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 #if BYTE_ORDER == LITTLE_ENDIAN
705 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
707 unsigned short mn_hi, mn_lo;
709 /** @defgroup mdb_node Node Flags
711 * Flags for node headers.
714 #define F_BIGDATA 0x01 /**< data put on overflow page */
715 #define F_SUBDATA 0x02 /**< data is a sub-database */
716 #define F_DUPDATA 0x04 /**< data has duplicates */
718 /** valid flags for #mdb_node_add() */
719 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
722 unsigned short mn_flags; /**< @ref mdb_node */
723 unsigned short mn_ksize; /**< key size */
724 char mn_data[1]; /**< key and data are appended here */
727 /** Size of the node header, excluding dynamic data at the end */
728 #define NODESIZE offsetof(MDB_node, mn_data)
730 /** Bit position of top word in page number, for shifting mn_flags */
731 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
733 /** Size of a node in a branch page with a given key.
734 * This is just the node header plus the key, there is no data.
736 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
738 /** Size of a node in a leaf page with a given key and data.
739 * This is node header plus key plus data size.
741 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
743 /** Address of node \b i in page \b p */
744 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
746 /** Address of the key for the node */
747 #define NODEKEY(node) (void *)((node)->mn_data)
749 /** Address of the data for a node */
750 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
752 /** Get the page number pointed to by a branch node */
753 #define NODEPGNO(node) \
754 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
755 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
756 /** Set the page number in a branch node */
757 #define SETPGNO(node,pgno) do { \
758 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
759 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
761 /** Get the size of the data in a leaf node */
762 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
763 /** Set the size of the data for a leaf node */
764 #define SETDSZ(node,size) do { \
765 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
766 /** The size of a key in a node */
767 #define NODEKSZ(node) ((node)->mn_ksize)
769 /** Copy a page number from src to dst */
771 #define COPY_PGNO(dst,src) dst = src
773 #if SIZE_MAX > 4294967295UL
774 #define COPY_PGNO(dst,src) do { \
775 unsigned short *s, *d; \
776 s = (unsigned short *)&(src); \
777 d = (unsigned short *)&(dst); \
784 #define COPY_PGNO(dst,src) do { \
785 unsigned short *s, *d; \
786 s = (unsigned short *)&(src); \
787 d = (unsigned short *)&(dst); \
793 /** The address of a key in a LEAF2 page.
794 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
795 * There are no node headers, keys are stored contiguously.
797 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
799 /** Set the \b node's key into \b keyptr, if requested. */
800 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
801 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
803 /** Set the \b node's key into \b key. */
804 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
806 /** Information about a single database in the environment. */
807 typedef struct MDB_db {
808 uint32_t md_pad; /**< also ksize for LEAF2 pages */
809 uint16_t md_flags; /**< @ref mdb_dbi_open */
810 uint16_t md_depth; /**< depth of this tree */
811 pgno_t md_branch_pages; /**< number of internal pages */
812 pgno_t md_leaf_pages; /**< number of leaf pages */
813 pgno_t md_overflow_pages; /**< number of overflow pages */
814 size_t md_entries; /**< number of data items */
815 pgno_t md_root; /**< the root page of this tree */
818 /** mdb_dbi_open flags */
819 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
820 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
821 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
822 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
824 /** Handle for the DB used to track free pages. */
826 /** Handle for the default DB. */
829 /** Meta page content.
830 * A meta page is the start point for accessing a database snapshot.
831 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
833 typedef struct MDB_meta {
834 /** Stamp identifying this as an MDB file. It must be set
837 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
839 void *mm_address; /**< address for fixed mapping */
840 size_t mm_mapsize; /**< size of mmap region */
841 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
842 /** The size of pages used in this DB */
843 #define mm_psize mm_dbs[0].md_pad
844 /** Any persistent environment flags. @ref mdb_env */
845 #define mm_flags mm_dbs[0].md_flags
846 pgno_t mm_last_pg; /**< last used page in file */
847 txnid_t mm_txnid; /**< txnid that committed this page */
850 /** Buffer for a stack-allocated meta page.
851 * The members define size and alignment, and silence type
852 * aliasing warnings. They are not used directly; that could
853 * mean incorrectly using several union members in parallel.
855 typedef union MDB_metabuf {
858 char mm_pad[PAGEHDRSZ];
863 /** Auxiliary DB info.
864 * The information here is mostly static/read-only. There is
865 * only a single copy of this record in the environment.
867 typedef struct MDB_dbx {
868 MDB_val md_name; /**< name of the database */
869 MDB_cmp_func *md_cmp; /**< function for comparing keys */
870 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
871 MDB_rel_func *md_rel; /**< user relocate function */
872 void *md_relctx; /**< user-provided context for md_rel */
875 /** A database transaction.
876 * Every operation requires a transaction handle.
879 MDB_txn *mt_parent; /**< parent of a nested txn */
880 MDB_txn *mt_child; /**< nested txn under this txn */
881 pgno_t mt_next_pgno; /**< next unallocated page */
882 /** The ID of this transaction. IDs are integers incrementing from 1.
883 * Only committed write transactions increment the ID. If a transaction
884 * aborts, the ID may be re-used by the next writer.
887 MDB_env *mt_env; /**< the DB environment */
888 /** The list of pages that became unused during this transaction.
891 /** The sorted list of dirty pages we temporarily wrote to disk
892 * because the dirty list was full. page numbers in here are
893 * shifted left by 1, deleted slots have the LSB set.
895 MDB_IDL mt_spill_pgs;
897 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
899 /** For read txns: This thread/txn's reader table slot, or NULL. */
902 /** Array of records for each DB known in the environment. */
904 /** Array of MDB_db records for each known DB */
906 /** @defgroup mt_dbflag Transaction DB Flags
910 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
911 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
912 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
913 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
915 /** In write txns, array of cursors for each DB */
916 MDB_cursor **mt_cursors;
917 /** Array of flags for each DB */
918 unsigned char *mt_dbflags;
919 /** Number of DB records in use. This number only ever increments;
920 * we don't decrement it when individual DB handles are closed.
