDB_txn *txn)
2292 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2295 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2296 DPUTS("environment had fatal error, must shutdown!");
2300 rc = mdb_txn_renew0(txn);
2301 if (rc == MDB_SUCCESS) {
2302 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2303 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2304 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2310 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2314 int rc, size, tsize = sizeof(MDB_txn);
2316 if (env->me_flags & MDB_FATAL_ERROR) {
2317 DPUTS("environment had fatal error, must shutdown!");
2320 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2323 /* Nested transactions: Max 1 child, write txns only, no writemap */
2324 if (parent->mt_child ||
2325 (flags & MDB_RDONLY) ||
2326 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2327 (env->me_flags & MDB_WRITEMAP))
2329 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2331 tsize = sizeof(MDB_ntxn);
2333 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2334 if (!(flags & MDB_RDONLY))
2335 size += env->me_maxdbs * sizeof(MDB_cursor *);
2337 if ((txn = calloc(1, size)) == NULL) {
2338 DPRINTF(("calloc: %s", strerror(ErrCode())));
2341 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2342 if (flags & MDB_RDONLY) {
2343 txn->mt_flags |= MDB_TXN_RDONLY;
2344 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2346 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2347 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2353 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2354 if (!txn->mt_u.dirty_list ||
2355 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2357 free(txn->mt_u.dirty_list);
2361 txn->mt_txnid = parent->mt_txnid;
2362 txn->mt_dirty_room = parent->mt_dirty_room;
2363 txn->mt_u.dirty_list[0].mid = 0;
2364 txn->mt_spill_pgs = NULL;
2365 txn->mt_next_pgno = parent->mt_next_pgno;
2366 parent->mt_child = txn;
2367 txn->mt_parent = parent;
2368 txn->mt_numdbs = parent->mt_numdbs;
2369 txn->mt_flags = parent->mt_flags;
2370 txn->mt_dbxs = parent->mt_dbxs;
2371 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2372 /* Copy parent's mt_dbflags, but clear DB_NEW */
2373 for (i=0; i<txn->mt_numdbs; i++)
2374 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2376 ntxn = (MDB_ntxn *)txn;
2377 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2378 if (env->me_pghead) {
2379 size = MDB_IDL_SIZEOF(env->me_pghead);
2380 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2382 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2387 rc = mdb_cursor_shadow(parent, txn);
2389 mdb_txn_reset0(txn, "beginchild-fail");
2391 rc = mdb_txn_renew0(txn);
2397 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2398 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2399 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2406 mdb_txn_env(MDB_txn *txn)
2408 if(!txn) return NULL;
2412 /** Export or close DBI handles opened in this txn. */
2414 mdb_dbis_update(MDB_txn *txn, int keep)
2417 MDB_dbi n = txn->mt_numdbs;
2418 MDB_env *env = txn->mt_env;
2419 unsigned char *tdbflags = txn->mt_dbflags;
2421 for (i = n; --i >= 2;) {
2422 if (tdbflags[i] & DB_NEW) {
2424 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2426 char *ptr = env->me_dbxs[i].md_name.mv_data;
2427 env->me_dbxs[i].md_name.mv_data = NULL;
2428 env->me_dbxs[i].md_name.mv_size = 0;
2429 env->me_dbflags[i] = 0;
2434 if (keep && env->me_numdbs < n)
2438 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2439 * May be called twice for readonly txns: First reset it, then abort.
2440 * @param[in] txn the transaction handle to reset
2441 * @param[in] act why the transaction is being reset
2444 mdb_txn_reset0(MDB_txn *txn, const char *act)
2446 MDB_env *env = txn->mt_env;
2448 /* Close any DBI handles opened in this txn */
2449 mdb_dbis_update(txn, 0);
2451 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2452 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2453 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2455 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2456 if (txn->mt_u.reader) {
2457 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2458 if (!(env->me_flags & MDB_NOTLS))
2459 txn->mt_u.reader = NULL; /* txn does not own reader */
2461 txn->mt_numdbs = 0; /* close nothing if called again */
2462 txn->mt_dbxs = NULL; /* mark txn as reset */
2464 mdb_cursors_close(txn, 0);
2466 if (!(env->me_flags & MDB_WRITEMAP)) {
2467 mdb_dlist_free(txn);
2469 mdb_midl_free(env->me_pghead);
2471 if (txn->mt_parent) {
2472 txn->mt_parent->mt_child = NULL;
2473 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2474 mdb_midl_free(txn->mt_free_pgs);
2475 mdb_midl_free(txn->mt_spill_pgs);
2476 free(txn->mt_u.dirty_list);
2480 if (mdb_midl_shrink(&txn->mt_free_pgs))
2481 env->me_free_pgs = txn->mt_free_pgs;
2482 env->me_pghead = NULL;
2486 /* The writer mutex was locked in mdb_txn_begin. */
2488 UNLOCK_MUTEX_W(env);
2493 mdb_txn_reset(MDB_txn *txn)
2498 /* This call is only valid for read-only txns */
2499 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2502 mdb_txn_reset0(txn, "reset");
2506 mdb_txn_abort(MDB_txn *txn)
2512 mdb_txn_abort(txn->mt_child);
2514 mdb_txn_reset0(txn, "abort");
2515 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2516 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2517 txn->mt_u.reader->mr_pid = 0;
2522 /** Save the freelist as of this transaction to the freeDB.
2523 * This changes the freelist. Keep trying until it stabilizes.
2526 mdb_freelist_save(MDB_txn *txn)
2528 /* env->me_pghead[] can grow and shrink during this call.
2529 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2530 * Page numbers cannot disappear from txn->mt_free_pgs[].
2533 MDB_env *env = txn->mt_env;
2534 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2535 txnid_t pglast = 0, head_id = 0;
2536 pgno_t freecnt = 0, *free_pgs, *mop;
2537 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2539 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2541 if (env->me_pghead) {
2542 /* Make sure first page of freeDB is touched and on freelist */
2543 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2544 if (rc && rc != MDB_NOTFOUND)
2548 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2549 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2550 ? SSIZE_MAX : maxfree_1pg;
2553 /* Come back here after each Put() in case freelist changed */
2558 /* If using records from freeDB which we have not yet
2559 * deleted, delete them and any we reserved for me_pghead.
2561 while (pglast < env->me_pglast) {
2562 rc = mdb_cursor_first(&mc, &key, NULL);
2565 pglast = head_id = *(txnid_t *)key.mv_data;
2566 total_room = head_room = 0;
2567 assert(pglast <= env->me_pglast);
2568 rc = mdb_cursor_del(&mc, 0);
2573 /* Save the IDL of pages freed by this txn, to a single record */
2574 if (freecnt < txn->mt_free_pgs[0]) {
2576 /* Make sure last page of freeDB is touched and on freelist */
2577 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2578 if (rc && rc != MDB_NOTFOUND)
2581 free_pgs = txn->mt_free_pgs;
2582 /* Write to last page of freeDB */
2583 key.mv_size = sizeof(txn->mt_txnid);
2584 key.mv_data = &txn->mt_txnid;
2586 freecnt = free_pgs[0];
2587 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2588 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2591 /* Retry if mt_free_pgs[] grew during the Put() */
2592 free_pgs = txn->mt_free_pgs;
2593 } while (freecnt < free_pgs[0]);
2594 mdb_midl_sort(free_pgs);
2595 memcpy(data.mv_data, free_pgs, data.mv_size);
2598 unsigned int i = free_pgs[0];
2599 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2600 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2602 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2608 mop = env->me_pghead;
2609 mop_len = mop ? mop[0] : 0;
2611 /* Reserve records for me_pghead[]. Split it if multi-page,
2612 * to avoid searching freeDB for a page range. Use keys in
2613 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2615 if (total_room >= mop_len) {
2616 if (total_room == mop_len || --more < 0)
2618 } else if (head_room >= maxfree_1pg && head_id > 1) {
2619 /* Keep current record (overflow page), add a new one */
2623 /* (Re)write {key = head_id, IDL length = head_room} */
2624 total_room -= head_room;
2625 head_room = mop_len - total_room;
2626 if (head_room > maxfree_1pg && head_id > 1) {
2627 /* Overflow multi-page for part of me_pghead */
2628 head_room /= head_id; /* amortize page sizes */
2629 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2630 } else if (head_room < 0) {
2631 /* Rare case, not bothering to delete this record */
2634 key.mv_size = sizeof(head_id);
2635 key.mv_data = &head_id;
2636 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2637 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2640 /* IDL is initially empty, zero out at least the length */
2641 pgs = (pgno_t *)data.mv_data;
2642 j = head_room > clean_limit ? head_room : 0;
2646 total_room += head_room;
2649 /* Fill in the reserved me_pghead records */
2655 rc = mdb_cursor_first(&mc, &key, &data);
2656 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2657 unsigned flags = MDB_CURRENT;
2658 txnid_t id = *(txnid_t *)key.mv_data;
2659 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2662 assert(len >= 0 && id <= env->me_pglast);
2664 if (len > mop_len) {
2666 data.mv_size = (len + 1) * sizeof(MDB_ID);
2669 data.mv_data = mop -= len;
2672 rc = mdb_cursor_put(&mc, &key, &data, flags);
2674 if (rc || !(mop_len -= len))
2681 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2682 * @param[in] txn the transaction that's being committed
2683 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2684 * @return 0 on success, non-zero on failure.
2687 mdb_page_flush(MDB_txn *txn, int keep)
2689 MDB_env *env = txn->mt_env;
2690 MDB_ID2L dl = txn->mt_u.dirty_list;
2691 unsigned psize = env->me_psize, j;
2692 int i, pagecount = dl[0].mid, rc;
2693 size_t size = 0, pos = 0;
2695 MDB_page *dp = NULL;
2699 struct iovec iov[MDB_COMMIT_PAGES];
2700 ssize_t wpos = 0, wsize = 0, wres;
2701 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2707 if (env->me_flags & MDB_WRITEMAP) {
2708 /* Clear dirty flags */
2709 while (++i <= pagecount) {
2711 /* Don't flush this page yet */
2712 if (dp->mp_flags & P_KEEP) {
2713 dp->mp_flags ^= P_KEEP;
2717 dp->mp_flags &= ~P_DIRTY;
2722 /* Write the pages */
2724 if (++i <= pagecount) {
2726 /* Don't flush this page yet */
2727 if (dp->mp_flags & P_KEEP) {
2728 dp->mp_flags ^= P_KEEP;
2733 /* clear dirty flag */
2734 dp->mp_flags &= ~P_DIRTY;
2737 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2742 /* Windows actually supports scatter/gather I/O, but only on
2743 * unbuffered file handles. Since we're relying on the OS page
2744 * cache for all our data, that's self-defeating. So we just
2745 * write pages one at a time. We use the ov structure to set
2746 * the write offset, to at least save the overhead of a Seek
2749 DPRINTF(("committing page %"Z"u", pgno));
2750 memset(&ov, 0, sizeof(ov));
2751 ov.Offset = pos & 0xffffffff;
2752 ov.OffsetHigh = pos >> 16 >> 16;
2753 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2755 DPRINTF(("WriteFile: %d", rc));
2759 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2760 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2762 /* Write previous page(s) */
2763 #ifdef MDB_USE_PWRITEV
2764 wres = pwritev(env->me_fd, iov, n, wpos);
2767 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2769 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2771 DPRINTF(("lseek: %s", strerror(rc)));
2774 wres = writev(env->me_fd, iov, n);
2777 if (wres != wsize) {
2780 DPRINTF(("Write error: %s", strerror(rc)));
2782 rc = EIO; /* TODO: Use which error code? */
2783 DPUTS("short write, filesystem full?");
2794 DPRINTF(("committing page %"Z"u", pgno));
2795 next_pos = pos + size;
2796 iov[n].iov_len = size;
2797 iov[n].iov_base = (char *)dp;
2803 for (i = keep; ++i <= pagecount; ) {
2805 /* This is a page we skipped above */
2808 dl[j].mid = dp->mp_pgno;
2811 mdb_dpage_free(env, dp);
2816 txn->mt_dirty_room += i - j;
2822 mdb_txn_commit(MDB_txn *txn)
2828 assert(txn != NULL);
2829 assert(txn->mt_env != NULL);
2831 if (txn->mt_child) {
2832 rc = mdb_txn_commit(txn->mt_child);
2833 txn->mt_child = NULL;
2840 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2841 mdb_dbis_update(txn, 1);
2842 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2847 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2848 DPUTS("error flag is set, can't commit");
2850 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2855 if (txn->mt_parent) {
2856 MDB_txn *parent = txn->mt_parent;
2859 unsigned x, y, len, ps_len;
2861 /* Append our free list to parent's */
2862 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2865 mdb_midl_free(txn->mt_free_pgs);
2866 /* Failures after this must either undo the changes
2867 * to the parent or set MDB_TXN_ERROR in the parent.
2870 parent->mt_next_pgno = txn->mt_next_pgno;
2871 parent->mt_flags = txn->mt_flags;
2873 /* Merge our cursors into parent's and close them */
2874 mdb_cursors_close(txn, 1);
2876 /* Update parent's DB table. */
2877 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2878 parent->mt_numdbs = txn->mt_numdbs;
2879 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2880 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2881 for (i=2; i<txn->mt_numdbs; i++) {
2882 /* preserve parent's DB_NEW status */
2883 x = parent->mt_dbflags[i] & DB_NEW;
2884 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2887 dst = parent->mt_u.dirty_list;
2888 src = txn->mt_u.dirty_list;
2889 /* Remove anything in our dirty list from parent's spill list */
2890 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
2892 pspill[0] = (pgno_t)-1;
2893 /* Mark our dirty pages as deleted in parent spill list */
2894 for (i=0, len=src[0].mid; ++i <= len; ) {
2895 MDB_ID pn = src[i].mid << 1;
2896 while (pn > pspill[x])
2898 if (pn == pspill[x]) {
2903 /* Squash deleted pagenums if we deleted any */
2904 for (x=y; ++x <= ps_len; )
2905 if (!(pspill[x] & 1))
2906 pspill[++y] = pspill[x];
2910 /* Find len = length of merging our dirty list with parent's */
2912 dst[0].mid = 0; /* simplify loops */
2913 if (parent->mt_parent) {
2914 len = x + src[0].mid;
2915 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2916 for (i = x; y && i; y--) {
2917 pgno_t yp = src[y].mid;
2918 while (yp < dst[i].mid)
2920 if (yp == dst[i].mid) {
2925 } else { /* Simplify the above for single-ancestor case */
2926 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2928 /* Merge our dirty list with parent's */
2930 for (i = len; y; dst[i--] = src[y--]) {
2931 pgno_t yp = src[y].mid;
2932 while (yp < dst[x].mid)
2933 dst[i--] = dst[x--];
2934 if (yp == dst[x].mid)
2935 free(dst[x--].mptr);
2939 free(txn->mt_u.dirty_list);
2940 parent->mt_dirty_room = txn->mt_dirty_room;
2941 if (txn->mt_spill_pgs) {
2942 if (parent->mt_spill_pgs) {
2943 /* TODO: Prevent failure here, so parent does not fail */
2944 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2946 parent->mt_flags |= MDB_TXN_ERROR;
2947 mdb_midl_free(txn->mt_spill_pgs);
2948 mdb_midl_sort(parent->mt_spill_pgs);
2950 parent->mt_spill_pgs = txn->mt_spill_pgs;
2954 parent->mt_child = NULL;
2955 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2960 if (txn != env->me_txn) {
2961 DPUTS("attempt to commit unknown transaction");
2966 mdb_cursors_close(txn, 0);
2968 if (!txn->mt_u.dirty_list[0].mid &&
2969 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2972 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2973 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2975 /* Update DB root pointers */
2976 if (txn->mt_numdbs > 2) {
2980 data.mv_size = sizeof(MDB_db);
2982 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2983 for (i = 2; i < txn->mt_numdbs; i++) {
2984 if (txn->mt_dbflags[i] & DB_DIRTY) {
2985 data.mv_data = &txn->mt_dbs[i];
2986 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2993 rc = mdb_freelist_save(txn);
2997 mdb_midl_free(env->me_pghead);
2998 env->me_pghead = NULL;
2999 if (mdb_midl_shrink(&txn->mt_free_pgs))
3000 env->me_free_pgs = txn->mt_free_pgs;
3006 if ((rc = mdb_page_flush(txn, 0)) ||
3007 (rc = mdb_env_sync(env, 0)) ||
3008 (rc = mdb_env_write_meta(txn)))
3014 mdb_dbis_update(txn, 1);
3017 UNLOCK_MUTEX_W(env);
3027 /** Read the environment parameters of a DB environment before
3028 * mapping it into memory.