924 /** @defgroup mdb_txn Transaction Flags
928 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
929 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
930 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
931 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
933 unsigned int mt_flags; /**< @ref mdb_txn */
934 /** dirty_list room: Array size - #dirty pages visible to this txn.
935 * Includes ancestor txns' dirty pages not hidden by other txns'
936 * dirty/spilled pages. Thus commit(nested txn) has room to merge
937 * dirty_list into mt_parent after freeing hidden mt_parent pages.
939 unsigned int mt_dirty_room;
942 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
943 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
944 * raise this on a 64 bit machine.
946 #define CURSOR_STACK 32
950 /** Cursors are used for all DB operations.
951 * A cursor holds a path of (page pointer, key index) from the DB
952 * root to a position in the DB, plus other state. #MDB_DUPSORT
953 * cursors include an xcursor to the current data item. Write txns
954 * track their cursors and keep them up to date when data moves.
955 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
956 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
959 /** Next cursor on this DB in this txn */
961 /** Backup of the original cursor if this cursor is a shadow */
962 MDB_cursor *mc_backup;
963 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
964 struct MDB_xcursor *mc_xcursor;
965 /** The transaction that owns this cursor */
967 /** The database handle this cursor operates on */
969 /** The database record for this cursor */
971 /** The database auxiliary record for this cursor */
973 /** The @ref mt_dbflag for this database */
974 unsigned char *mc_dbflag;
975 unsigned short mc_snum; /**< number of pushed pages */
976 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
977 /** @defgroup mdb_cursor Cursor Flags
979 * Cursor state flags.
982 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
983 #define C_EOF 0x02 /**< No more data */
984 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
985 #define C_DEL 0x08 /**< last op was a cursor_del */
986 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
987 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
989 unsigned int mc_flags; /**< @ref mdb_cursor */
990 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
991 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
994 /** Context for sorted-dup records.
995 * We could have gone to a fully recursive design, with arbitrarily
996 * deep nesting of sub-databases. But for now we only handle these
997 * levels - main DB, optional sub-DB, sorted-duplicate DB.
999 typedef struct MDB_xcursor {
1000 /** A sub-cursor for traversing the Dup DB */
1001 MDB_cursor mx_cursor;
1002 /** The database record for this Dup DB */
1004 /** The auxiliary DB record for this Dup DB */
1006 /** The @ref mt_dbflag for this Dup DB */
1007 unsigned char mx_dbflag;
1010 /** State of FreeDB old pages, stored in the MDB_env */
1011 typedef struct MDB_pgstate {
1012 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1013 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1016 /** The database environment. */
1018 HANDLE me_fd; /**< The main data file */
1019 HANDLE me_lfd; /**< The lock file */
1020 HANDLE me_mfd; /**< just for writing the meta pages */
1021 /** Failed to update the meta page. Probably an I/O error. */
1022 #define MDB_FATAL_ERROR 0x80000000U
1023 /** Some fields are initialized. */
1024 #define MDB_ENV_ACTIVE 0x20000000U
1025 /** me_txkey is set */
1026 #define MDB_ENV_TXKEY 0x10000000U
1027 /** Have liveness lock in reader table */
1028 #define MDB_LIVE_READER 0x08000000U
1029 uint32_t me_flags; /**< @ref mdb_env */
1030 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1031 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1032 unsigned int me_maxreaders; /**< size of the reader table */
1033 unsigned int me_numreaders; /**< max numreaders set by this env */
1034 MDB_dbi me_numdbs; /**< number of DBs opened */
1035 MDB_dbi me_maxdbs; /**< size of the DB table */
1036 MDB_PID_T me_pid; /**< process ID of this env */
1037 char *me_path; /**< path to the DB files */
1038 char *me_map; /**< the memory map of the data file */
1039 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1040 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1041 void *me_pbuf; /**< scratch area for DUPSORT put() */
1042 MDB_txn *me_txn; /**< current write transaction */
1043 size_t me_mapsize; /**< size of the data memory map */
1044 off_t me_size; /**< current file size */
1045 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1046 MDB_dbx *me_dbxs; /**< array of static DB info */
1047 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1048 pthread_key_t me_txkey; /**< thread-key for readers */
1049 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1050 # define me_pglast me_pgstate.mf_pglast
1051 # define me_pghead me_pgstate.mf_pghead
1052 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1053 /** IDL of pages that became unused in a write txn */
1054 MDB_IDL me_free_pgs;
1055 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1056 MDB_ID2L me_dirty_list;
1057 /** Max number of freelist items that can fit in a single overflow page */
1059 /** Max size of a node on a page */
1060 unsigned int me_nodemax;
1061 #if !(MDB_MAXKEYSIZE)
1062 unsigned int me_maxkey; /**< max size of a key */
1065 int me_pidquery; /**< Used in OpenProcess */
1066 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1068 #elif defined(MDB_USE_POSIX_SEM)
1069 sem_t *me_rmutex; /* Shared mutexes are not supported */
1074 /** Nested transaction */
1075 typedef struct MDB_ntxn {
1076 MDB_txn mnt_txn; /**< the transaction */
1077 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1080 /** max number of pages to commit in one writev() call */
1081 #define MDB_COMMIT_PAGES 64
1082 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1083 #undef MDB_COMMIT_PAGES
1084 #define MDB_COMMIT_PAGES IOV_MAX
1087 /* max bytes to write in one call */
1088 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1090 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1091 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1092 static int mdb_page_touch(MDB_cursor *mc);
1094 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1095 static int mdb_page_search_root(MDB_cursor *mc,
1096 MDB_val *key, int modify);
1097 #define MDB_PS_MODIFY 1
1098 #define MDB_PS_ROOTONLY 2
1099 #define MDB_PS_FIRST 4
1100 #define MDB_PS_LAST 8
1101 static int mdb_page_search(MDB_cursor *mc,
1102 MDB_val *key, int flags);
1103 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1105 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1106 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1107 pgno_t newpgno, unsigned int nflags);
1109 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1110 static int mdb_env_pick_meta(const MDB_env *env);
1111 static int mdb_env_write_meta(MDB_txn *txn);
1112 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1113 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1115 static void mdb_env_close0(MDB_env *env, int excl);
1117 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1118 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1119 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1120 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1121 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1122 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1123 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1124 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1125 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1127 static int mdb_rebalance(MDB_cursor *mc);
1128 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1130 static void mdb_cursor_pop(MDB_cursor *mc);
1131 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1133 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1134 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1135 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1136 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1137 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1139 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1140 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1142 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1143 static void mdb_xcursor_init0(MDB_cursor *mc);
1144 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1146 static int mdb_drop0(MDB_cursor *mc, int subs);
1147 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1150 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1154 static SECURITY_DESCRIPTOR mdb_null_sd;
1155 static SECURITY_ATTRIBUTES mdb_all_sa;
1156 static int mdb_sec_inited;
1159 /** Return the library version info. */
1161 mdb_version(int *major, int *minor, int *patch)
1163 if (major) *major = MDB_VERSION_MAJOR;
1164 if (minor) *minor = MDB_VERSION_MINOR;
1165 if (patch) *patch = MDB_VERSION_PATCH;
1166 return MDB_VERSION_STRING;
1169 /** Table of descriptions for MDB @ref errors */
1170 static char *const mdb_errstr[] = {
1171 "MDB_KEYEXIST: Key/data pair already exists",
1172 "MDB_NOTFOUND: No matching key/data pair found",
1173 "MDB_PAGE_NOTFOUND: Requested page not found",
1174 "MDB_CORRUPTED: Located page was wrong type",
1175 "MDB_PANIC: Update of meta page failed",
1176 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1177 "MDB_INVALID: File is not an MDB file",
1178 "MDB_MAP_FULL: Environment mapsize limit reached",
1179 "MDB_DBS_FULL: Environment maxdbs limit reached",
1180 "MDB_READERS_FULL: Environment maxreaders limit reached",
1181 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1182 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1183 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1184 "MDB_PAGE_FULL: Internal error - page has no more space",
1185 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1186 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1187 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1188 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1189 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1193 mdb_strerror(int err)
1197 return ("Successful return: 0");
1199 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1200 i = err - MDB_KEYEXIST;
1201 return mdb_errstr[i];
1204 return strerror(err);
1208 /** Display a key in hexadecimal and return the address of the result.
1209 * @param[in] key the key to display
1210 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1211 * @return The key in hexadecimal form.
1214 mdb_dkey(MDB_val *key, char *buf)
1217 unsigned char *c = key->mv_data;
1223 if (key->mv_size > DKBUF_MAXKEYSIZE)
1224 return "MDB_MAXKEYSIZE";
1225 /* may want to make this a dynamic check: if the key is mostly
1226 * printable characters, print it as-is instead of converting to hex.
1230 for (i=0; i<key->mv_size; i++)
1231 ptr += sprintf(ptr, "%02x", *c++);
1233 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1238 /** Display all the keys in the page. */
1240 mdb_page_list(MDB_page *mp)
1243 unsigned int i, nkeys, nsize, total = 0;
1247 nkeys = NUMKEYS(mp);
1248 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1249 for (i=0; i<nkeys; i++) {
1250 node = NODEPTR(mp, i);
1251 key.mv_size = node->mn_ksize;
1252 key.mv_data = node->mn_data;
1253 nsize = NODESIZE + key.mv_size;
1254 if (IS_BRANCH(mp)) {
1255 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1259 if (F_ISSET(node->mn_flags, F_BIGDATA))
1260 nsize += sizeof(pgno_t);
1262 nsize += NODEDSZ(node);
1264 nsize += sizeof(indx_t);
1265 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1267 total = EVEN(total);
1269 fprintf(stderr, "Total: %d\n", total);
1273 mdb_cursor_chk(MDB_cursor *mc)
1279 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1280 for (i=0; i<mc->mc_top; i++) {
1282 node = NODEPTR(mp, mc->mc_ki[i]);
1283 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1286 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1292 /** Count all the pages in each DB and in the freelist
1293 * and make sure it matches the actual number of pages
1296 static void mdb_audit(MDB_txn *txn)
1300 MDB_ID freecount, count;
1305 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1306 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1307 freecount += *(MDB_ID *)data.mv_data;
1310 for (i = 0; i<txn->mt_numdbs; i++) {
1312 mdb_cursor_init(&mc, txn, i, &mx);
1313 if (txn->mt_dbs[i].md_root == P_INVALID)
1315 count += txn->mt_dbs[i].md_branch_pages +
1316 txn->mt_dbs[i].md_leaf_pages +
1317 txn->mt_dbs[i].md_overflow_pages;
1318 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1319 mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1323 mp = mc.mc_pg[mc.mc_top];
1324 for (j=0; j<NUMKEYS(mp); j++) {
1325 MDB_node *leaf = NODEPTR(mp, j);
1326 if (leaf->mn_flags & F_SUBDATA) {
1328 memcpy(&db, NODEDATA(leaf), sizeof(db));
1329 count += db.md_branch_pages + db.md_leaf_pages +
1330 db.md_overflow_pages;
1334 while (mdb_cursor_sibling(&mc, 1) == 0);
1337 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1338 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1339 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1345 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1347 return txn->mt_dbxs[dbi].md_cmp(a, b);
1351 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1353 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1356 /** Allocate memory for a page.