3029 * @param[in] env the environment handle
3030 * @param[out] meta address of where to store the meta information
3031 * @return 0 on success, non-zero on failure.
3034 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3040 enum { Size = sizeof(pbuf) };
3042 /* We don't know the page size yet, so use a minimum value.
3043 * Read both meta pages so we can use the latest one.
3046 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3050 memset(&ov, 0, sizeof(ov));
3052 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3053 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3056 rc = pread(env->me_fd, &pbuf, Size, off);
3059 if (rc == 0 && off == 0)
3061 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3062 DPRINTF(("read: %s", mdb_strerror(rc)));
3066 p = (MDB_page *)&pbuf;
3068 if (!F_ISSET(p->mp_flags, P_META)) {
3069 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3074 if (m->mm_magic != MDB_MAGIC) {
3075 DPUTS("meta has invalid magic");
3079 if (m->mm_version != MDB_DATA_VERSION) {
3080 DPRINTF(("database is version %u, expected version %u",
3081 m->mm_version, MDB_DATA_VERSION));
3082 return MDB_VERSION_MISMATCH;
3085 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3091 /** Write the environment parameters of a freshly created DB environment.
3092 * @param[in] env the environment handle
3093 * @param[out] meta address of where to store the meta information
3094 * @return 0 on success, non-zero on failure.
3097 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3105 memset(&ov, 0, sizeof(ov));
3106 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3108 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3111 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3112 len = pwrite(fd, ptr, size, pos); \
3113 rc = (len >= 0); } while(0)
3116 DPUTS("writing new meta page");
3118 psize = env->me_psize;
3120 meta->mm_magic = MDB_MAGIC;
3121 meta->mm_version = MDB_DATA_VERSION;
3122 meta->mm_mapsize = env->me_mapsize;
3123 meta->mm_psize = psize;
3124 meta->mm_last_pg = 1;
3125 meta->mm_flags = env->me_flags & 0xffff;
3126 meta->mm_flags |= MDB_INTEGERKEY;
3127 meta->mm_dbs[0].md_root = P_INVALID;
3128 meta->mm_dbs[1].md_root = P_INVALID;
3130 p = calloc(2, psize);
3132 p->mp_flags = P_META;
3133 *(MDB_meta *)METADATA(p) = *meta;
3135 q = (MDB_page *)((char *)p + psize);
3137 q->mp_flags = P_META;
3138 *(MDB_meta *)METADATA(q) = *meta;
3140 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3143 else if ((unsigned) len == psize * 2)
3151 /** Update the environment info to commit a transaction.
3152 * @param[in] txn the transaction that's being committed
3153 * @return 0 on success, non-zero on failure.
3156 mdb_env_write_meta(MDB_txn *txn)
3159 MDB_meta meta, metab, *mp;
3161 int rc, len, toggle;
3170 assert(txn != NULL);
3171 assert(txn->mt_env != NULL);
3173 toggle = txn->mt_txnid & 1;
3174 DPRINTF(("writing meta page %d for root page %"Z"u",
3175 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3178 mp = env->me_metas[toggle];
3180 if (env->me_flags & MDB_WRITEMAP) {
3181 /* Persist any increases of mapsize config */
3182 if (env->me_mapsize > mp->mm_mapsize)
3183 mp->mm_mapsize = env->me_mapsize;
3184 mp->mm_dbs[0] = txn->mt_dbs[0];
3185 mp->mm_dbs[1] = txn->mt_dbs[1];
3186 mp->mm_last_pg = txn->mt_next_pgno - 1;
3187 mp->mm_txnid = txn->mt_txnid;
3188 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3189 unsigned meta_size = env->me_psize;
3190 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3193 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3194 if (meta_size < env->me_os_psize)
3195 meta_size += meta_size;
3200 if (MDB_MSYNC(ptr, meta_size, rc)) {
3207 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3208 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3210 ptr = (char *)&meta;
3211 if (env->me_mapsize > mp->mm_mapsize) {
3212 /* Persist any increases of mapsize config */
3213 meta.mm_mapsize = env->me_mapsize;
3214 off = offsetof(MDB_meta, mm_mapsize);
3216 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3218 len = sizeof(MDB_meta) - off;
3221 meta.mm_dbs[0] = txn->mt_dbs[0];
3222 meta.mm_dbs[1] = txn->mt_dbs[1];
3223 meta.mm_last_pg = txn->mt_next_pgno - 1;
3224 meta.mm_txnid = txn->mt_txnid;
3227 off += env->me_psize;
3230 /* Write to the SYNC fd */
3231 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3232 env->me_fd : env->me_mfd;
3235 memset(&ov, 0, sizeof(ov));
3237 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3241 rc = pwrite(mfd, ptr, len, off);
3244 rc = rc < 0 ? ErrCode() : EIO;
3245 DPUTS("write failed, disk error?");
3246 /* On a failure, the pagecache still contains the new data.
3247 * Write some old data back, to prevent it from being used.
3248 * Use the non-SYNC fd; we know it will fail anyway.
3250 meta.mm_last_pg = metab.mm_last_pg;
3251 meta.mm_txnid = metab.mm_txnid;
3253 memset(&ov, 0, sizeof(ov));
3255 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3257 r2 = pwrite(env->me_fd, ptr, len, off);
3258 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3261 env->me_flags |= MDB_FATAL_ERROR;
3265 /* Memory ordering issues are irrelevant; since the entire writer
3266 * is wrapped by wmutex, all of these changes will become visible
3267 * after the wmutex is unlocked. Since the DB is multi-version,
3268 * readers will get consistent data regardless of how fresh or
3269 * how stale their view of these values is.
3272 env->me_txns->mti_txnid = txn->mt_txnid;
3277 /** Check both meta pages to see which one is newer.
3278 * @param[in] env the environment handle
3279 * @return meta toggle (0 or 1).
3282 mdb_env_pick_meta(const MDB_env *env)
3284 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3288 mdb_env_create(MDB_env **env)
3292 e = calloc(1, sizeof(MDB_env));
3296 e->me_maxreaders = DEFAULT_READERS;
3297 e->me_maxdbs = e->me_numdbs = 2;
3298 e->me_fd = INVALID_HANDLE_VALUE;
3299 e->me_lfd = INVALID_HANDLE_VALUE;
3300 e->me_mfd = INVALID_HANDLE_VALUE;
3301 #ifdef MDB_USE_POSIX_SEM
3302 e->me_rmutex = SEM_FAILED;
3303 e->me_wmutex = SEM_FAILED;
3305 e->me_pid = getpid();
3306 GET_PAGESIZE(e->me_os_psize);
3307 VGMEMP_CREATE(e,0,0);
3313 mdb_env_map(MDB_env *env, void *addr, int newsize)
3316 unsigned int flags = env->me_flags;
3320 LONG sizelo, sizehi;
3321 sizelo = env->me_mapsize & 0xffffffff;
3322 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3324 /* Windows won't create mappings for zero length files.
3325 * Just allocate the maxsize right now.
3328 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3329 || !SetEndOfFile(env->me_fd)
3330 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3333 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3334 PAGE_READWRITE : PAGE_READONLY,
3335 sizehi, sizelo, NULL);
3338 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3339 FILE_MAP_WRITE : FILE_MAP_READ,
3340 0, 0, env->me_mapsize, addr);
3341 rc = env->me_map ? 0 : ErrCode();
3346 int prot = PROT_READ;
3347 if (flags & MDB_WRITEMAP) {
3349 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3352 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3354 if (env->me_map == MAP_FAILED) {
3359 if (flags & MDB_NORDAHEAD) {
3360 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3362 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3364 #ifdef POSIX_MADV_RANDOM
3365 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3366 #endif /* POSIX_MADV_RANDOM */
3367 #endif /* MADV_RANDOM */
3371 /* Can happen because the address argument to mmap() is just a
3372 * hint. mmap() can pick another, e.g. if the range is in use.
3373 * The MAP_FIXED flag would prevent that, but then mmap could
3374 * instead unmap existing pages to make room for the new map.
3376 if (addr && env->me_map != addr)
3377 return EBUSY; /* TODO: Make a new MDB_* error code? */
3379 p = (MDB_page *)env->me_map;
3380 env->me_metas[0] = METADATA(p);
3381 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3387 mdb_env_set_mapsize(MDB_env *env, size_t size)
3389 /* If env is already open, caller is responsible for making
3390 * sure there are no active txns.
3398 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3399 else if (size < env->me_mapsize) {
3400 /* If the configured size is smaller, make sure it's
3401 * still big enough. Silently round up to minimum if not.
3403 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3407 munmap(env->me_map, env->me_mapsize);
3408 env->me_mapsize = size;
3409 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3410 rc = mdb_env_map(env, old, 1);
3414 env->me_mapsize = size;
3416 env->me_maxpg = env->me_mapsize / env->me_psize;
3421 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3425 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3430 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3432 if (env->me_map || readers < 1)
3434 env->me_maxreaders = readers;
3439 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3441 if (!env || !readers)
3443 *readers = env->me_maxreaders;
3447 /** Further setup required for opening an MDB environment
3450 mdb_env_open2(MDB_env *env)
3452 unsigned int flags = env->me_flags;
3453 int i, newenv = 0, rc;
3457 /* See if we should use QueryLimited */
3459 if ((rc & 0xff) > 5)
3460 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3462 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3465 memset(&meta, 0, sizeof(meta));
3467 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3470 DPUTS("new mdbenv");
3472 env->me_psize = env->me_os_psize;
3473 if (env->me_psize > MAX_PAGESIZE)
3474 env->me_psize = MAX_PAGESIZE;
3476 env->me_psize = meta.mm_psize;
3479 /* Was a mapsize configured? */
3480 if (!env->me_mapsize) {
3481 /* If this is a new environment, take the default,
3482 * else use the size recorded in the existing env.
3484 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3485 } else if (env->me_mapsize < meta.mm_mapsize) {
3486 /* If the configured size is smaller, make sure it's
3487 * still big enough. Silently round up to minimum if not.
3489 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3490 if (env->me_mapsize < minsize)
3491 env->me_mapsize = minsize;
3494 rc = mdb_env_map(env, meta.mm_address, newenv);
3499 if (flags & MDB_FIXEDMAP)
3500 meta.mm_address = env->me_map;
3501 i = mdb_env_init_meta(env, &meta);
3502 if (i != MDB_SUCCESS) {
3507 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3508 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3510 #if !(MDB_MAXKEYSIZE)
3511 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3513 env->me_maxpg = env->me_mapsize / env->me_psize;
3517 int toggle = mdb_env_pick_meta(env);
3518 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3520 DPRINTF(("opened database version %u, pagesize %u",
3521 env->me_metas[0]->mm_version, env->me_psize));
3522 DPRINTF(("using meta page %d", toggle));
3523 DPRINTF(("depth: %u", db->md_depth));
3524 DPRINTF(("entries: %"Z"u", db->md_entries));
3525 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3526 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3527 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3528 DPRINTF(("root: %"Z"u", db->md_root));
3536 /** Release a reader thread's slot in the reader lock table.
3537 * This function is called automatically when a thread exits.
3538 * @param[in] ptr This points to the slot in the reader lock table.
3541 mdb_env_reader_dest(void *ptr)
3543 MDB_reader *reader = ptr;
3549 /** Junk for arranging thread-specific callbacks on Windows. This is
3550 * necessarily platform and compiler-specific. Windows supports up
3551 * to 1088 keys. Let's assume nobody opens more than 64 environments
3552 * in a single process, for now. They can override this if needed.
3554 #ifndef MAX_TLS_KEYS
3555 #define MAX_TLS_KEYS 64
3557 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3558 static int mdb_tls_nkeys;
3560 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3564 case DLL_PROCESS_ATTACH: break;
3565 case DLL_THREAD_ATTACH: break;
3566 case DLL_THREAD_DETACH:
3567 for (i=0; i<mdb_tls_nkeys; i++) {
3568 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3569 mdb_env_reader_dest(r);
3572 case DLL_PROCESS_DETACH: break;
3577 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3579 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3583 /* Force some symbol references.
3584 * _tls_used forces the linker to create the TLS directory if not already done
3585 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3587 #pragma comment(linker, "/INCLUDE:_tls_used")
3588 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3589 #pragma const_seg(".CRT$XLB")
3590 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3591 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3594 #pragma comment(linker, "/INCLUDE:__tls_used")
3595 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3596 #pragma data_seg(".CRT$XLB")
3597 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3599 #endif /* WIN 32/64 */
3600 #endif /* !__GNUC__ */
3603 /** Downgrade the exclusive lock on the region back to shared */
3605 mdb_env_share_locks(MDB_env *env, int *excl)
3607 int rc = 0, toggle = mdb_env_pick_meta(env);
3609 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3614 /* First acquire a shared lock. The Unlock will
3615 * then release the existing exclusive lock.
3617 memset(&ov, 0, sizeof(ov));
3618 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3621 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3627 struct flock lock_info;
3628 /* The shared lock replaces the existing lock */
3629 memset((void *)&lock_info, 0, sizeof(lock_info));
3630 lock_info.l_type = F_RDLCK;
3631 lock_info.l_whence = SEEK_SET;
3632 lock_info.l_start = 0;
3633 lock_info.l_len = 1;
3634 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3635 (rc = ErrCode()) == EINTR) ;
3636 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3643 /** Try to get exlusive lock, otherwise shared.
3644 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3647 mdb_env_excl_lock(MDB_env *env, int *excl)
3651 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3655 memset(&ov, 0, sizeof(ov));
3656 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3663 struct flock lock_info;
3664 memset((void *)&lock_info, 0, sizeof(lock_info));
3665 lock_info.l_type = F_WRLCK;
3666 lock_info.l_whence = SEEK_SET;
3667 lock_info.l_start = 0;
3668 lock_info.l_len = 1;
3669 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3670 (rc = ErrCode()) == EINTR) ;
3674 # ifdef MDB_USE_POSIX_SEM
3675 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3678 lock_info.l_type = F_RDLCK;
3679 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3680 (rc = ErrCode()) == EINTR) ;
3690 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3692 * @(#) $Revision: 5.1 $
3693 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3694 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3696 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3700 * Please do not copyright this code. This code is in the public domain.
3702 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3703 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3704 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3705 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3706 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3707 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3708 * PERFORMANCE OF THIS SOFTWARE.
3711 * chongo <Landon Curt Noll> /\oo/\
3712 * http://www.isthe.com/chongo/
3714 * Share and Enjoy! :-)
3717 typedef unsigned long long mdb_hash_t;
3718 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3720 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3721 * @param[in] val value to hash
3722 * @param[in] hval initial value for hash
3723 * @return 64 bit hash
3725 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3726 * hval arg on the first call.
3729 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3731 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3732 unsigned char *end = s + val->mv_size;
3734 * FNV-1a hash each octet of the string
3737 /* xor the bottom with the current octet */
3738 hval ^= (mdb_hash_t)*s++;
3740 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3741 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3742 (hval << 7) + (hval << 8) + (hval << 40);
3744 /* return our new hash value */
3748 /** Hash the string and output the encoded hash.