1357 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1360 mdb_page_malloc(MDB_txn *txn, unsigned num)
1362 MDB_env *env = txn->mt_env;
1363 MDB_page *ret = env->me_dpages;
1364 size_t psize = env->me_psize, sz = psize, off;
1365 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1366 * For a single page alloc, we init everything after the page header.
1367 * For multi-page, we init the final page; if the caller needed that
1368 * many pages they will be filling in at least up to the last page.
1372 VGMEMP_ALLOC(env, ret, sz);
1373 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1374 env->me_dpages = ret->mp_next;
1377 psize -= off = PAGEHDRSZ;
1382 if ((ret = malloc(sz)) != NULL) {
1383 if (!(env->me_flags & MDB_NOMEMINIT)) {
1384 memset((char *)ret + off, 0, psize);
1387 VGMEMP_ALLOC(env, ret, sz);
1392 /** Free a single page.
1393 * Saves single pages to a list, for future reuse.
1394 * (This is not used for multi-page overflow pages.)
1397 mdb_page_free(MDB_env *env, MDB_page *mp)
1399 mp->mp_next = env->me_dpages;
1400 VGMEMP_FREE(env, mp);
1401 env->me_dpages = mp;
1404 /** Free a dirty page */
1406 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1408 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1409 mdb_page_free(env, dp);
1411 /* large pages just get freed directly */
1412 VGMEMP_FREE(env, dp);
1417 /** Return all dirty pages to dpage list */
1419 mdb_dlist_free(MDB_txn *txn)
1421 MDB_env *env = txn->mt_env;
1422 MDB_ID2L dl = txn->mt_u.dirty_list;
1423 unsigned i, n = dl[0].mid;
1425 for (i = 1; i <= n; i++) {
1426 mdb_dpage_free(env, dl[i].mptr);
1431 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1432 * @param[in] mc A cursor handle for the current operation.
1433 * @param[in] pflags Flags of the pages to update:
1434 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1435 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1436 * @return 0 on success, non-zero on failure.
1439 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1441 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1442 MDB_txn *txn = mc->mc_txn;
1448 int rc = MDB_SUCCESS, level;
1450 /* Mark pages seen by cursors */
1451 if (mc->mc_flags & C_UNTRACK)
1452 mc = NULL; /* will find mc in mt_cursors */
1453 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1454 for (; mc; mc=mc->mc_next) {
1455 if (!(mc->mc_flags & C_INITIALIZED))
1457 for (m3 = mc;; m3 = &mx->mx_cursor) {
1459 for (j=0; j<m3->mc_snum; j++) {
1461 if ((mp->mp_flags & Mask) == pflags)
1462 mp->mp_flags ^= P_KEEP;
1464 mx = m3->mc_xcursor;
1465 /* Proceed to mx if it is at a sub-database */
1466 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1468 if (! (mp && (mp->mp_flags & P_LEAF)))
1470 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1471 if (!(leaf->mn_flags & F_SUBDATA))
1480 /* Mark dirty root pages */
1481 for (i=0; i<txn->mt_numdbs; i++) {
1482 if (txn->mt_dbflags[i] & DB_DIRTY) {
1483 pgno_t pgno = txn->mt_dbs[i].md_root;
1484 if (pgno == P_INVALID)
1486 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1488 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1489 dp->mp_flags ^= P_KEEP;
1497 static int mdb_page_flush(MDB_txn *txn, int keep);
1499 /** Spill pages from the dirty list back to disk.
1500 * This is intended to prevent running into #MDB_TXN_FULL situations,
1501 * but note that they may still occur in a few cases:
1502 * 1) our estimate of the txn size could be too small. Currently this
1503 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1504 * 2) child txns may run out of space if their parents dirtied a
1505 * lot of pages and never spilled them. TODO: we probably should do
1506 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1507 * the parent's dirty_room is below a given threshold.
1509 * Otherwise, if not using nested txns, it is expected that apps will
1510 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1511 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1512 * If the txn never references them again, they can be left alone.
1513 * If the txn only reads them, they can be used without any fuss.
1514 * If the txn writes them again, they can be dirtied immediately without
1515 * going thru all of the work of #mdb_page_touch(). Such references are
1516 * handled by #mdb_page_unspill().
1518 * Also note, we never spill DB root pages, nor pages of active cursors,
1519 * because we'll need these back again soon anyway. And in nested txns,
1520 * we can't spill a page in a child txn if it was already spilled in a
1521 * parent txn. That would alter the parent txns' data even though
1522 * the child hasn't committed yet, and we'd have no way to undo it if
1523 * the child aborted.
1525 * @param[in] m0 cursor A cursor handle identifying the transaction and
1526 * database for which we are checking space.
1527 * @param[in] key For a put operation, the key being stored.
1528 * @param[in] data For a put operation, the data being stored.
1529 * @return 0 on success, non-zero on failure.
1532 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1534 MDB_txn *txn = m0->mc_txn;
1536 MDB_ID2L dl = txn->mt_u.dirty_list;
1537 unsigned int i, j, need;
1540 if (m0->mc_flags & C_SUB)
1543 /* Estimate how much space this op will take */
1544 i = m0->mc_db->md_depth;
1545 /* Named DBs also dirty the main DB */
1546 if (m0->mc_dbi > MAIN_DBI)
1547 i += txn->mt_dbs[MAIN_DBI].md_depth;
1548 /* For puts, roughly factor in the key+data size */
1550 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1551 i += i; /* double it for good measure */
1554 if (txn->mt_dirty_room > i)
1557 if (!txn->mt_spill_pgs) {
1558 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1559 if (!txn->mt_spill_pgs)
1562 /* purge deleted slots */
1563 MDB_IDL sl = txn->mt_spill_pgs;
1564 unsigned int num = sl[0];
1566 for (i=1; i<=num; i++) {
1573 /* Preserve pages which may soon be dirtied again */
1574 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1577 /* Less aggressive spill - we originally spilled the entire dirty list,
1578 * with a few exceptions for cursor pages and DB root pages. But this
1579 * turns out to be a lot of wasted effort because in a large txn many
1580 * of those pages will need to be used again. So now we spill only 1/8th
1581 * of the dirty pages. Testing revealed this to be a good tradeoff,
1582 * better than 1/2, 1/4, or 1/10.
1584 if (need < MDB_IDL_UM_MAX / 8)
1585 need = MDB_IDL_UM_MAX / 8;
1587 /* Save the page IDs of all the pages we're flushing */
1588 /* flush from the tail forward, this saves a lot of shifting later on. */
1589 for (i=dl[0].mid; i && need; i--) {
1590 MDB_ID pn = dl[i].mid << 1;
1592 if (dp->mp_flags & P_KEEP)
1594 /* Can't spill twice, make sure it's not already in a parent's
1597 if (txn->mt_parent) {
1599 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1600 if (tx2->mt_spill_pgs) {
1601 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1602 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1603 dp->mp_flags |= P_KEEP;
1611 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1615 mdb_midl_sort(txn->mt_spill_pgs);
1617 /* Flush the spilled part of dirty list */
1618 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1621 /* Reset any dirty pages we kept that page_flush didn't see */
1622 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1625 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1629 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1631 mdb_find_oldest(MDB_txn *txn)
1634 txnid_t mr, oldest = txn->mt_txnid - 1;
1635 if (txn->mt_env->me_txns) {
1636 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1637 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1648 /** Add a page to the txn's dirty list */
1650 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1653 int (*insert)(MDB_ID2L, MDB_ID2 *);
1655 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1656 insert = mdb_mid2l_append;
1658 insert = mdb_mid2l_insert;
1660 mid.mid = mp->mp_pgno;
1662 insert(txn->mt_u.dirty_list, &mid);
1663 txn->mt_dirty_room--;
1666 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1667 * me_pghead and mt_next_pgno.
1669 * If there are free pages available from older transactions, they
1670 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1671 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1672 * and move me_pglast to say which records were consumed. Only this
1673 * function can create me_pghead and move me_pglast/mt_next_pgno.
1674 * @param[in] mc cursor A cursor handle identifying the transaction and
1675 * database for which we are allocating.
1676 * @param[in] num the number of pages to allocate.
1677 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1678 * will always be satisfied by a single contiguous chunk of memory.
1679 * @return 0 on success, non-zero on failure.
1682 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1684 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1685 /* Get at most <Max_retries> more freeDB records once me_pghead
1686 * has enough pages. If not enough, use new pages from the map.
1687 * If <Paranoid> and mc is updating the freeDB, only get new
1688 * records if me_pghead is empty. Then the freelist cannot play
1689 * catch-up with itself by growing while trying to save it.
1691 enum { Paranoid = 1, Max_retries = 500 };
1693 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1695 int rc, retry = Max_retries;
1696 MDB_txn *txn = mc->mc_txn;
1697 MDB_env *env = txn->mt_env;
1698 pgno_t pgno, *mop = env->me_pghead;
1699 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1701 txnid_t oldest = 0, last;
1707 /* If our dirty list is already full, we can't do anything */
1708 if (txn->mt_dirty_room == 0)
1709 return MDB_TXN_FULL;
1711 for (op = MDB_FIRST;; op = MDB_NEXT) {
1714 pgno_t *idl, old_id, new_id;
1716 /* Seek a big enough contiguous page range. Prefer
1717 * pages at the tail, just truncating the list.
1723 if (mop[i-n2] == pgno+n2)
1726 if (Max_retries < INT_MAX && --retry < 0)
1730 if (op == MDB_FIRST) { /* 1st iteration */
1731 /* Prepare to fetch more and coalesce */
1732 oldest = mdb_find_oldest(txn);
1733 last = env->me_pglast;
1734 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1737 key.mv_data = &last; /* will look up last+1 */
1738 key.mv_size = sizeof(last);
1740 if (Paranoid && mc->mc_dbi == FREE_DBI)
1743 if (Paranoid && retry < 0 && mop_len)
1747 /* Do not fetch more if the record will be too recent */
1750 rc = mdb_cursor_get(&m2, &key, NULL, op);
1752 if (rc == MDB_NOTFOUND)
1756 last = *(txnid_t*)key.mv_data;
1759 np = m2.mc_pg[m2.mc_top];
1760 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1761 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1764 idl = (MDB_ID *) data.mv_data;
1767 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1770 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1772 mop = env->me_pghead;
1774 env->me_pglast = last;
1776 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1777 last, txn->mt_dbs[FREE_DBI].md_root, i));
1779 DPRINTF(("IDL %"Z"u", idl[k]));
1781 /* Merge in descending sorted order */
1784 mop[0] = (pgno_t)-1;
1788 for (; old_id < new_id; old_id = mop[--j])
1795 /* Use new pages from the map when nothing suitable in the freeDB */
1797 pgno = txn->mt_next_pgno;
1798 if (pgno + num >= env->me_maxpg) {
1799 DPUTS("DB size maxed out");
1800 return MDB_MAP_FULL;
1804 if (env->me_flags & MDB_WRITEMAP) {
1805 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1807 if (!(np = mdb_page_malloc(txn, num)))
1811 mop[0] = mop_len -= num;
1812 /* Move any stragglers down */
1813 for (j = i-num; j < mop_len; )
1814 mop[++j] = mop[++i];
1816 txn->mt_next_pgno = pgno + num;
1819 mdb_page_dirty(txn, np);
1825 /** Copy the used portions of a non-overflow page.
1826 * @param[in] dst page to copy into
1827 * @param[in] src page to copy from
1828 * @param[in] psize size of a page
1831 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1833 enum { Align = sizeof(pgno_t) };
1834 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1836 /* If page isn't full, just copy the used portion. Adjust
1837 * alignment so memcpy may copy words instead of bytes.