3749 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3750 * very short name limits. We don't care about the encoding being reversible,
3751 * we just want to preserve as many bits of the input as possible in a
3752 * small printable string.
3753 * @param[in] str string to hash
3754 * @param[out] encbuf an array of 11 chars to hold the hash
3756 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3759 mdb_pack85(unsigned long l, char *out)
3763 for (i=0; i<5; i++) {
3764 *out++ = mdb_a85[l % 85];
3770 mdb_hash_enc(MDB_val *val, char *encbuf)
3772 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3774 mdb_pack85(h, encbuf);
3775 mdb_pack85(h>>32, encbuf+5);
3780 /** Open and/or initialize the lock region for the environment.
3781 * @param[in] env The MDB environment.
3782 * @param[in] lpath The pathname of the file used for the lock region.
3783 * @param[in] mode The Unix permissions for the file, if we create it.
3784 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3785 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3786 * @return 0 on success, non-zero on failure.
3789 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3792 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3794 # define MDB_ERRCODE_ROFS EROFS
3795 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3796 # define MDB_CLOEXEC O_CLOEXEC
3799 # define MDB_CLOEXEC 0
3806 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3807 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3808 FILE_ATTRIBUTE_NORMAL, NULL);
3810 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3812 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3814 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3819 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3820 /* Lose record locks when exec*() */
3821 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3822 fcntl(env->me_lfd, F_SETFD, fdflags);
3825 if (!(env->me_flags & MDB_NOTLS)) {
3826 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3829 env->me_flags |= MDB_ENV_TXKEY;
3831 /* Windows TLS callbacks need help finding their TLS info. */
3832 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3836 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3840 /* Try to get exclusive lock. If we succeed, then
3841 * nobody is using the lock region and we should initialize it.
3843 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3846 size = GetFileSize(env->me_lfd, NULL);
3848 size = lseek(env->me_lfd, 0, SEEK_END);
3849 if (size == -1) goto fail_errno;
3851 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3852 if (size < rsize && *excl > 0) {
3854 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
3855 || !SetEndOfFile(env->me_lfd))
3858 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3862 size = rsize - sizeof(MDB_txninfo);
3863 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3868 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3870 if (!mh) goto fail_errno;
3871 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3873 if (!env->me_txns) goto fail_errno;
3875 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3877 if (m == MAP_FAILED) goto fail_errno;
3883 BY_HANDLE_FILE_INFORMATION stbuf;
3892 if (!mdb_sec_inited) {
3893 InitializeSecurityDescriptor(&mdb_null_sd,
3894 SECURITY_DESCRIPTOR_REVISION);
3895 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3896 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3897 mdb_all_sa.bInheritHandle = FALSE;
3898 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3901 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3902 idbuf.volume = stbuf.dwVolumeSerialNumber;
3903 idbuf.nhigh = stbuf.nFileIndexHigh;
3904 idbuf.nlow = stbuf.nFileIndexLow;
3905 val.mv_data = &idbuf;
3906 val.mv_size = sizeof(idbuf);
3907 mdb_hash_enc(&val, encbuf);
3908 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3909 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3910 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3911 if (!env->me_rmutex) goto fail_errno;
3912 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3913 if (!env->me_wmutex) goto fail_errno;
3914 #elif defined(MDB_USE_POSIX_SEM)
3923 #if defined(__NetBSD__)
3924 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3926 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3927 idbuf.dev = stbuf.st_dev;
3928 idbuf.ino = stbuf.st_ino;
3929 val.mv_data = &idbuf;
3930 val.mv_size = sizeof(idbuf);
3931 mdb_hash_enc(&val, encbuf);
3932 #ifdef MDB_SHORT_SEMNAMES
3933 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3935 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3936 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3937 /* Clean up after a previous run, if needed: Try to
3938 * remove both semaphores before doing anything else.
3940 sem_unlink(env->me_txns->mti_rmname);
3941 sem_unlink(env->me_txns->mti_wmname);
3942 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3943 O_CREAT|O_EXCL, mode, 1);
3944 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3945 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3946 O_CREAT|O_EXCL, mode, 1);
3947 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3948 #else /* MDB_USE_POSIX_SEM */
3949 pthread_mutexattr_t mattr;
3951 if ((rc = pthread_mutexattr_init(&mattr))
3952 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3953 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3954 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3956 pthread_mutexattr_destroy(&mattr);
3957 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3959 env->me_txns->mti_magic = MDB_MAGIC;
3960 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3961 env->me_txns->mti_txnid = 0;
3962 env->me_txns->mti_numreaders = 0;
3965 if (env->me_txns->mti_magic != MDB_MAGIC) {
3966 DPUTS("lock region has invalid magic");
3970 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3971 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3972 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3973 rc = MDB_VERSION_MISMATCH;
3977 if (rc && rc != EACCES && rc != EAGAIN) {
3981 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3982 if (!env->me_rmutex) goto fail_errno;
3983 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3984 if (!env->me_wmutex) goto fail_errno;
3985 #elif defined(MDB_USE_POSIX_SEM)
3986 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3987 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3988 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3989 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4000 /** The name of the lock file in the DB environment */
4001 #define LOCKNAME "/lock.mdb"
4002 /** The name of the data file in the DB environment */
4003 #define DATANAME "/data.mdb"
4004 /** The suffix of the lock file when no subdir is used */
4005 #define LOCKSUFF "-lock"
4006 /** Only a subset of the @ref mdb_env flags can be changed
4007 * at runtime. Changing other flags requires closing the
4008 * environment and re-opening it with the new flags.
4010 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4011 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4012 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4015 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4017 int oflags, rc, len, excl = -1;
4018 char *lpath, *dpath;
4020 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4024 if (flags & MDB_NOSUBDIR) {
4025 rc = len + sizeof(LOCKSUFF) + len + 1;
4027 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4032 if (flags & MDB_NOSUBDIR) {
4033 dpath = lpath + len + sizeof(LOCKSUFF);
4034 sprintf(lpath, "%s" LOCKSUFF, path);
4035 strcpy(dpath, path);
4037 dpath = lpath + len + sizeof(LOCKNAME);
4038 sprintf(lpath, "%s" LOCKNAME, path);
4039 sprintf(dpath, "%s" DATANAME, path);
4043 flags |= env->me_flags;
4044 if (flags & MDB_RDONLY) {
4045 /* silently ignore WRITEMAP when we're only getting read access */
4046 flags &= ~MDB_WRITEMAP;
4048 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4049 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4052 env->me_flags = flags |= MDB_ENV_ACTIVE;
4056 env->me_path = strdup(path);
4057 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4058 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4059 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4064 /* For RDONLY, get lockfile after we know datafile exists */
4065 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4066 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4072 if (F_ISSET(flags, MDB_RDONLY)) {
4073 oflags = GENERIC_READ;
4074 len = OPEN_EXISTING;
4076 oflags = GENERIC_READ|GENERIC_WRITE;
4079 mode = FILE_ATTRIBUTE_NORMAL;
4080 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4081 NULL, len, mode, NULL);
4083 if (F_ISSET(flags, MDB_RDONLY))
4086 oflags = O_RDWR | O_CREAT;
4088 env->me_fd = open(dpath, oflags, mode);
4090 if (env->me_fd == INVALID_HANDLE_VALUE) {
4095 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4096 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4101 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4102 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4103 env->me_mfd = env->me_fd;
4105 /* Synchronous fd for meta writes. Needed even with
4106 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4109 len = OPEN_EXISTING;
4110 env->me_mfd = CreateFile(dpath, oflags,
4111 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4112 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4115 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4117 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4122 DPRINTF(("opened dbenv %p", (void *) env));
4124 rc = mdb_env_share_locks(env, &excl);
4128 if (!((flags & MDB_RDONLY) ||
4129 (env->me_pbuf = calloc(1, env->me_psize))))
4135 mdb_env_close0(env, excl);
4141 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4143 mdb_env_close0(MDB_env *env, int excl)
4147 if (!(env->me_flags & MDB_ENV_ACTIVE))
4150 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4151 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4152 free(env->me_dbxs[i].md_name.mv_data);
4155 free(env->me_dbflags);
4158 free(env->me_dirty_list);
4159 mdb_midl_free(env->me_free_pgs);
4161 if (env->me_flags & MDB_ENV_TXKEY) {
4162 pthread_key_delete(env->me_txkey);
4164 /* Delete our key from the global list */
4165 for (i=0; i<mdb_tls_nkeys; i++)
4166 if (mdb_tls_keys[i] == env->me_txkey) {
4167 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4175 munmap(env->me_map, env->me_mapsize);
4177 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4178 (void) close(env->me_mfd);
4179 if (env->me_fd != INVALID_HANDLE_VALUE)
4180 (void) close(env->me_fd);
4182 MDB_PID_T pid = env->me_pid;
4183 /* Clearing readers is done in this function because
4184 * me_txkey with its destructor must be disabled first.
4186 for (i = env->me_numreaders; --i >= 0; )
4187 if (env->me_txns->mti_readers[i].mr_pid == pid)
4188 env->me_txns->mti_readers[i].mr_pid = 0;
4190 if (env->me_rmutex) {
4191 CloseHandle(env->me_rmutex);
4192 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4194 /* Windows automatically destroys the mutexes when
4195 * the last handle closes.
4197 #elif defined(MDB_USE_POSIX_SEM)
4198 if (env->me_rmutex != SEM_FAILED) {
4199 sem_close(env->me_rmutex);
4200 if (env->me_wmutex != SEM_FAILED)
4201 sem_close(env->me_wmutex);
4202 /* If we have the filelock: If we are the
4203 * only remaining user, clean up semaphores.
4206 mdb_env_excl_lock(env, &excl);
4208 sem_unlink(env->me_txns->mti_rmname);
4209 sem_unlink(env->me_txns->mti_wmname);
4213 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4215 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4218 /* Unlock the lockfile. Windows would have unlocked it
4219 * after closing anyway, but not necessarily at once.
4221 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4224 (void) close(env->me_lfd);
4227 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4231 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4233 MDB_txn *txn = NULL;
4239 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4243 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4246 /* Do the lock/unlock of the reader mutex before starting the
4247 * write txn. Otherwise other read txns could block writers.
4249 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4254 /* We must start the actual read txn after blocking writers */
4255 mdb_txn_reset0(txn, "reset-stage1");
4257 /* Temporarily block writers until we snapshot the meta pages */
4260 rc = mdb_txn_renew0(txn);
4262 UNLOCK_MUTEX_W(env);
4267 wsize = env->me_psize * 2;
4271 DO_WRITE(rc, fd, ptr, w2, len);
4275 } else if (len > 0) {
4281 /* Non-blocking or async handles are not supported */
4287 UNLOCK_MUTEX_W(env);
4292 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4294 if (wsize > MAX_WRITE)
4298 DO_WRITE(rc, fd, ptr, w2, len);
4302 } else if (len > 0) {
4319 mdb_env_copy(MDB_env *env, const char *path)
4323 HANDLE newfd = INVALID_HANDLE_VALUE;
4325 if (env->me_flags & MDB_NOSUBDIR) {
4326 lpath = (char *)path;
4329 len += sizeof(DATANAME);
4330 lpath = malloc(len);
4333 sprintf(lpath, "%s" DATANAME, path);
4336 /* The destination path must exist, but the destination file must not.
4337 * We don't want the OS to cache the writes, since the source data is
4338 * already in the OS cache.
4341 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4342 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4344 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4346 if (newfd == INVALID_HANDLE_VALUE) {
4352 /* Set O_DIRECT if the file system supports it */
4353 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4354 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4356 #ifdef F_NOCACHE /* __APPLE__ */
4357 rc = fcntl(newfd, F_NOCACHE, 1);
4364 rc = mdb_env_copyfd(env, newfd);
4367 if (!(env->me_flags & MDB_NOSUBDIR))
4369 if (newfd != INVALID_HANDLE_VALUE)
4370 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4377 mdb_env_close(MDB_env *env)
4384 VGMEMP_DESTROY(env);
4385 while ((dp = env->me_dpages) != NULL) {
4386 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4387 env->me_dpages = dp->mp_next;
4391 mdb_env_close0(env, 0);
4395 /** Compare two items pointing at aligned size_t's */
4397 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4399 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4400 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4403 /** Compare two items pointing at aligned unsigned int's */
4405 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4407 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4408 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4411 /** Compare two items pointing at unsigned ints of unknown alignment.
4412 * Nodes and keys are guaranteed to be 2-byte aligned.
4415 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4417 #if BYTE_ORDER == LITTLE_ENDIAN
4418 unsigned short *u, *c;
4421 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4422 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4425 } while(!x && u > (unsigned short *)a->mv_data);
4428 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4432 /** Compare two items lexically */
4434 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4441 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4447 diff = memcmp(a->mv_data, b->mv_data, len);
4448 return diff ? diff : len_diff<0 ? -1 : len_diff;
4451 /** Compare two items in reverse byte order */
4453 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4455 const unsigned char *p1, *p2, *p1_lim;
4459 p1_lim = (const unsigned char *)a->mv_data;
4460 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4461 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4463 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4469 while (p1 > p1_lim) {
4470 diff = *--p1 - *--p2;
4474 return len_diff<0 ? -1 : len_diff;
4477 /** Search for key within a page, using binary search.
4478 * Returns the smallest entry larger or equal to the key.
4479 * If exactp is non-null, stores whether the found entry was an exact match
4480 * in *exactp (1 or 0).
4481 * Updates the cursor index with the index of the found entry.
4482 * If no entry larger or equal to the key is found, returns NULL.
4485 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4487 unsigned int i = 0, nkeys;
4490 MDB_page *mp = mc->mc_pg[mc->mc_top];
4491 MDB_node *node = NULL;
4496 nkeys = NUMKEYS(mp);
4501 COPY_PGNO(pgno, mp->mp_pgno);
4502 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4503 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4510 low = IS_LEAF(mp) ? 0 : 1;
4512 cmp = mc->mc_dbx->md_cmp;
4514 /* Branch pages have no data, so if using integer keys,
4515 * alignment is guaranteed. Use faster mdb_cmp_int.
4517 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4518 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4525 nodekey.mv_size = mc->mc_db->md_pad;
4526 node = NODEPTR(mp, 0); /* fake */
4527 while (low <= high) {
4528 i = (low + high) >> 1;
4529 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4530 rc = cmp(key, &nodekey);
4531 DPRINTF(("found leaf index %u [%s], rc = %i",
4532 i, DKEY(&nodekey), rc));
4541 while (low <= high) {
4542 i = (low + high) >> 1;
4544 node = NODEPTR(mp, i);
4545 nodekey.mv_size = NODEKSZ(node);
4546 nodekey.mv_data = NODEKEY(node);
4548 rc = cmp(key, &nodekey);
4551 DPRINTF(("found leaf index %u [%s], rc = %i",
4552 i, DKEY(&nodekey), rc));
4554 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4555 i, DKEY(&nodekey), NODEPGNO(node), rc));
4566 if (rc > 0) { /* Found entry is less than the key. */
4567 i++; /* Skip to get the smallest entry larger than key. */
4569 node = NODEPTR(mp, i);
4572 *exactp = (rc == 0);
4573 /* store the key index */
4574 mc->mc_ki[mc->mc_top] = i;
4576 /* There is no entry larger or equal to the key. */
4579 /* nodeptr is fake for LEAF2 */
4585 mdb_cursor_adjust(MDB_cursor *mc, func)
4589 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4590 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4597 /** Pop a page off the top of the cursor's stack. */
4599 mdb_cursor_pop(MDB_cursor *mc)
4603 MDB_page *top = mc->mc_pg[mc->mc_top];
4609 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4610 DDBI(mc), (void *) mc));
4614 /** Push a page onto the top of the cursor's stack. */
4616 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4618 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4619 DDBI(mc), (void *) mc));
4621 if (mc->mc_snum >= CURSOR_STACK) {
4622 assert(mc->mc_snum < CURSOR_STACK);
4623 return MDB_CURSOR_FULL;
4626 mc->mc_top = mc->mc_snum++;
4627 mc->mc_pg[mc->mc_top] = mp;
4628 mc->mc_ki[mc->mc_top] = 0;
4633 /** Find the address of the page corresponding to a given page number.