1839 if ((unused &= -Align) && !IS_LEAF2(src)) {
1841 memcpy(dst, src, (lower + (Align-1)) & -Align);
1842 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1845 memcpy(dst, src, psize - unused);
1849 /** Pull a page off the txn's spill list, if present.
1850 * If a page being referenced was spilled to disk in this txn, bring
1851 * it back and make it dirty/writable again.
1852 * @param[in] txn the transaction handle.
1853 * @param[in] mp the page being referenced. It must not be dirty.
1854 * @param[out] ret the writable page, if any. ret is unchanged if
1855 * mp wasn't spilled.
1858 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1860 MDB_env *env = txn->mt_env;
1863 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1865 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1866 if (!tx2->mt_spill_pgs)
1868 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1869 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1872 if (txn->mt_dirty_room == 0)
1873 return MDB_TXN_FULL;
1874 if (IS_OVERFLOW(mp))
1878 if (env->me_flags & MDB_WRITEMAP) {
1881 np = mdb_page_malloc(txn, num);
1885 memcpy(np, mp, num * env->me_psize);
1887 mdb_page_copy(np, mp, env->me_psize);
1890 /* If in current txn, this page is no longer spilled.
1891 * If it happens to be the last page, truncate the spill list.
1892 * Otherwise mark it as deleted by setting the LSB.
1894 if (x == txn->mt_spill_pgs[0])
1895 txn->mt_spill_pgs[0]--;
1897 txn->mt_spill_pgs[x] |= 1;
1898 } /* otherwise, if belonging to a parent txn, the
1899 * page remains spilled until child commits
1902 mdb_page_dirty(txn, np);
1903 np->mp_flags |= P_DIRTY;
1911 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1912 * @param[in] mc cursor pointing to the page to be touched
1913 * @return 0 on success, non-zero on failure.
1916 mdb_page_touch(MDB_cursor *mc)
1918 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1919 MDB_txn *txn = mc->mc_txn;
1920 MDB_cursor *m2, *m3;
1924 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1925 if (txn->mt_flags & MDB_TXN_SPILLS) {
1927 rc = mdb_page_unspill(txn, mp, &np);
1933 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1934 (rc = mdb_page_alloc(mc, 1, &np)))
1937 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
1938 mp->mp_pgno, pgno));
1939 assert(mp->mp_pgno != pgno);
1940 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1941 /* Update the parent page, if any, to point to the new page */
1943 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1944 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1945 SETPGNO(node, pgno);
1947 mc->mc_db->md_root = pgno;
1949 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1950 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1952 /* If txn has a parent, make sure the page is in our
1956 unsigned x = mdb_mid2l_search(dl, pgno);
1957 if (x <= dl[0].mid && dl[x].mid == pgno) {
1958 if (mp != dl[x].mptr) { /* bad cursor? */
1959 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1960 return MDB_CORRUPTED;
1965 assert(dl[0].mid < MDB_IDL_UM_MAX);
1967 np = mdb_page_malloc(txn, 1);
1972 mdb_mid2l_insert(dl, &mid);
1977 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1979 np->mp_flags |= P_DIRTY;
1982 /* Adjust cursors pointing to mp */
1983 mc->mc_pg[mc->mc_top] = np;
1984 m2 = txn->mt_cursors[mc->mc_dbi];
1985 if (mc->mc_flags & C_SUB) {
1986 for (; m2; m2=m2->mc_next) {
1987 m3 = &m2->mc_xcursor->mx_cursor;
1988 if (m3->mc_snum < mc->mc_snum) continue;
1989 if (m3->mc_pg[mc->mc_top] == mp)
1990 m3->mc_pg[mc->mc_top] = np;
1993 for (; m2; m2=m2->mc_next) {
1994 if (m2->mc_snum < mc->mc_snum) continue;
1995 if (m2->mc_pg[mc->mc_top] == mp) {
1996 m2->mc_pg[mc->mc_top] = np;
1997 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1998 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2000 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2001 if (!(leaf->mn_flags & F_SUBDATA))
2002 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2011 mdb_env_sync(MDB_env *env, int force)
2014 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2015 if (env->me_flags & MDB_WRITEMAP) {
2016 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2017 ? MS_ASYNC : MS_SYNC;
2018 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2021 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2025 if (MDB_FDATASYNC(env->me_fd))
2032 /** Back up parent txn's cursors, then grab the originals for tracking */
2034 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2036 MDB_cursor *mc, *bk;
2041 for (i = src->mt_numdbs; --i >= 0; ) {
2042 if ((mc = src->mt_cursors[i]) != NULL) {
2043 size = sizeof(MDB_cursor);
2045 size += sizeof(MDB_xcursor);
2046 for (; mc; mc = bk->mc_next) {
2052 mc->mc_db = &dst->mt_dbs[i];
2053 /* Kill pointers into src - and dst to reduce abuse: The
2054 * user may not use mc until dst ends. Otherwise we'd...
2056 mc->mc_txn = NULL; /* ...set this to dst */
2057 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2058 if ((mx = mc->mc_xcursor) != NULL) {
2059 *(MDB_xcursor *)(bk+1) = *mx;
2060 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2062 mc->mc_next = dst->mt_cursors[i];
2063 dst->mt_cursors[i] = mc;
2070 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2071 * @param[in] txn the transaction handle.
2072 * @param[in] merge true to keep changes to parent cursors, false to revert.
2073 * @return 0 on success, non-zero on failure.
2076 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2078 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2082 for (i = txn->mt_numdbs; --i >= 0; ) {
2083 for (mc = cursors[i]; mc; mc = next) {
2085 if ((bk = mc->mc_backup) != NULL) {
2087 /* Commit changes to parent txn */
2088 mc->mc_next = bk->mc_next;
2089 mc->mc_backup = bk->mc_backup;
2090 mc->mc_txn = bk->mc_txn;
2091 mc->mc_db = bk->mc_db;
2092 mc->mc_dbflag = bk->mc_dbflag;
2093 if ((mx = mc->mc_xcursor) != NULL)
2094 mx->mx_cursor.mc_txn = bk->mc_txn;
2096 /* Abort nested txn */
2098 if ((mx = mc->mc_xcursor) != NULL)
2099 *mx = *(MDB_xcursor *)(bk+1);
2103 /* Only malloced cursors are permanently tracked. */
2111 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2114 mdb_txn_reset0(MDB_txn *txn, const char *act);
2116 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2122 Pidset = F_SETLK, Pidcheck = F_GETLK
2126 /** Set or check a pid lock. Set returns 0 on success.
2127 * Check returns 0 if the process is certainly dead, nonzero if it may
2128 * be alive (the lock exists or an error happened so we do not know).
2130 * On Windows Pidset is a no-op, we merely check for the existence
2131 * of the process with the given pid. On POSIX we use a single byte
2132 * lock on the lockfile, set at an offset equal to the pid.
2135 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2137 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2140 if (op == Pidcheck) {
2141 h = OpenProcess(env->me_pidquery, FALSE, pid);
2142 /* No documented "no such process" code, but other program use this: */
2144 return ErrCode() != ERROR_INVALID_PARAMETER;
2145 /* A process exists until all handles to it close. Has it exited? */
2146 ret = WaitForSingleObject(h, 0) != 0;
2153 struct flock lock_info;
2154 memset(&lock_info, 0, sizeof(lock_info));
2155 lock_info.l_type = F_WRLCK;
2156 lock_info.l_whence = SEEK_SET;
2157 lock_info.l_start = pid;
2158 lock_info.l_len = 1;
2159 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2160 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2162 } else if ((rc = ErrCode()) == EINTR) {
2170 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2171 * @param[in] txn the transaction handle to initialize
2172 * @return 0 on success, non-zero on failure.
2175 mdb_txn_renew0(MDB_txn *txn)
2177 MDB_env *env = txn->mt_env;
2178 MDB_txninfo *ti = env->me_txns;
2182 int rc, new_notls = 0;
2185 txn->mt_numdbs = env->me_numdbs;
2186 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2188 if (txn->mt_flags & MDB_TXN_RDONLY) {
2190 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2191 txn->mt_txnid = meta->mm_txnid;
2192 txn->mt_u.reader = NULL;
2194 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2195 pthread_getspecific(env->me_txkey);
2197 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2198 return MDB_BAD_RSLOT;
2200 MDB_PID_T pid = env->me_pid;
2201 pthread_t tid = pthread_self();
2203 if (!(env->me_flags & MDB_LIVE_READER)) {
2204 rc = mdb_reader_pid(env, Pidset, pid);
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)
4014 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4015 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4019 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4021 int oflags, rc, len, excl = -1;
4022 char *lpath, *dpath;
4024 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4028 if (flags & MDB_NOSUBDIR) {
4029 rc = len + sizeof(LOCKSUFF) + len + 1;
4031 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4036 if (flags & MDB_NOSUBDIR) {
4037 dpath = lpath + len + sizeof(LOCKSUFF);
4038 sprintf(lpath, "%s" LOCKSUFF, path);
4039 strcpy(dpath, path);
4041 dpath = lpath + len + sizeof(LOCKNAME);
4042 sprintf(lpath, "%s" LOCKNAME, path);
4043 sprintf(dpath, "%s" DATANAME, path);
4047 flags |= env->me_flags;
4048 if (flags & MDB_RDONLY) {
4049 /* silently ignore WRITEMAP when we're only getting read access */
4050 flags &= ~MDB_WRITEMAP;
4052 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4053 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4056 env->me_flags = flags |= MDB_ENV_ACTIVE;
4060 env->me_path = strdup(path);
4061 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4062 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4063 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4068 /* For RDONLY, get lockfile after we know datafile exists */
4069 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4070 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4076 if (F_ISSET(flags, MDB_RDONLY)) {
4077 oflags = GENERIC_READ;
4078 len = OPEN_EXISTING;
4080 oflags = GENERIC_READ|GENERIC_WRITE;
4083 mode = FILE_ATTRIBUTE_NORMAL;
4084 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4085 NULL, len, mode, NULL);
4087 if (F_ISSET(flags, MDB_RDONLY))
4090 oflags = O_RDWR | O_CREAT;
4092 env->me_fd = open(dpath, oflags, mode);
4094 if (env->me_fd == INVALID_HANDLE_VALUE) {
4099 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4100 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4105 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4106 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4107 env->me_mfd = env->me_fd;
4109 /* Synchronous fd for meta writes. Needed even with
4110 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4113 len = OPEN_EXISTING;
4114 env->me_mfd = CreateFile(dpath, oflags,
4115 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4116 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4119 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4121 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4126 DPRINTF(("opened dbenv %p", (void *) env));
4128 rc = mdb_env_share_locks(env, &excl);
4132 if (!((flags & MDB_RDONLY) ||
4133 (env->me_pbuf = calloc(1, env->me_psize))))
4139 mdb_env_close0(env, excl);
4145 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4147 mdb_env_close0(MDB_env *env, int excl)
4151 if (!(env->me_flags & MDB_ENV_ACTIVE))
4154 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4155 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4156 free(env->me_dbxs[i].md_name.mv_data);
4159 free(env->me_dbflags);
4162 free(env->me_dirty_list);
4163 mdb_midl_free(env->me_free_pgs);
4165 if (env->me_flags & MDB_ENV_TXKEY) {
4166 pthread_key_delete(env->me_txkey);
4168 /* Delete our key from the global list */
4169 for (i=0; i<mdb_tls_nkeys; i++)
4170 if (mdb_tls_keys[i] == env->me_txkey) {
4171 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4179 munmap(env->me_map, env->me_mapsize);
4181 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4182 (void) close(env->me_mfd);
4183 if (env->me_fd != INVALID_HANDLE_VALUE)
4184 (void) close(env->me_fd);
4186 MDB_PID_T pid = env->me_pid;
4187 /* Clearing readers is done in this function because
4188 * me_txkey with its destructor must be disabled first.