4634 * @param[in] txn the transaction for this access.
4635 * @param[in] pgno the page number for the page to retrieve.
4636 * @param[out] ret address of a pointer where the page's address will be stored.
4637 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4638 * @return 0 on success, non-zero on failure.
4641 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4643 MDB_env *env = txn->mt_env;
4647 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4651 MDB_ID2L dl = tx2->mt_u.dirty_list;
4653 /* Spilled pages were dirtied in this txn and flushed
4654 * because the dirty list got full. Bring this page
4655 * back in from the map (but don't unspill it here,
4656 * leave that unless page_touch happens again).
4658 if (tx2->mt_spill_pgs) {
4659 MDB_ID pn = pgno << 1;
4660 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4661 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4662 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4667 unsigned x = mdb_mid2l_search(dl, pgno);
4668 if (x <= dl[0].mid && dl[x].mid == pgno) {
4674 } while ((tx2 = tx2->mt_parent) != NULL);
4677 if (pgno < txn->mt_next_pgno) {
4679 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4681 DPRINTF(("page %"Z"u not found", pgno));
4683 return MDB_PAGE_NOTFOUND;
4693 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4694 * The cursor is at the root page, set up the rest of it.
4697 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4699 MDB_page *mp = mc->mc_pg[mc->mc_top];
4703 while (IS_BRANCH(mp)) {
4707 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4708 assert(NUMKEYS(mp) > 1);
4709 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4711 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4713 if (flags & MDB_PS_LAST)
4714 i = NUMKEYS(mp) - 1;
4717 node = mdb_node_search(mc, key, &exact);
4719 i = NUMKEYS(mp) - 1;
4721 i = mc->mc_ki[mc->mc_top];
4727 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4730 assert(i < NUMKEYS(mp));
4731 node = NODEPTR(mp, i);
4733 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4736 mc->mc_ki[mc->mc_top] = i;
4737 if ((rc = mdb_cursor_push(mc, mp)))
4740 if (flags & MDB_PS_MODIFY) {
4741 if ((rc = mdb_page_touch(mc)) != 0)
4743 mp = mc->mc_pg[mc->mc_top];
4748 DPRINTF(("internal error, index points to a %02X page!?",
4750 return MDB_CORRUPTED;
4753 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4754 key ? DKEY(key) : "null"));
4755 mc->mc_flags |= C_INITIALIZED;
4756 mc->mc_flags &= ~C_EOF;
4761 /** Search for the lowest key under the current branch page.
4762 * This just bypasses a NUMKEYS check in the current page
4763 * before calling mdb_page_search_root(), because the callers
4764 * are all in situations where the current page is known to
4768 mdb_page_search_lowest(MDB_cursor *mc)
4770 MDB_page *mp = mc->mc_pg[mc->mc_top];
4771 MDB_node *node = NODEPTR(mp, 0);
4774 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4777 mc->mc_ki[mc->mc_top] = 0;
4778 if ((rc = mdb_cursor_push(mc, mp)))
4780 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4783 /** Search for the page a given key should be in.
4784 * Push it and its parent pages on the cursor stack.
4785 * @param[in,out] mc the cursor for this operation.
4786 * @param[in] key the key to search for, or NULL for first/last page.
4787 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4788 * are touched (updated with new page numbers).
4789 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4790 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4791 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4792 * @return 0 on success, non-zero on failure.
4795 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4800 /* Make sure the txn is still viable, then find the root from
4801 * the txn's db table and set it as the root of the cursor's stack.
4803 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4804 DPUTS("transaction has failed, must abort");
4807 /* Make sure we're using an up-to-date root */
4808 if (*mc->mc_dbflag & DB_STALE) {
4810 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4811 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4818 MDB_node *leaf = mdb_node_search(&mc2,
4819 &mc->mc_dbx->md_name, &exact);
4821 return MDB_NOTFOUND;
4822 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4825 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4827 /* The txn may not know this DBI, or another process may
4828 * have dropped and recreated the DB with other flags.
4830 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4831 return MDB_INCOMPATIBLE;
4832 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4834 *mc->mc_dbflag &= ~DB_STALE;
4836 root = mc->mc_db->md_root;
4838 if (root == P_INVALID) { /* Tree is empty. */
4839 DPUTS("tree is empty");
4840 return MDB_NOTFOUND;
4845 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4846 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4852 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4853 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4855 if (flags & MDB_PS_MODIFY) {
4856 if ((rc = mdb_page_touch(mc)))
4860 if (flags & MDB_PS_ROOTONLY)
4863 return mdb_page_search_root(mc, key, flags);
4867 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4869 MDB_txn *txn = mc->mc_txn;
4870 pgno_t pg = mp->mp_pgno;
4871 unsigned x = 0, ovpages = mp->mp_pages;
4872 MDB_env *env = txn->mt_env;
4873 MDB_IDL sl = txn->mt_spill_pgs;
4874 MDB_ID pn = pg << 1;
4877 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4878 /* If the page is dirty or on the spill list we just acquired it,
4879 * so we should give it back to our current free list, if any.
4880 * Otherwise put it onto the list of pages we freed in this txn.
4882 * Won't create me_pghead: me_pglast must be inited along with it.
4883 * Unsupported in nested txns: They would need to hide the page
4884 * range in ancestor txns' dirty and spilled lists.
4886 if (env->me_pghead &&
4888 ((mp->mp_flags & P_DIRTY) ||
4889 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4893 MDB_ID2 *dl, ix, iy;
4894 rc = mdb_midl_need(&env->me_pghead, ovpages);
4897 if (!(mp->mp_flags & P_DIRTY)) {
4898 /* This page is no longer spilled */
4905 /* Remove from dirty list */
4906 dl = txn->mt_u.dirty_list;
4908 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4916 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4917 txn->mt_flags |= MDB_TXN_ERROR;
4918 return MDB_CORRUPTED;
4921 if (!(env->me_flags & MDB_WRITEMAP))
4922 mdb_dpage_free(env, mp);
4924 /* Insert in me_pghead */
4925 mop = env->me_pghead;
4926 j = mop[0] + ovpages;
4927 for (i = mop[0]; i && mop[i] < pg; i--)
4933 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4937 mc->mc_db->md_overflow_pages -= ovpages;
4941 /** Return the data associated with a given node.
4942 * @param[in] txn The transaction for this operation.
4943 * @param[in] leaf The node being read.
4944 * @param[out] data Updated to point to the node's data.
4945 * @return 0 on success, non-zero on failure.
4948 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4950 MDB_page *omp; /* overflow page */
4954 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4955 data->mv_size = NODEDSZ(leaf);
4956 data->mv_data = NODEDATA(leaf);
4960 /* Read overflow data.
4962 data->mv_size = NODEDSZ(leaf);
4963 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4964 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4965 DPRINTF(("read overflow page %"Z"u failed", pgno));
4968 data->mv_data = METADATA(omp);
4974 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4975 MDB_val *key, MDB_val *data)
4984 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4986 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4989 if (txn->mt_flags & MDB_TXN_ERROR)
4992 mdb_cursor_init(&mc, txn, dbi, &mx);
4993 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4996 /** Find a sibling for a page.
4997 * Replaces the page at the top of the cursor's stack with the
4998 * specified sibling, if one exists.
4999 * @param[in] mc The cursor for this operation.
5000 * @param[in] move_right Non-zero if the right sibling is requested,
5001 * otherwise the left sibling.
5002 * @return 0 on success, non-zero on failure.
5005 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5011 if (mc->mc_snum < 2) {
5012 return MDB_NOTFOUND; /* root has no siblings */
5016 DPRINTF(("parent page is page %"Z"u, index %u",
5017 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5019 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5020 : (mc->mc_ki[mc->mc_top] == 0)) {
5021 DPRINTF(("no more keys left, moving to %s sibling",
5022 move_right ? "right" : "left"));
5023 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5024 /* undo cursor_pop before returning */
5031 mc->mc_ki[mc->mc_top]++;
5033 mc->mc_ki[mc->mc_top]--;
5034 DPRINTF(("just moving to %s index key %u",
5035 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5037 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
5039 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5040 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5041 /* mc will be inconsistent if caller does mc_snum++ as above */
5042 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5046 mdb_cursor_push(mc, mp);
5048 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5053 /** Move the cursor to the next data item. */
5055 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5061 if (mc->mc_flags & C_EOF) {
5062 return MDB_NOTFOUND;
5065 assert(mc->mc_flags & C_INITIALIZED);
5067 mp = mc->mc_pg[mc->mc_top];
5069 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5070 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5071 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5072 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5073 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5074 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5075 if (rc == MDB_SUCCESS)
5076 MDB_GET_KEY(leaf, key);
5081 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5082 if (op == MDB_NEXT_DUP)
5083 return MDB_NOTFOUND;
5087 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5088 if (mc->mc_flags & C_DEL)
5091 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5092 DPUTS("=====> move to next sibling page");
5093 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5094 mc->mc_flags |= C_EOF;
5097 mp = mc->mc_pg[mc->mc_top];
5098 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5100 mc->mc_ki[mc->mc_top]++;
5103 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5104 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5107 key->mv_size = mc->mc_db->md_pad;
5108 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5112 assert(IS_LEAF(mp));
5113 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5115 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5116 mdb_xcursor_init1(mc, leaf);
5119 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5122 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5123 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5124 if (rc != MDB_SUCCESS)
5129 MDB_GET_KEY(leaf, key);
5133 /** Move the cursor to the previous data item. */
5135 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5141 assert(mc->mc_flags & C_INITIALIZED);
5143 mp = mc->mc_pg[mc->mc_top];
5145 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5146 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5147 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5148 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5149 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5150 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5151 if (rc == MDB_SUCCESS)
5152 MDB_GET_KEY(leaf, key);
5156 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5157 if (op == MDB_PREV_DUP)
5158 return MDB_NOTFOUND;
5163 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5165 if (mc->mc_ki[mc->mc_top] == 0) {
5166 DPUTS("=====> move to prev sibling page");
5167 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5170 mp = mc->mc_pg[mc->mc_top];
5171 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5172 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5174 mc->mc_ki[mc->mc_top]--;
5176 mc->mc_flags &= ~C_EOF;
5178 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5179 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5182 key->mv_size = mc->mc_db->md_pad;
5183 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5187 assert(IS_LEAF(mp));
5188 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5190 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5191 mdb_xcursor_init1(mc, leaf);
5194 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5197 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5198 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5199 if (rc != MDB_SUCCESS)
5204 MDB_GET_KEY(leaf, key);
5208 /** Set the cursor on a specific data item. */
5210 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5211 MDB_cursor_op op, int *exactp)
5215 MDB_node *leaf = NULL;
5220 if (key->mv_size == 0)
5221 return MDB_BAD_VALSIZE;
5224 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5226 /* See if we're already on the right page */
5227 if (mc->mc_flags & C_INITIALIZED) {
5230 mp = mc->mc_pg[mc->mc_top];
5232 mc->mc_ki[mc->mc_top] = 0;
5233 return MDB_NOTFOUND;
5235 if (mp->mp_flags & P_LEAF2) {
5236 nodekey.mv_size = mc->mc_db->md_pad;
5237 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5239 leaf = NODEPTR(mp, 0);
5240 MDB_GET_KEY2(leaf, nodekey);
5242 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5244 /* Probably happens rarely, but first node on the page
5245 * was the one we wanted.
5247 mc->mc_ki[mc->mc_top] = 0;
5254 unsigned int nkeys = NUMKEYS(mp);
5256 if (mp->mp_flags & P_LEAF2) {
5257 nodekey.mv_data = LEAF2KEY(mp,
5258 nkeys-1, nodekey.mv_size);
5260 leaf = NODEPTR(mp, nkeys-1);
5261 MDB_GET_KEY2(leaf, nodekey);
5263 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5265 /* last node was the one we wanted */
5266 mc->mc_ki[mc->mc_top] = nkeys-1;
5272 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5273 /* This is definitely the right page, skip search_page */
5274 if (mp->mp_flags & P_LEAF2) {
5275 nodekey.mv_data = LEAF2KEY(mp,
5276 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5278 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5279 MDB_GET_KEY2(leaf, nodekey);
5281 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5283 /* current node was the one we wanted */
5293 /* If any parents have right-sibs, search.
5294 * Otherwise, there's nothing further.
5296 for (i=0; i<mc->mc_top; i++)
5298 NUMKEYS(mc->mc_pg[i])-1)
5300 if (i == mc->mc_top) {
5301 /* There are no other pages */
5302 mc->mc_ki[mc->mc_top] = nkeys;
5303 return MDB_NOTFOUND;
5307 /* There are no other pages */
5308 mc->mc_ki[mc->mc_top] = 0;
5309 if (op == MDB_SET_RANGE) {
5313 return MDB_NOTFOUND;
5317 rc = mdb_page_search(mc, key, 0);
5318 if (rc != MDB_SUCCESS)
5321 mp = mc->mc_pg[mc->mc_top];
5322 assert(IS_LEAF(mp));
5325 leaf = mdb_node_search(mc, key, exactp);
5326 if (exactp != NULL && !*exactp) {
5327 /* MDB_SET specified and not an exact match. */
5328 return MDB_NOTFOUND;
5332 DPUTS("===> inexact leaf not found, goto sibling");
5333 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5334 return rc; /* no entries matched */
5335 mp = mc->mc_pg[mc->mc_top];
5336 assert(IS_LEAF(mp));
5337 leaf = NODEPTR(mp, 0);
5341 mc->mc_flags |= C_INITIALIZED;
5342 mc->mc_flags &= ~C_EOF;
5345 key->mv_size = mc->mc_db->md_pad;
5346 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5350 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5351 mdb_xcursor_init1(mc, leaf);
5354 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5355 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5356 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5359 if (op == MDB_GET_BOTH) {
5365 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5366 if (rc != MDB_SUCCESS)
5369 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5371 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5373 rc = mc->mc_dbx->md_dcmp(data, &d2);
5375 if (op == MDB_GET_BOTH || rc > 0)
5376 return MDB_NOTFOUND;
5383 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5384 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5389 /* The key already matches in all other cases */
5390 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5391 MDB_GET_KEY(leaf, key);
5392 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5397 /** Move the cursor to the first item in the database. */
5399 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5405 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5407 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5408 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5409 if (rc != MDB_SUCCESS)
5412 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5414 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5415 mc->mc_flags |= C_INITIALIZED;
5416 mc->mc_flags &= ~C_EOF;
5418 mc->mc_ki[mc->mc_top] = 0;
5420 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5421 key->mv_size = mc->mc_db->md_pad;
5422 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5427 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5428 mdb_xcursor_init1(mc, leaf);
5429 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5433 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5437 MDB_GET_KEY(leaf, key);
5441 /** Move the cursor to the last item in the database. */
5443 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5449 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5451 if (!(mc->mc_flags & C_EOF)) {
5453 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5454 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5455 if (rc != MDB_SUCCESS)
5458 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5461 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5462 mc->mc_flags |= C_INITIALIZED|C_EOF;
5463 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5465 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5466 key->mv_size = mc->mc_db->md_pad;
5467 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5472 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5473 mdb_xcursor_init1(mc, leaf);
5474 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5478 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5483 MDB_GET_KEY(leaf, key);
5488 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5493 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5497 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5501 case MDB_GET_CURRENT:
5502 if (!(mc->mc_flags & C_INITIALIZED)) {
5505 MDB_page *mp = mc->mc_pg[mc->mc_top];
5506 int nkeys = NUMKEYS(mp);
5507 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5508 mc->mc_ki[mc->mc_top] = nkeys;
5514 key->mv_size = mc->mc_db->md_pad;
5515 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5517 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5518 MDB_GET_KEY(leaf, key);
5520 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5521 if (mc->mc_flags & C_DEL)
5522 mdb_xcursor_init1(mc, leaf);
5523 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5525 rc = mdb_node_read(mc->mc_txn, leaf, data);
5532 case MDB_GET_BOTH_RANGE:
5537 if (mc->mc_xcursor == NULL) {
5538 rc = MDB_INCOMPATIBLE;
5548 rc = mdb_cursor_set(mc, key, data, op,
5549 op == MDB_SET_RANGE ? NULL : &exact);
5552 case MDB_GET_MULTIPLE:
5553 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5557 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5558 rc = MDB_INCOMPATIBLE;
5562 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5563 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5566 case MDB_NEXT_MULTIPLE:
5571 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5572 rc = MDB_INCOMPATIBLE;
5575 if (!(mc->mc_flags & C_INITIALIZED))
5576 rc = mdb_cursor_first(mc, key, data);
5578 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5579 if (rc == MDB_SUCCESS) {
5580 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5583 mx = &mc->mc_xcursor->mx_cursor;
5584 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5586 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5587 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5595 case MDB_NEXT_NODUP:
5596 if (!(mc->mc_flags & C_INITIALIZED))
5597 rc = mdb_cursor_first(mc, key, data);
5599 rc = mdb_cursor_next(mc, key, data, op);
5603 case MDB_PREV_NODUP:
5604 if (!(mc->mc_flags & C_INITIALIZED)) {
5605 rc = mdb_cursor_last(mc, key, data);
5608 mc->mc_flags |= C_INITIALIZED;
5609 mc->mc_ki[mc->mc_top]++;
5611 rc = mdb_cursor_prev(mc, key, data, op);
5614 rc = mdb_cursor_first(mc, key, data);
5617 mfunc = mdb_cursor_first;
5619 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5623 if (mc->mc_xcursor == NULL) {
5624 rc = MDB_INCOMPATIBLE;
5627 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5631 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5634 rc = mdb_cursor_last(mc, key, data);
5637 mfunc = mdb_cursor_last;
5640 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5645 if (mc->mc_flags & C_DEL)
5646 mc->mc_flags ^= C_DEL;
5651 /** Touch all the pages in the cursor stack. Set mc_top.
5652 * Makes sure all the pages are writable, before attempting a write operation.
5653 * @param[in] mc The cursor to operate on.
5656 mdb_cursor_touch(MDB_cursor *mc)
5658 int rc = MDB_SUCCESS;
5660 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5663 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5664 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5667 *mc->mc_dbflag |= DB_DIRTY;
5672 rc = mdb_page_touch(mc);
5673 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5674 mc->mc_top = mc->mc_snum-1;
5679 /** Do not spill pages to disk if txn is getting full, may fail instead */
5680 #define MDB_NOSPILL 0x8000
5683 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5686 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5687 MDB_env *env = mc->mc_txn->mt_env;
5688 MDB_node *leaf = NULL;
5691 MDB_val xdata, *rdata, dkey, olddata;
5693 int do_sub = 0, insert;
5694 unsigned int mcount = 0, dcount = 0, nospill;
5697 unsigned int nflags;
5700 /* Check this first so counter will always be zero on any
5703 if (flags & MDB_MULTIPLE) {
5704 dcount = data[1].mv_size;
5705 data[1].mv_size = 0;
5706 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5707 return MDB_INCOMPATIBLE;
5710 nospill = flags & MDB_NOSPILL;
5711 flags &= ~MDB_NOSPILL;
5713 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5714 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5716 if (flags != MDB_CURRENT && key->mv_size-1 >= ENV_MAXKEY(env))
5717 return MDB_BAD_VALSIZE;
5719 #if SIZE_MAX > MAXDATASIZE
5720 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5721 return MDB_BAD_VALSIZE;
5723 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5724 return MDB_BAD_VALSIZE;
5727 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5728 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5732 if (flags == MDB_CURRENT) {
5733 if (!(mc->mc_flags & C_INITIALIZED))
5736 } else if (mc->mc_db->md_root == P_INVALID) {
5737 /* new database, cursor has nothing to point to */
5740 mc->mc_flags &= ~C_INITIALIZED;
5745 if (flags & MDB_APPEND) {
5747 rc = mdb_cursor_last(mc, &k2, &d2);
5749 rc = mc->mc_dbx->md_cmp(key, &k2);
5752 mc->mc_ki[mc->mc_top]++;
5754 /* new key is <= last key */
5759 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5761 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5762 DPRINTF(("duplicate key [%s]", DKEY(key)));
5764 return MDB_KEYEXIST;
5766 if (rc && rc != MDB_NOTFOUND)
5770 if (mc->mc_flags & C_DEL)
5771 mc->mc_flags ^= C_DEL;
5773 /* Cursor is positioned, check for room in the dirty list */
5775 if (flags & MDB_MULTIPLE) {
5777 xdata.mv_size = data->mv_size * dcount;
5781 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5785 if (rc == MDB_NO_ROOT) {
5787 /* new database, write a root leaf page */
5788 DPUTS("allocating new root leaf page");
5789 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5792 mdb_cursor_push(mc, np);
5793 mc->mc_db->md_root = np->mp_pgno;
5794 mc->mc_db->md_depth++;
5795 *mc->mc_dbflag |= DB_DIRTY;
5796 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5798 np->mp_flags |= P_LEAF2;
5799 mc->mc_flags |= C_INITIALIZED;
5801 /* make sure all cursor pages are writable */
5802 rc2 = mdb_cursor_touch(mc);
5809 /* The key does not exist */
5810 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5811 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5812 LEAFSIZE(key, data) > env->me_nodemax)
5814 /* Too big for a node, insert in sub-DB */
5815 fp_flags = P_LEAF|P_DIRTY;
5817 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5818 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
5822 /* there's only a key anyway, so this is a no-op */
5823 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5824 unsigned int ksize = mc->mc_db->md_pad;
5825 if (key->mv_size != ksize)
5826 return MDB_BAD_VALSIZE;
5827 if (flags == MDB_CURRENT) {
5828 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5829 memcpy(ptr, key->mv_data, ksize);
5835 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5836 olddata.mv_size = NODEDSZ(leaf);
5837 olddata.mv_data = NODEDATA(leaf);
5840 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5841 /* Prepare (sub-)page/sub-DB to accept the new item,
5842 * if needed. fp: old sub-page or a header faking
5843 * it. mp: new (sub-)page. offset: growth in page
5844 * size. xdata: node data with new page or DB.
5846 unsigned i, offset = 0;
5847 mp = fp = xdata.mv_data = env->me_pbuf;
5848 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5850 /* Was a single item before, must convert now */
5851 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5852 /* Just overwrite the current item */
5853 if (flags == MDB_CURRENT)
5856 #if UINT_MAX < SIZE_MAX
5857 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5858 #ifdef MISALIGNED_OK
5859 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5861 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5864 /* if data matches, skip it */
5865 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5866 if (flags & MDB_NODUPDATA)
5868 else if (flags & MDB_MULTIPLE)
5875 /* Back up original data item */
5876 dkey.mv_size = olddata.mv_size;
5877 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5879 /* Make sub-page header for the dup items, with dummy body */
5880 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5881 fp->mp_lower = PAGEHDRSZ;
5882 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5883 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5884 fp->mp_flags |= P_LEAF2;
5885 fp->mp_pad = data->mv_size;
5886 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
5888 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
5889 (dkey.mv_size & 1) + (data->mv_size & 1);
5891 fp->mp_upper = xdata.mv_size;
5892 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
5893 } else if (leaf->mn_flags & F_SUBDATA) {
5894 /* Data is on sub-DB, just store it */
5895 flags |= F_DUPDATA|F_SUBDATA;
5898 /* Data is on sub-page */
5899 fp = olddata.mv_data;
5902 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5903 offset = EVEN(NODESIZE + sizeof(indx_t) +
5907 offset = fp->mp_pad;
5908 if (SIZELEFT(fp) < offset) {
5909 offset *= 4; /* space for 4 more */
5912 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
5914 fp->mp_flags |= P_DIRTY;
5915 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
5916 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5920 xdata.mv_size = olddata.mv_size + offset;
5923 fp_flags = fp->mp_flags;
5924 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
5925 /* Too big for a sub-page, convert to sub-DB */
5926 fp_flags &= ~P_SUBP;
5928 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5929 fp_flags |= P_LEAF2;
5930 dummy.md_pad = fp->mp_pad;
5931 dummy.md_flags = MDB_DUPFIXED;
5932 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5933 dummy.md_flags |= MDB_INTEGERKEY;
5939 dummy.md_branch_pages = 0;
5940 dummy.md_leaf_pages = 1;
5941 dummy.md_overflow_pages = 0;
5942 dummy.md_entries = NUMKEYS(fp);
5943 xdata.mv_size = sizeof(MDB_db);
5944 xdata.mv_data = &dummy;
5945 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5947 offset = env->me_psize - olddata.mv_size;
5948 flags |= F_DUPDATA|F_SUBDATA;
5949 dummy.md_root = mp->mp_pgno;
5952 mp->mp_flags = fp_flags | P_DIRTY;
5953 mp->mp_pad = fp->mp_pad;
5954 mp->mp_lower = fp->mp_lower;
5955 mp->mp_upper = fp->mp_upper + offset;
5956 if (fp_flags & P_LEAF2) {
5957 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5959 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
5960 olddata.mv_size - fp->mp_upper);
5961 for (i=0; i<NUMKEYS(fp); i++)
5962 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5969 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5973 /* overflow page overwrites need special handling */
5974 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5977 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
5979 memcpy(&pg, olddata.mv_data, sizeof(pg));
5980 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5982 ovpages = omp->mp_pages;
5984 /* Is the ov page large enough? */
5985 if (ovpages >= dpages) {
5986 if (!(omp->mp_flags & P_DIRTY) &&
5987 (level || (env->me_flags & MDB_WRITEMAP)))
5989 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5992 level = 0; /* dirty in this txn or clean */
5995 if (omp->mp_flags & P_DIRTY) {
5996 /* yes, overwrite it. Note in this case we don't
5997 * bother to try shrinking the page if the new data
5998 * is smaller than the overflow threshold.
6001 /* It is writable only in a parent txn */
6002 size_t sz = (size_t) env->me_psize * ovpages, off;
6003 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6009 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6010 if (!(flags & MDB_RESERVE)) {
6011 /* Copy end of page, adjusting alignment so
6012 * compiler may copy words instead of bytes.
6014 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6015 memcpy((size_t *)((char *)np + off),
6016 (size_t *)((char *)omp + off), sz - off);
6019 memcpy(np, omp, sz); /* Copy beginning of page */
6022 SETDSZ(leaf, data->mv_size);
6023 if (F_ISSET(flags, MDB_RESERVE))
6024 data->mv_data = METADATA(omp);
6026 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6030 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6032 } else if (data->mv_size == olddata.mv_size) {
6033 /* same size, just replace it. Note that we could
6034 * also reuse this node if the new data is smaller,
6035 * but instead we opt to shrink the node in that case.
6037 if (F_ISSET(flags, MDB_RESERVE))
6038 data->mv_data = olddata.mv_data;
6039 else if (data->mv_size)
6040 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6042 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6045 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6046 mc->mc_db->md_entries--;
6052 nflags = flags & NODE_ADD_FLAGS;
6053 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6054 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6055 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6056 nflags &= ~MDB_APPEND;
6058 nflags |= MDB_SPLIT_REPLACE;
6059 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6061 /* There is room already in this leaf page. */
6062 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6063 if (rc == 0 && !do_sub && insert) {
6064 /* Adjust other cursors pointing to mp */
6065 MDB_cursor *m2, *m3;
6066 MDB_dbi dbi = mc->mc_dbi;
6067 unsigned i = mc->mc_top;
6068 MDB_page *mp = mc->mc_pg[i];
6070 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6071 if (mc->mc_flags & C_SUB)
6072 m3 = &m2->mc_xcursor->mx_cursor;
6075 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6076 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6083 if (rc != MDB_SUCCESS)
6084 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6086 /* Now store the actual data in the child DB. Note that we're
6087 * storing the user data in the keys field, so there are strict
6088 * size limits on dupdata. The actual data fields of the child
6089 * DB are all zero size.
6096 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6097 if (flags & MDB_CURRENT) {
6098 xflags = MDB_CURRENT|MDB_NOSPILL;
6100 mdb_xcursor_init1(mc, leaf);
6101 xflags = (flags & MDB_NODUPDATA) ?
6102 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6104 /* converted, write the original data first */
6106 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6110 /* Adjust other cursors pointing to mp */
6112 unsigned i = mc->mc_top;
6113 MDB_page *mp = mc->mc_pg[i];
6115 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6116 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6117 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6118 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6119 mdb_xcursor_init1(m2, leaf);
6123 /* we've done our job */
6126 if (flags & MDB_APPENDDUP)
6127 xflags |= MDB_APPEND;
6128 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6129 if (flags & F_SUBDATA) {
6130 void *db = NODEDATA(leaf);
6131 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6134 /* sub-writes might have failed so check rc again.
6135 * Don't increment count if we just replaced an existing item.
6137 if (!rc && !(flags & MDB_CURRENT))
6138 mc->mc_db->md_entries++;
6139 if (flags & MDB_MULTIPLE) {
6143 /* let caller know how many succeeded, if any */
6144 data[1].mv_size = mcount;
6145 if (mcount < dcount) {
6146 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6153 /* If we succeeded and the key didn't exist before, make sure
6154 * the cursor is marked valid.
6157 mc->mc_flags |= C_INITIALIZED;
6162 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6168 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6169 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6171 if (!(mc->mc_flags & C_INITIALIZED))
6174 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6175 return MDB_NOTFOUND;
6177 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6180 rc = mdb_cursor_touch(mc);
6184 mp = mc->mc_pg[mc->mc_top];
6185 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6187 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6188 if (!(flags & MDB_NODUPDATA)) {
6189 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6190 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6192 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6193 /* If sub-DB still has entries, we're done */
6194 if (mc->mc_xcursor->mx_db.md_entries) {
6195 if (leaf->mn_flags & F_SUBDATA) {
6196 /* update subDB info */
6197 void *db = NODEDATA(leaf);
6198 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6201 /* shrink fake page */
6202 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6203 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6204 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6205 /* fix other sub-DB cursors pointed at this fake page */
6206 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6207 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6208 if (m2->mc_pg[mc->mc_top] == mp &&
6209 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6210 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6213 mc->mc_db->md_entries--;
6214 mc->mc_flags |= C_DEL;
6217 /* otherwise fall thru and delete the sub-DB */
6220 if (leaf->mn_flags & F_SUBDATA) {
6221 /* add all the child DB's pages to the free list */
6222 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6223 if (rc == MDB_SUCCESS) {
6224 mc->mc_db->md_entries -=
6225 mc->mc_xcursor->mx_db.md_entries;
6230 return mdb_cursor_del0(mc, leaf);
6233 /** Allocate and initialize new pages for a database.
6234 * @param[in] mc a cursor on the database being added to.
6235 * @param[in] flags flags defining what type of page is being allocated.
6236 * @param[in] num the number of pages to allocate. This is usually 1,
6237 * unless allocating overflow pages for a large record.
6238 * @param[out] mp Address of a page, or NULL on failure.
6239 * @return 0 on success, non-zero on failure.
6242 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6247 if ((rc = mdb_page_alloc(mc, num, &np)))
6249 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6250 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6251 np->mp_flags = flags | P_DIRTY;
6252 np->mp_lower = PAGEHDRSZ;
6253 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6256 mc->mc_db->md_branch_pages++;
6257 else if (IS_LEAF(np))
6258 mc->mc_db->md_leaf_pages++;
6259 else if (IS_OVERFLOW(np)) {
6260 mc->mc_db->md_overflow_pages += num;
6268 /** Calculate the size of a leaf node.
6269 * The size depends on the environment's page size; if a data item
6270 * is too large it will be put onto an overflow page and the node
6271 * size will only include the key and not the data. Sizes are always
6272 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6273 * of the #MDB_node headers.
6274 * @param[in] env The environment handle.
6275 * @param[in] key The key for the node.
6276 * @param[in] data The data for the node.
6277 * @return The number of bytes needed to store the node.
6280 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6284 sz = LEAFSIZE(key, data);
6285 if (sz > env->me_nodemax) {
6286 /* put on overflow page */
6287 sz -= data->mv_size - sizeof(pgno_t);
6290 return EVEN(sz + sizeof(indx_t));
6293 /** Calculate the size of a branch node.
6294 * The size should depend on the environment's page size but since
6295 * we currently don't support spilling large keys onto overflow
6296 * pages, it's simply the size of the #MDB_node header plus the
6297 * size of the key. Sizes are always rounded up to an even number
6298 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6299 * @param[in] env The environment handle.
6300 * @param[in] key The key for the node.
6301 * @return The number of bytes needed to store the node.
6304 mdb_branch_size(MDB_env *env, MDB_val *key)
6309 if (sz > env->me_nodemax) {
6310 /* put on overflow page */
6311 /* not implemented */
6312 /* sz -= key->size - sizeof(pgno_t); */
6315 return sz + sizeof(indx_t);
6318 /** Add a node to the page pointed to by the cursor.
6319 * @param[in] mc The cursor for this operation.
6320 * @param[in] indx The index on the page where the new node should be added.
6321 * @param[in] key The key for the new node.
6322 * @param[in] data The data for the new node, if any.
6323 * @param[in] pgno The page number, if adding a branch node.
6324 * @param[in] flags Flags for the node.
6325 * @return 0 on success, non-zero on failure. Possible errors are:
6327 * <li>ENOMEM - failed to allocate overflow pages for the node.
6328 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6329 * should never happen since all callers already calculate the
6330 * page's free space before calling this function.
6334 mdb_node_add(MDB_cursor *mc, indx_t indx,
6335 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6338 size_t node_size = NODESIZE;
6342 MDB_page *mp = mc->mc_pg[mc->mc_top];
6343 MDB_page *ofp = NULL; /* overflow page */
6346 assert(mp->mp_upper >= mp->mp_lower);
6348 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6349 IS_LEAF(mp) ? "leaf" : "branch",
6350 IS_SUBP(mp) ? "sub-" : "",
6351 mp->mp_pgno, indx, data ? data->mv_size : 0,
6352 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6355 /* Move higher keys up one slot. */
6356 int ksize = mc->mc_db->md_pad, dif;
6357 char *ptr = LEAF2KEY(mp, indx, ksize);
6358 dif = NUMKEYS(mp) - indx;
6360 memmove(ptr+ksize, ptr, dif*ksize);
6361 /* insert new key */
6362 memcpy(ptr, key->mv_data, ksize);
6364 /* Just using these for counting */
6365 mp->mp_lower += sizeof(indx_t);
6366 mp->mp_upper -= ksize - sizeof(indx_t);
6370 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6372 node_size += key->mv_size;
6375 if (F_ISSET(flags, F_BIGDATA)) {
6376 /* Data already on overflow page. */
6377 node_size += sizeof(pgno_t);
6378 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6379 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6381 /* Put data on overflow page. */
6382 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6383 data->mv_size, node_size+data->mv_size));
6384 node_size = EVEN(node_size + sizeof(pgno_t));
6385 if ((ssize_t)node_size > room)
6387 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6389 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6393 node_size += data->mv_size;
6396 node_size = EVEN(node_size);
6397 if ((ssize_t)node_size > room)
6401 /* Move higher pointers up one slot. */
6402 for (i = NUMKEYS(mp); i > indx; i--)
6403 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6405 /* Adjust free space offsets. */
6406 ofs = mp->mp_upper - node_size;
6407 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6408 mp->mp_ptrs[indx] = ofs;
6410 mp->mp_lower += sizeof(indx_t);
6412 /* Write the node data. */
6413 node = NODEPTR(mp, indx);
6414 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6415 node->mn_flags = flags;
6417 SETDSZ(node,data->mv_size);
6422 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6427 if (F_ISSET(flags, F_BIGDATA))
6428 memcpy(node->mn_data + key->mv_size, data->mv_data,
6430 else if (F_ISSET(flags, MDB_RESERVE))
6431 data->mv_data = node->mn_data + key->mv_size;
6433 memcpy(node->mn_data + key->mv_size, data->mv_data,
6436 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6438 if (F_ISSET(flags, MDB_RESERVE))
6439 data->mv_data = METADATA(ofp);
6441 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6448 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6449 mp->mp_pgno, NUMKEYS(mp)));
6450 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6451 DPRINTF(("node size = %"Z"u", node_size));
6452 return MDB_PAGE_FULL;
6455 /** Delete the specified node from a page.
6456 * @param[in] mp The page to operate on.
6457 * @param[in] indx The index of the node to delete.
6458 * @param[in] ksize The size of a node. Only used if the page is
6459 * part of a #MDB_DUPFIXED database.
6462 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6465 indx_t i, j, numkeys, ptr;
6472 COPY_PGNO(pgno, mp->mp_pgno);
6473 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6474 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6477 assert(indx < NUMKEYS(mp));
6480 int x = NUMKEYS(mp) - 1 - indx;
6481 base = LEAF2KEY(mp, indx, ksize);
6483 memmove(base, base + ksize, x * ksize);
6484 mp->mp_lower -= sizeof(indx_t);
6485 mp->mp_upper += ksize - sizeof(indx_t);
6489 node = NODEPTR(mp, indx);
6490 sz = NODESIZE + node->mn_ksize;
6492 if (F_ISSET(node->mn_flags, F_BIGDATA))
6493 sz += sizeof(pgno_t);
6495 sz += NODEDSZ(node);
6499 ptr = mp->mp_ptrs[indx];
6500 numkeys = NUMKEYS(mp);
6501 for (i = j = 0; i < numkeys; i++) {
6503 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6504 if (mp->mp_ptrs[i] < ptr)
6505 mp->mp_ptrs[j] += sz;
6510 base = (char *)mp + mp->mp_upper;
6511 memmove(base + sz, base, ptr - mp->mp_upper);
6513 mp->mp_lower -= sizeof(indx_t);
6517 /** Compact the main page after deleting a node on a subpage.
6518 * @param[in] mp The main page to operate on.
6519 * @param[in] indx The index of the subpage on the main page.
6522 mdb_node_shrink(MDB_page *mp, indx_t indx)
6528 indx_t i, numkeys, ptr;
6530 node = NODEPTR(mp, indx);
6531 sp = (MDB_page *)NODEDATA(node);
6532 delta = SIZELEFT(sp);
6533 xp = (MDB_page *)((char *)sp + delta);
6535 /* shift subpage upward */
6537 nsize = NUMKEYS(sp) * sp->mp_pad;
6539 return; /* do not make the node uneven-sized */
6540 memmove(METADATA(xp), METADATA(sp), nsize);
6543 numkeys = NUMKEYS(sp);
6544 for (i=numkeys-1; i>=0; i--)
6545 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6547 xp->mp_upper = sp->mp_lower;
6548 xp->mp_lower = sp->mp_lower;
6549 xp->mp_flags = sp->mp_flags;
6550 xp->mp_pad = sp->mp_pad;
6551 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6553 nsize = NODEDSZ(node) - delta;
6554 SETDSZ(node, nsize);
6556 /* shift lower nodes upward */
6557 ptr = mp->mp_ptrs[indx];
6558 numkeys = NUMKEYS(mp);
6559 for (i = 0; i < numkeys; i++) {
6560 if (mp->mp_ptrs[i] <= ptr)
6561 mp->mp_ptrs[i] += delta;
6564 base = (char *)mp + mp->mp_upper;
6565 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6566 mp->mp_upper += delta;
6569 /** Initial setup of a sorted-dups cursor.
6570 * Sorted duplicates are implemented as a sub-database for the given key.
6571 * The duplicate data items are actually keys of the sub-database.
6572 * Operations on the duplicate data items are performed using a sub-cursor
6573 * initialized when the sub-database is first accessed. This function does
6574 * the preliminary setup of the sub-cursor, filling in the fields that
6575 * depend only on the parent DB.
6576 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6579 mdb_xcursor_init0(MDB_cursor *mc)
6581 MDB_xcursor *mx = mc->mc_xcursor;
6583 mx->mx_cursor.mc_xcursor = NULL;
6584 mx->mx_cursor.mc_txn = mc->mc_txn;
6585 mx->mx_cursor.mc_db = &mx->mx_db;
6586 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6587 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6588 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6589 mx->mx_cursor.mc_snum = 0;
6590 mx->mx_cursor.mc_top = 0;
6591 mx->mx_cursor.mc_flags = C_SUB;
6592 mx->mx_dbx.md_name.mv_size = 0;
6593 mx->mx_dbx.md_name.mv_data = NULL;
6594 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6595 mx->mx_dbx.md_dcmp = NULL;
6596 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6599 /** Final setup of a sorted-dups cursor.
6600 * Sets up the fields that depend on the data from the main cursor.
6601 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6602 * @param[in] node The data containing the #MDB_db record for the
6603 * sorted-dup database.
6606 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6608 MDB_xcursor *mx = mc->mc_xcursor;
6610 if (node->mn_flags & F_SUBDATA) {
6611 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6612 mx->mx_cursor.mc_pg[0] = 0;
6613 mx->mx_cursor.mc_snum = 0;
6614 mx->mx_cursor.mc_top = 0;
6615 mx->mx_cursor.mc_flags = C_SUB;
6617 MDB_page *fp = NODEDATA(node);
6618 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6619 mx->mx_db.md_flags = 0;
6620 mx->mx_db.md_depth = 1;
6621 mx->mx_db.md_branch_pages = 0;
6622 mx->mx_db.md_leaf_pages = 1;
6623 mx->mx_db.md_overflow_pages = 0;
6624 mx->mx_db.md_entries = NUMKEYS(fp);
6625 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6626 mx->mx_cursor.mc_snum = 1;
6627 mx->mx_cursor.mc_top = 0;
6628 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6629 mx->mx_cursor.mc_pg[0] = fp;
6630 mx->mx_cursor.mc_ki[0] = 0;
6631 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6632 mx->mx_db.md_flags = MDB_DUPFIXED;
6633 mx->mx_db.md_pad = fp->mp_pad;
6634 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6635 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6638 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6639 mx->mx_db.md_root));
6640 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6641 #if UINT_MAX < SIZE_MAX
6642 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6643 #ifdef MISALIGNED_OK
6644 mx->mx_dbx.md_cmp = mdb_cmp_long;
6646 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6651 /** Initialize a cursor for a given transaction and database. */
6653 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6656 mc->mc_backup = NULL;
6659 mc->mc_db = &txn->mt_dbs[dbi];
6660 mc->mc_dbx = &txn->mt_dbxs[dbi];
6661 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6666 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6668 mc->mc_xcursor = mx;
6669 mdb_xcursor_init0(mc);
6671 mc->mc_xcursor = NULL;
6673 if (*mc->mc_dbflag & DB_STALE) {
6674 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6679 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6682 size_t size = sizeof(MDB_cursor);
6684 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6687 if (txn->mt_flags & MDB_TXN_ERROR)
6690 /* Allow read access to the freelist */
6691 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6694 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6695 size += sizeof(MDB_xcursor);
6697 if ((mc = malloc(size)) != NULL) {
6698 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6699 if (txn->mt_cursors) {
6700 mc->mc_next = txn->mt_cursors[dbi];
6701 txn->mt_cursors[dbi] = mc;
6702 mc->mc_flags |= C_UNTRACK;
6714 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6716 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6719 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6722 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6726 /* Return the count of duplicate data items for the current key */
6728 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6732 if (mc == NULL || countp == NULL)
6735 if (mc->mc_xcursor == NULL)
6736 return MDB_INCOMPATIBLE;
6738 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6739 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6742 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6745 *countp = mc->mc_xcursor->mx_db.md_entries;
6751 mdb_cursor_close(MDB_cursor *mc)
6753 if (mc && !mc->mc_backup) {
6754 /* remove from txn, if tracked */
6755 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6756 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6757 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6759 *prev = mc->mc_next;
6766 mdb_cursor_txn(MDB_cursor *mc)
6768 if (!mc) return NULL;
6773 mdb_cursor_dbi(MDB_cursor *mc)
6779 /** Replace the key for a branch node with a new key.
6780 * @param[in] mc Cursor pointing to the node to operate on.
6781 * @param[in] key The new key to use.
6782 * @return 0 on success, non-zero on failure.
6785 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6791 int delta, ksize, oksize;
6792 indx_t ptr, i, numkeys, indx;
6795 indx = mc->mc_ki[mc->mc_top];
6796 mp = mc->mc_pg[mc->mc_top];
6797 node = NODEPTR(mp, indx);
6798 ptr = mp->mp_ptrs[indx];
6802 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6803 k2.mv_data = NODEKEY(node);
6804 k2.mv_size = node->mn_ksize;
6805 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6807 mdb_dkey(&k2, kbuf2),
6813 /* Sizes must be 2-byte aligned. */
6814 ksize = EVEN(key->mv_size);
6815 oksize = EVEN(node->mn_ksize);
6816 delta = ksize - oksize;
6818 /* Shift node contents if EVEN(key length) changed. */
6820 if (delta > 0 && SIZELEFT(mp) < delta) {
6822 /* not enough space left, do a delete and split */
6823 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6824 pgno = NODEPGNO(node);
6825 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6826 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6829 numkeys = NUMKEYS(mp);
6830 for (i = 0; i < numkeys; i++) {
6831 if (mp->mp_ptrs[i] <= ptr)
6832 mp->mp_ptrs[i] -= delta;
6835 base = (char *)mp + mp->mp_upper;
6836 len = ptr - mp->mp_upper + NODESIZE;
6837 memmove(base - delta, base, len);
6838 mp->mp_upper -= delta;
6840 node = NODEPTR(mp, indx);
6843 /* But even if no shift was needed, update ksize */
6844 if (node->mn_ksize != key->mv_size)
6845 node->mn_ksize = key->mv_size;
6848 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6854 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6856 /** Move a node from csrc to cdst.
6859 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6866 unsigned short flags;
6870 /* Mark src and dst as dirty. */
6871 if ((rc = mdb_page_touch(csrc)) ||
6872 (rc = mdb_page_touch(cdst)))
6875 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6876 key.mv_size = csrc->mc_db->md_pad;
6877 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6879 data.mv_data = NULL;
6883 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6884 assert(!((size_t)srcnode&1));
6885 srcpg = NODEPGNO(srcnode);
6886 flags = srcnode->mn_flags;
6887 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6888 unsigned int snum = csrc->mc_snum;
6890 /* must find the lowest key below src */
6891 mdb_page_search_lowest(csrc);
6892 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6893 key.mv_size = csrc->mc_db->md_pad;
6894 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6896 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6897 key.mv_size = NODEKSZ(s2);
6898 key.mv_data = NODEKEY(s2);
6900 csrc->mc_snum = snum--;
6901 csrc->mc_top = snum;
6903 key.mv_size = NODEKSZ(srcnode);
6904 key.mv_data = NODEKEY(srcnode);
6906 data.mv_size = NODEDSZ(srcnode);
6907 data.mv_data = NODEDATA(srcnode);
6909 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6910 unsigned int snum = cdst->mc_snum;
6913 /* must find the lowest key below dst */
6914 mdb_page_search_lowest(cdst);
6915 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6916 bkey.mv_size = cdst->mc_db->md_pad;
6917 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6919 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6920 bkey.mv_size = NODEKSZ(s2);
6921 bkey.mv_data = NODEKEY(s2);
6923 cdst->mc_snum = snum--;
6924 cdst->mc_top = snum;
6925 mdb_cursor_copy(cdst, &mn);
6927 rc = mdb_update_key(&mn, &bkey);
6932 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6933 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6934 csrc->mc_ki[csrc->mc_top],
6936 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6937 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6939 /* Add the node to the destination page.
6941 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6942 if (rc != MDB_SUCCESS)
6945 /* Delete the node from the source page.
6947 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6950 /* Adjust other cursors pointing to mp */
6951 MDB_cursor *m2, *m3;
6952 MDB_dbi dbi = csrc->mc_dbi;
6953 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6955 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6956 if (csrc->mc_flags & C_SUB)
6957 m3 = &m2->mc_xcursor->mx_cursor;
6960 if (m3 == csrc) continue;
6961 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6962 csrc->mc_ki[csrc->mc_top]) {
6963 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6964 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6969 /* Update the parent separators.
6971 if (csrc->mc_ki[csrc->mc_top] == 0) {
6972 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6973 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6974 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6976 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6977 key.mv_size = NODEKSZ(srcnode);
6978 key.mv_data = NODEKEY(srcnode);
6980 DPRINTF(("update separator for source page %"Z"u to [%s]",
6981 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6982 mdb_cursor_copy(csrc, &mn);
6985 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6988 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6990 indx_t ix = csrc->mc_ki[csrc->mc_top];
6991 nullkey.mv_size = 0;
6992 csrc->mc_ki[csrc->mc_top] = 0;
6993 rc = mdb_update_key(csrc, &nullkey);
6994 csrc->mc_ki[csrc->mc_top] = ix;
6995 assert(rc == MDB_SUCCESS);
6999 if (cdst->mc_ki[cdst->mc_top] == 0) {
7000 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7001 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7002 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7004 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7005 key.mv_size = NODEKSZ(srcnode);
7006 key.mv_data = NODEKEY(srcnode);
7008 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7009 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7010 mdb_cursor_copy(cdst, &mn);
7013 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7016 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7018 indx_t ix = cdst->mc_ki[cdst->mc_top];
7019 nullkey.mv_size = 0;
7020 cdst->mc_ki[cdst->mc_top] = 0;
7021 rc = mdb_update_key(cdst, &nullkey);
7022 cdst->mc_ki[cdst->mc_top] = ix;
7023 assert(rc == MDB_SUCCESS);
7030 /** Merge one page into another.
7031 * The nodes from the page pointed to by \b csrc will
7032 * be copied to the page pointed to by \b cdst and then
7033 * the \b csrc page will be freed.
7034 * @param[in] csrc Cursor pointing to the source page.
7035 * @param[in] cdst Cursor pointing to the destination page.
7038 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7046 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
7047 cdst->mc_pg[cdst->mc_top]->mp_pgno));
7049 assert(csrc->mc_snum > 1); /* can't merge root page */
7050 assert(cdst->mc_snum > 1);
7052 /* Mark dst as dirty. */
7053 if ((rc = mdb_page_touch(cdst)))
7056 /* Move all nodes from src to dst.
7058 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
7059 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7060 key.mv_size = csrc->mc_db->md_pad;
7061 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
7062 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7063 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7064 if (rc != MDB_SUCCESS)
7066 key.mv_data = (char *)key.mv_data + key.mv_size;
7069 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7070 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7071 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7072 unsigned int snum = csrc->mc_snum;
7074 /* must find the lowest key below src */
7075 mdb_page_search_lowest(csrc);
7076 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7077 key.mv_size = csrc->mc_db->md_pad;
7078 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7080 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7081 key.mv_size = NODEKSZ(s2);
7082 key.mv_data = NODEKEY(s2);
7084 csrc->mc_snum = snum--;
7085 csrc->mc_top = snum;
7087 key.mv_size = srcnode->mn_ksize;
7088 key.mv_data = NODEKEY(srcnode);
7091 data.mv_size = NODEDSZ(srcnode);
7092 data.mv_data = NODEDATA(srcnode);
7093 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7094 if (rc != MDB_SUCCESS)
7099 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7100 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7101 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7103 /* Unlink the src page from parent and add to free list.
7105 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
7106 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
7109 rc = mdb_update_key(csrc, &key);
7115 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7116 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7119 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7120 csrc->mc_db->md_leaf_pages--;
7122 csrc->mc_db->md_branch_pages--;
7124 /* Adjust other cursors pointing to mp */
7125 MDB_cursor *m2, *m3;
7126 MDB_dbi dbi = csrc->mc_dbi;
7127 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7129 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7130 if (csrc->mc_flags & C_SUB)
7131 m3 = &m2->mc_xcursor->mx_cursor;
7134 if (m3 == csrc) continue;
7135 if (m3->mc_snum < csrc->mc_snum) continue;
7136 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7137 m3->mc_pg[csrc->mc_top] = mp;
7138 m3->mc_ki[csrc->mc_top] += nkeys;
7142 mdb_cursor_pop(csrc);
7144 return mdb_rebalance(csrc);
7147 /** Copy the contents of a cursor.
7148 * @param[in] csrc The cursor to copy from.
7149 * @param[out] cdst The cursor to copy to.
7152 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7156 cdst->mc_txn = csrc->mc_txn;
7157 cdst->mc_dbi = csrc->mc_dbi;
7158 cdst->mc_db = csrc->mc_db;
7159 cdst->mc_dbx = csrc->mc_dbx;
7160 cdst->mc_snum = csrc->mc_snum;
7161 cdst->mc_top = csrc->mc_top;
7162 cdst->mc_flags = csrc->mc_flags;
7164 for (i=0; i<csrc->mc_snum; i++) {
7165 cdst->mc_pg[i] = csrc->mc_pg[i];
7166 cdst->mc_ki[i] = csrc->mc_ki[i];
7170 /** Rebalance the tree after a delete operation.
7171 * @param[in] mc Cursor pointing to the page where rebalancing
7173 * @return 0 on success, non-zero on failure.
7176 mdb_rebalance(MDB_cursor *mc)
7180 unsigned int ptop, minkeys;
7183 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7187 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7188 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7189 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7190 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7191 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7195 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7196 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7199 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7200 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7206 if (mc->mc_snum < 2) {
7207 MDB_page *mp = mc->mc_pg[0];
7209 DPUTS("Can't rebalance a subpage, ignoring");
7212 if (NUMKEYS(mp) == 0) {
7213 DPUTS("tree is completely empty");
7214 mc->mc_db->md_root = P_INVALID;
7215 mc->mc_db->md_depth = 0;
7216 mc->mc_db->md_leaf_pages = 0;
7217 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7220 /* Adjust cursors pointing to mp */
7223 mc->mc_flags &= ~C_INITIALIZED;
7225 MDB_cursor *m2, *m3;
7226 MDB_dbi dbi = mc->mc_dbi;
7228 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7229 if (mc->mc_flags & C_SUB)
7230 m3 = &m2->mc_xcursor->mx_cursor;
7233 if (m3->mc_snum < mc->mc_snum) continue;
7234 if (m3->mc_pg[0] == mp) {
7237 m3->mc_flags &= ~C_INITIALIZED;
7241 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7242 DPUTS("collapsing root page!");
7243 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7246 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7247 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7250 mc->mc_db->md_depth--;
7251 mc->mc_db->md_branch_pages--;
7252 mc->mc_ki[0] = mc->mc_ki[1];
7254 /* Adjust other cursors pointing to mp */
7255 MDB_cursor *m2, *m3;
7256 MDB_dbi dbi = mc->mc_dbi;
7258 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7259 if (mc->mc_flags & C_SUB)
7260 m3 = &m2->mc_xcursor->mx_cursor;
7263 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7264 if (m3->mc_pg[0] == mp) {
7268 for (i=0; i<m3->mc_snum; i++) {
7269 m3->mc_pg[i] = m3->mc_pg[i+1];
7270 m3->mc_ki[i] = m3->mc_ki[i+1];
7276 DPUTS("root page doesn't need rebalancing");
7280 /* The parent (branch page) must have at least 2 pointers,
7281 * otherwise the tree is invalid.
7283 ptop = mc->mc_top-1;
7284 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7286 /* Leaf page fill factor is below the threshold.
7287 * Try to move keys from left or right neighbor, or
7288 * merge with a neighbor page.
7293 mdb_cursor_copy(mc, &mn);
7294 mn.mc_xcursor = NULL;
7296 if (mc->mc_ki[ptop] == 0) {
7297 /* We're the leftmost leaf in our parent.
7299 DPUTS("reading right neighbor");
7301 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7302 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7305 mn.mc_ki[mn.mc_top] = 0;
7306 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7308 /* There is at least one neighbor to the left.
7310 DPUTS("reading left neighbor");
7312 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7313 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7316 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7317 mc->mc_ki[mc->mc_top] = 0;
7320 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7321 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7322 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7324 /* If the neighbor page is above threshold and has enough keys,
7325 * move one key from it. Otherwise we should try to merge them.
7326 * (A branch page must never have less than 2 keys.)
7328 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7329 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7330 return mdb_node_move(&mn, mc);
7332 if (mc->mc_ki[ptop] == 0)
7333 rc = mdb_page_merge(&mn, mc);
7335 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7336 rc = mdb_page_merge(mc, &mn);
7337 mdb_cursor_copy(&mn, mc);
7339 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7344 /** Complete a delete operation started by #mdb_cursor_del(). */
7346 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7353 mp = mc->mc_pg[mc->mc_top];
7354 ki = mc->mc_ki[mc->mc_top];
7356 /* add overflow pages to free list */
7357 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7361 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7362 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7363 (rc = mdb_ovpage_free(mc, omp)))
7366 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7367 mc->mc_db->md_entries--;
7368 rc = mdb_rebalance(mc);
7369 if (rc != MDB_SUCCESS)
7370 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7373 MDB_dbi dbi = mc->mc_dbi;
7375 mp = mc->mc_pg[mc->mc_top];
7376 nkeys = NUMKEYS(mp);
7378 /* if mc points past last node in page, find next sibling */
7379 if (mc->mc_ki[mc->mc_top] >= nkeys)
7380 mdb_cursor_sibling(mc, 1);
7382 /* Adjust other cursors pointing to mp */
7383 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7384 if (m2 == mc || m2->mc_snum < mc->mc_snum)
7386 if (!(m2->mc_flags & C_INITIALIZED))
7388 if (m2->mc_pg[mc->mc_top] == mp) {
7389 if (m2->mc_ki[mc->mc_top] >= ki) {
7390 m2->mc_flags |= C_DEL;
7391 if (m2->mc_ki[mc->mc_top] > ki)
7392 m2->mc_ki[mc->mc_top]--;
7394 if (m2->mc_ki[mc->mc_top] >= nkeys)
7395 mdb_cursor_sibling(m2, 1);
7398 mc->mc_flags |= C_DEL;
7405 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7406 MDB_val *key, MDB_val *data)
7411 MDB_val rdata, *xdata;
7415 assert(key != NULL);
7417 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7419 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7422 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7423 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7425 mdb_cursor_init(&mc, txn, dbi, &mx);
7428 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7429 /* must ignore any data */
7440 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7442 /* let mdb_page_split know about this cursor if needed:
7443 * delete will trigger a rebalance; if it needs to move
7444 * a node from one page to another, it will have to
7445 * update the parent's separator key(s). If the new sepkey
7446 * is larger than the current one, the parent page may
7447 * run out of space, triggering a split. We need this
7448 * cursor to be consistent until the end of the rebalance.
7450 mc.mc_flags |= C_UNTRACK;
7451 mc.mc_next = txn->mt_cursors[dbi];
7452 txn->mt_cursors[dbi] = &mc;
7453 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7454 txn->mt_cursors[dbi] = mc.mc_next;
7459 /** Split a page and insert a new node.
7460 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7461 * The cursor will be updated to point to the actual page and index where
7462 * the node got inserted after the split.
7463 * @param[in] newkey The key for the newly inserted node.
7464 * @param[in] newdata The data for the newly inserted node.
7465 * @param[in] newpgno The page number, if the new node is a branch node.
7466 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7467 * @return 0 on success, non-zero on failure.
7470 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7471 unsigned int nflags)
7474 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7477 int i, j, split_indx, nkeys, pmax;
7478 MDB_env *env = mc->mc_txn->mt_env;
7480 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7481 MDB_page *copy = NULL;
7482 MDB_page *mp, *rp, *pp;
7487 mp = mc->mc_pg[mc->mc_top];
7488 newindx = mc->mc_ki[mc->mc_top];
7489 nkeys = NUMKEYS(mp);
7491 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7492 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7493 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7495 /* Create a right sibling. */
7496 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7498 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7500 if (mc->mc_snum < 2) {
7501 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7503 /* shift current top to make room for new parent */
7504 mc->mc_pg[1] = mc->mc_pg[0];
7505 mc->mc_ki[1] = mc->mc_ki[0];
7508 mc->mc_db->md_root = pp->mp_pgno;
7509 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7510 mc->mc_db->md_depth++;
7513 /* Add left (implicit) pointer. */
7514 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7515 /* undo the pre-push */
7516 mc->mc_pg[0] = mc->mc_pg[1];
7517 mc->mc_ki[0] = mc->mc_ki[1];
7518 mc->mc_db->md_root = mp->mp_pgno;
7519 mc->mc_db->md_depth--;
7526 ptop = mc->mc_top-1;
7527 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7530 mc->mc_flags |= C_SPLITTING;
7531 mdb_cursor_copy(mc, &mn);
7532 mn.mc_pg[mn.mc_top] = rp;
7533 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7535 if (nflags & MDB_APPEND) {
7536 mn.mc_ki[mn.mc_top] = 0;
7538 split_indx = newindx;
7542 split_indx = (nkeys+1) / 2;
7547 unsigned int lsize, rsize, ksize;
7548 /* Move half of the keys to the right sibling */
7550 x = mc->mc_ki[mc->mc_top] - split_indx;
7551 ksize = mc->mc_db->md_pad;
7552 split = LEAF2KEY(mp, split_indx, ksize);
7553 rsize = (nkeys - split_indx) * ksize;
7554 lsize = (nkeys - split_indx) * sizeof(indx_t);
7555 mp->mp_lower -= lsize;
7556 rp->mp_lower += lsize;
7557 mp->mp_upper += rsize - lsize;
7558 rp->mp_upper -= rsize - lsize;
7559 sepkey.mv_size = ksize;
7560 if (newindx == split_indx) {
7561 sepkey.mv_data = newkey->mv_data;
7563 sepkey.mv_data = split;
7566 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7567 memcpy(rp->mp_ptrs, split, rsize);
7568 sepkey.mv_data = rp->mp_ptrs;
7569 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7570 memcpy(ins, newkey->mv_data, ksize);
7571 mp->mp_lower += sizeof(indx_t);
7572 mp->mp_upper -= ksize - sizeof(indx_t);
7575 memcpy(rp->mp_ptrs, split, x * ksize);
7576 ins = LEAF2KEY(rp, x, ksize);
7577 memcpy(ins, newkey->mv_data, ksize);
7578 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7579 rp->mp_lower += sizeof(indx_t);
7580 rp->mp_upper -= ksize - sizeof(indx_t);
7581 mc->mc_ki[mc->mc_top] = x;
7582 mc->mc_pg[mc->mc_top] = rp;
7585 int psize, nsize, k;
7586 /* Maximum free space in an empty page */
7587 pmax = env->me_psize - PAGEHDRSZ;
7589 nsize = mdb_leaf_size(env, newkey, newdata);
7591 nsize = mdb_branch_size(env, newkey);
7592 nsize = EVEN(nsize);
7594 /* grab a page to hold a temporary copy */
7595 copy = mdb_page_malloc(mc->mc_txn, 1);
7598 copy->mp_pgno = mp->mp_pgno;
7599 copy->mp_flags = mp->mp_flags;
7600 copy->mp_lower = PAGEHDRSZ;
7601 copy->mp_upper = env->me_psize;
7603 /* prepare to insert */
7604 for (i=0, j=0; i<nkeys; i++) {
7606 copy->mp_ptrs[j++] = 0;
7608 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7611 /* When items are relatively large the split point needs
7612 * to be checked, because being off-by-one will make the
7613 * difference between success or failure in mdb_node_add.
7615 * It's also relevant if a page happens to be laid out
7616 * such that one half of its nodes are all "small" and
7617 * the other half of its nodes are "large." If the new
7618 * item is also "large" and falls on the half with
7619 * "large" nodes, it also may not fit.
7621 * As a final tweak, if the new item goes on the last
7622 * spot on the page (and thus, onto the new page), bias
7623 * the split so the new page is emptier than the old page.
7624 * This yields better packing during sequential inserts.
7626 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7627 /* Find split point */
7629 if (newindx <= split_indx || newindx >= nkeys) {
7631 k = newindx >= nkeys ? nkeys : split_indx+2;
7636 for (; i!=k; i+=j) {
7641 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7642 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7644 if (F_ISSET(node->mn_flags, F_BIGDATA))
7645 psize += sizeof(pgno_t);
7647 psize += NODEDSZ(node);
7649 psize = EVEN(psize);
7651 if (psize > pmax || i == k-j) {
7652 split_indx = i + (j<0);
7657 if (split_indx == newindx) {
7658 sepkey.mv_size = newkey->mv_size;
7659 sepkey.mv_data = newkey->mv_data;
7661 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7662 sepkey.mv_size = node->mn_ksize;
7663 sepkey.mv_data = NODEKEY(node);
7668 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7670 /* Copy separator key to the parent.
7672 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7676 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7679 if (mn.mc_snum == mc->mc_snum) {
7680 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7681 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7682 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7683 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7688 /* Right page might now have changed parent.
7689 * Check if left page also changed parent.
7691 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7692 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7693 for (i=0; i<ptop; i++) {
7694 mc->mc_pg[i] = mn.mc_pg[i];
7695 mc->mc_ki[i] = mn.mc_ki[i];
7697 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7698 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7702 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7705 mc->mc_flags ^= C_SPLITTING;
7706 if (rc != MDB_SUCCESS) {
7709 if (nflags & MDB_APPEND) {
7710 mc->mc_pg[mc->mc_top] = rp;
7711 mc->mc_ki[mc->mc_top] = 0;
7712 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7715 for (i=0; i<mc->mc_top; i++)
7716 mc->mc_ki[i] = mn.mc_ki[i];
7717 } else if (!IS_LEAF2(mp)) {
7719 mc->mc_pg[mc->mc_top] = rp;
7724 rkey.mv_data = newkey->mv_data;
7725 rkey.mv_size = newkey->mv_size;
7731 /* Update index for the new key. */
7732 mc->mc_ki[mc->mc_top] = j;
7734 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7735 rkey.mv_data = NODEKEY(node);
7736 rkey.mv_size = node->mn_ksize;
7738 xdata.mv_data = NODEDATA(node);
7739 xdata.mv_size = NODEDSZ(node);
7742 pgno = NODEPGNO(node);
7743 flags = node->mn_flags;
7746 if (!IS_LEAF(mp) && j == 0) {
7747 /* First branch index doesn't need key data. */
7751 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7753 /* return tmp page to freelist */
7754 mdb_page_free(env, copy);
7760 mc->mc_pg[mc->mc_top] = copy;
7765 } while (i != split_indx);
7767 nkeys = NUMKEYS(copy);
7768 for (i=0; i<nkeys; i++)
7769 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7770 mp->mp_lower = copy->mp_lower;
7771 mp->mp_upper = copy->mp_upper;
7772 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7773 env->me_psize - copy->mp_upper);
7775 /* reset back to original page */
7776 if (newindx < split_indx) {
7777 mc->mc_pg[mc->mc_top] = mp;
7778 if (nflags & MDB_RESERVE) {
7779 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7780 if (!(node->mn_flags & F_BIGDATA))
7781 newdata->mv_data = NODEDATA(node);
7784 mc->mc_pg[mc->mc_top] = rp;
7786 /* Make sure mc_ki is still valid.
7788 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7789 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7790 for (i=0; i<ptop; i++) {
7791 mc->mc_pg[i] = mn.mc_pg[i];
7792 mc->mc_ki[i] = mn.mc_ki[i];
7794 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7795 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7798 /* return tmp page to freelist */
7799 mdb_page_free(env, copy);
7803 /* Adjust other cursors pointing to mp */
7804 MDB_cursor *m2, *m3;
7805 MDB_dbi dbi = mc->mc_dbi;
7806 int fixup = NUMKEYS(mp);
7808 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7809 if (mc->mc_flags & C_SUB)
7810 m3 = &m2->mc_xcursor->mx_cursor;
7815 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7817 if (m3->mc_flags & C_SPLITTING)
7822 for (k=m3->mc_top; k>=0; k--) {
7823 m3->mc_ki[k+1] = m3->mc_ki[k];
7824 m3->mc_pg[k+1] = m3->mc_pg[k];
7826 if (m3->mc_ki[0] >= split_indx) {
7831 m3->mc_pg[0] = mc->mc_pg[0];
7835 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7836 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7837 m3->mc_ki[mc->mc_top]++;
7838 if (m3->mc_ki[mc->mc_top] >= fixup) {
7839 m3->mc_pg[mc->mc_top] = rp;
7840 m3->mc_ki[mc->mc_top] -= fixup;
7841 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7843 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7844 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7849 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
7854 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7855 MDB_val *key, MDB_val *data, unsigned int flags)
7860 assert(key != NULL);
7861 assert(data != NULL);
7863 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7866 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7869 mdb_cursor_init(&mc, txn, dbi, &mx);
7870 return mdb_cursor_put(&mc, key, data, flags);
7874 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7876 if ((flag & CHANGEABLE) != flag)
7879 env->me_flags |= flag;
7881 env->me_flags &= ~flag;
7886 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7891 *arg = env->me_flags;
7896 mdb_env_get_path(MDB_env *env, const char **arg)
7901 *arg = env->me_path;
7906 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
7915 /** Common code for #mdb_stat() and #mdb_env_stat().
7916 * @param[in] env the environment to operate in.
7917 * @param[in] db the #MDB_db record containing the stats to return.
7918 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7919 * @return 0, this function always succeeds.
7922 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7924 arg->ms_psize = env->me_psize;
7925 arg->ms_depth = db->md_depth;
7926 arg->ms_branch_pages = db->md_branch_pages;
7927 arg->ms_leaf_pages = db->md_leaf_pages;
7928 arg->ms_overflow_pages = db->md_overflow_pages;
7929 arg->ms_entries = db->md_entries;
7934 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7938 if (env == NULL || arg == NULL)
7941 toggle = mdb_env_pick_meta(env);
7943 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7947 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7951 if (env == NULL || arg == NULL)
7954 toggle = mdb_env_pick_meta(env);
7955 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7956 arg->me_mapsize = env->me_mapsize;
7957 arg->me_maxreaders = env->me_maxreaders;
7959 /* me_numreaders may be zero if this process never used any readers. Use
7960 * the shared numreader count if it exists.
7962 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7964 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7965 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7969 /** Set the default comparison functions for a database.
7970 * Called immediately after a database is opened to set the defaults.
7971 * The user can then override them with #mdb_set_compare() or
7972 * #mdb_set_dupsort().
7973 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7974 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7977 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7979 uint16_t f = txn->mt_dbs[dbi].md_flags;
7981 txn->mt_dbxs[dbi].md_cmp =
7982 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7983 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7985 txn->mt_dbxs[dbi].md_dcmp =
7986 !(f & MDB_DUPSORT) ? 0 :
7987 ((f & MDB_INTEGERDUP)
7988 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7989 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7992 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7997 int rc, dbflag, exact;
7998 unsigned int unused = 0;
8001 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8002 mdb_default_cmp(txn, FREE_DBI);
8005 if ((flags & VALID_FLAGS) != flags)
8007 if (txn->mt_flags & MDB_TXN_ERROR)
8013 if (flags & PERSISTENT_FLAGS) {
8014 uint16_t f2 = flags & PERSISTENT_FLAGS;
8015 /* make sure flag changes get committed */
8016 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8017 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8018 txn->mt_flags |= MDB_TXN_DIRTY;
8021 mdb_default_cmp(txn, MAIN_DBI);
8025 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8026 mdb_default_cmp(txn, MAIN_DBI);
8029 /* Is the DB already open? */
8031 for (i=2; i<txn->mt_numdbs; i++) {
8032 if (!txn->mt_dbxs[i].md_name.mv_size) {
8033 /* Remember this free slot */
8034 if (!unused) unused = i;
8037 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8038 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8044 /* If no free slot and max hit, fail */
8045 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8046 return MDB_DBS_FULL;
8048 /* Cannot mix named databases with some mainDB flags */
8049 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8050 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8052 /* Find the DB info */
8053 dbflag = DB_NEW|DB_VALID;
8056 key.mv_data = (void *)name;
8057 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8058 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8059 if (rc == MDB_SUCCESS) {
8060 /* make sure this is actually a DB */
8061 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8062 if (!(node->mn_flags & F_SUBDATA))
8063 return MDB_INCOMPATIBLE;
8064 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8065 /* Create if requested */
8067 data.mv_size = sizeof(MDB_db);
8068 data.mv_data = &dummy;
8069 memset(&dummy, 0, sizeof(dummy));
8070 dummy.md_root = P_INVALID;
8071 dummy.md_flags = flags & PERSISTENT_FLAGS;
8072 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8076 /* OK, got info, add to table */
8077 if (rc == MDB_SUCCESS) {
8078 unsigned int slot = unused ? unused : txn->mt_numdbs;
8079 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8080 txn->mt_dbxs[slot].md_name.mv_size = len;
8081 txn->mt_dbxs[slot].md_rel = NULL;
8082 txn->mt_dbflags[slot] = dbflag;
8083 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8085 mdb_default_cmp(txn, slot);
8094 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8096 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8099 if (txn->mt_dbflags[dbi] & DB_STALE) {
8102 /* Stale, must read the DB's root. cursor_init does it for us. */
8103 mdb_cursor_init(&mc, txn, dbi, &mx);
8105 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8108 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8111 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8113 ptr = env->me_dbxs[dbi].md_name.mv_data;
8114 env->me_dbxs[dbi].md_name.mv_data = NULL;
8115 env->me_dbxs[dbi].md_name.mv_size = 0;
8116 env->me_dbflags[dbi] = 0;
8120 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8122 /* We could return the flags for the FREE_DBI too but what's the point? */
8123 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8125 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8129 /** Add all the DB's pages to the free list.
8130 * @param[in] mc Cursor on the DB to free.
8131 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8132 * @return 0 on success, non-zero on failure.
8135 mdb_drop0(MDB_cursor *mc, int subs)
8139 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8140 if (rc == MDB_SUCCESS) {
8141 MDB_txn *txn = mc->mc_txn;
8146 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8147 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8150 mdb_cursor_copy(mc, &mx);
8151 while (mc->mc_snum > 0) {
8152 MDB_page *mp = mc->mc_pg[mc->mc_top];
8153 unsigned n = NUMKEYS(mp);
8155 for (i=0; i<n; i++) {
8156 ni = NODEPTR(mp, i);
8157 if (ni->mn_flags & F_BIGDATA) {
8160 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8161 rc = mdb_page_get(txn, pg, &omp, NULL);
8164 assert(IS_OVERFLOW(omp));
8165 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8169 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8170 mdb_xcursor_init1(mc, ni);
8171 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8177 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8179 for (i=0; i<n; i++) {
8181 ni = NODEPTR(mp, i);
8184 mdb_midl_xappend(txn->mt_free_pgs, pg);
8189 mc->mc_ki[mc->mc_top] = i;
8190 rc = mdb_cursor_sibling(mc, 1);
8192 /* no more siblings, go back to beginning
8193 * of previous level.
8197 for (i=1; i<mc->mc_snum; i++) {
8199 mc->mc_pg[i] = mx.mc_pg[i];
8204 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8205 } else if (rc == MDB_NOTFOUND) {
8211 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8213 MDB_cursor *mc, *m2;
8216 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8219 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8222 rc = mdb_cursor_open(txn, dbi, &mc);
8226 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8227 /* Invalidate the dropped DB's cursors */
8228 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8229 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8233 /* Can't delete the main DB */
8234 if (del && dbi > MAIN_DBI) {
8235 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8237 txn->mt_dbflags[dbi] = DB_STALE;
8238 mdb_dbi_close(txn->mt_env, dbi);
8241 /* reset the DB record, mark it dirty */
8242 txn->mt_dbflags[dbi] |= DB_DIRTY;
8243 txn->mt_dbs[dbi].md_depth = 0;
8244 txn->mt_dbs[dbi].md_branch_pages = 0;
8245 txn->mt_dbs[dbi].md_leaf_pages = 0;
8246 txn->mt_dbs[dbi].md_overflow_pages = 0;
8247 txn->mt_dbs[dbi].md_entries = 0;
8248 txn->mt_dbs[dbi].md_root = P_INVALID;
8250 txn->mt_flags |= MDB_TXN_DIRTY;
8253 mdb_cursor_close(mc);
8257 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8259 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8262 txn->mt_dbxs[dbi].md_cmp = cmp;
8266 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8268 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8271 txn->mt_dbxs[dbi].md_dcmp = cmp;
8275 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8277 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8280 txn->mt_dbxs[dbi].md_rel = rel;
8284 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8286 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8289 txn->mt_dbxs[dbi].md_relctx = ctx;
8293 int mdb_env_get_maxkeysize(MDB_env *env)
8295 return ENV_MAXKEY(env);
8298 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8300 unsigned int i, rdrs;
8303 int rc = 0, first = 1;
8307 if (!env->me_txns) {
8308 return func("(no reader locks)\n", ctx);
8310 rdrs = env->me_txns->mti_numreaders;
8311 mr = env->me_txns->mti_readers;
8312 for (i=0; i<rdrs; i++) {
8316 if (mr[i].mr_txnid == (txnid_t)-1) {
8317 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8319 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8323 rc = func(" pid thread txnid\n", ctx);
8327 rc = func(buf, ctx);
8333 rc = func("(no active readers)\n", ctx);
8338 /** Insert pid into list if not already present.
8339 * return -1 if already present.
8341 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
8343 /* binary search of pid in list */
8345 unsigned cursor = 1;
8347 unsigned n = ids[0];
8350 unsigned pivot = n >> 1;
8351 cursor = base + pivot + 1;
8352 val = pid - ids[cursor];
8357 } else if ( val > 0 ) {
8362 /* found, so it's a duplicate */
8371 for (n = ids[0]; n > cursor; n--)
8377 int mdb_reader_check(MDB_env *env, int *dead)
8379 unsigned int i, j, rdrs;
8381 MDB_PID_T *pids, pid;
8390 rdrs = env->me_txns->mti_numreaders;
8391 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
8395 mr = env->me_txns->mti_readers;
8396 for (i=0; i<rdrs; i++) {
8397 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8399 if (mdb_pid_insert(pids, pid) == 0) {
8400 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8402 /* Recheck, a new process may have reused pid */
8403 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8404 for (j=i; j<rdrs; j++)
8405 if (mr[j].mr_pid == pid) {
8406 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8407 (unsigned) pid, mr[j].mr_txnid));
8412 UNLOCK_MUTEX_R(env);