4190 for (i = env->me_numreaders; --i >= 0; )
4191 if (env->me_txns->mti_readers[i].mr_pid == pid)
4192 env->me_txns->mti_readers[i].mr_pid = 0;
4194 if (env->me_rmutex) {
4195 CloseHandle(env->me_rmutex);
4196 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4198 /* Windows automatically destroys the mutexes when
4199 * the last handle closes.
4201 #elif defined(MDB_USE_POSIX_SEM)
4202 if (env->me_rmutex != SEM_FAILED) {
4203 sem_close(env->me_rmutex);
4204 if (env->me_wmutex != SEM_FAILED)
4205 sem_close(env->me_wmutex);
4206 /* If we have the filelock: If we are the
4207 * only remaining user, clean up semaphores.
4210 mdb_env_excl_lock(env, &excl);
4212 sem_unlink(env->me_txns->mti_rmname);
4213 sem_unlink(env->me_txns->mti_wmname);
4217 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4219 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4222 /* Unlock the lockfile. Windows would have unlocked it
4223 * after closing anyway, but not necessarily at once.
4225 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4228 (void) close(env->me_lfd);
4231 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4235 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4237 MDB_txn *txn = NULL;
4243 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4247 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4250 /* Do the lock/unlock of the reader mutex before starting the
4251 * write txn. Otherwise other read txns could block writers.
4253 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4258 /* We must start the actual read txn after blocking writers */
4259 mdb_txn_reset0(txn, "reset-stage1");
4261 /* Temporarily block writers until we snapshot the meta pages */
4264 rc = mdb_txn_renew0(txn);
4266 UNLOCK_MUTEX_W(env);
4271 wsize = env->me_psize * 2;
4275 DO_WRITE(rc, fd, ptr, w2, len);
4279 } else if (len > 0) {
4285 /* Non-blocking or async handles are not supported */
4291 UNLOCK_MUTEX_W(env);
4296 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4298 if (wsize > MAX_WRITE)
4302 DO_WRITE(rc, fd, ptr, w2, len);
4306 } else if (len > 0) {
4323 mdb_env_copy(MDB_env *env, const char *path)
4327 HANDLE newfd = INVALID_HANDLE_VALUE;
4329 if (env->me_flags & MDB_NOSUBDIR) {
4330 lpath = (char *)path;
4333 len += sizeof(DATANAME);
4334 lpath = malloc(len);
4337 sprintf(lpath, "%s" DATANAME, path);
4340 /* The destination path must exist, but the destination file must not.
4341 * We don't want the OS to cache the writes, since the source data is
4342 * already in the OS cache.
4345 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4346 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4348 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4350 if (newfd == INVALID_HANDLE_VALUE) {
4356 /* Set O_DIRECT if the file system supports it */
4357 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4358 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4360 #ifdef F_NOCACHE /* __APPLE__ */
4361 rc = fcntl(newfd, F_NOCACHE, 1);
4368 rc = mdb_env_copyfd(env, newfd);
4371 if (!(env->me_flags & MDB_NOSUBDIR))
4373 if (newfd != INVALID_HANDLE_VALUE)
4374 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4381 mdb_env_close(MDB_env *env)
4388 VGMEMP_DESTROY(env);
4389 while ((dp = env->me_dpages) != NULL) {
4390 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4391 env->me_dpages = dp->mp_next;
4395 mdb_env_close0(env, 0);
4399 /** Compare two items pointing at aligned size_t's */
4401 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4403 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4404 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4407 /** Compare two items pointing at aligned unsigned int's */
4409 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4411 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4412 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4415 /** Compare two items pointing at unsigned ints of unknown alignment.
4416 * Nodes and keys are guaranteed to be 2-byte aligned.
4419 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4421 #if BYTE_ORDER == LITTLE_ENDIAN
4422 unsigned short *u, *c;
4425 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4426 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4429 } while(!x && u > (unsigned short *)a->mv_data);
4432 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4436 /** Compare two items lexically */
4438 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4445 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4451 diff = memcmp(a->mv_data, b->mv_data, len);
4452 return diff ? diff : len_diff<0 ? -1 : len_diff;
4455 /** Compare two items in reverse byte order */
4457 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4459 const unsigned char *p1, *p2, *p1_lim;
4463 p1_lim = (const unsigned char *)a->mv_data;
4464 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4465 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4467 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4473 while (p1 > p1_lim) {
4474 diff = *--p1 - *--p2;
4478 return len_diff<0 ? -1 : len_diff;
4481 /** Search for key within a page, using binary search.
4482 * Returns the smallest entry larger or equal to the key.
4483 * If exactp is non-null, stores whether the found entry was an exact match
4484 * in *exactp (1 or 0).
4485 * Updates the cursor index with the index of the found entry.
4486 * If no entry larger or equal to the key is found, returns NULL.
4489 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4491 unsigned int i = 0, nkeys;
4494 MDB_page *mp = mc->mc_pg[mc->mc_top];
4495 MDB_node *node = NULL;
4500 nkeys = NUMKEYS(mp);
4505 COPY_PGNO(pgno, mp->mp_pgno);
4506 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4507 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4512 low = IS_LEAF(mp) ? 0 : 1;
4514 cmp = mc->mc_dbx->md_cmp;
4516 /* Branch pages have no data, so if using integer keys,
4517 * alignment is guaranteed. Use faster mdb_cmp_int.
4519 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4520 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4527 nodekey.mv_size = mc->mc_db->md_pad;
4528 node = NODEPTR(mp, 0); /* fake */
4529 while (low <= high) {
4530 i = (low + high) >> 1;
4531 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4532 rc = cmp(key, &nodekey);
4533 DPRINTF(("found leaf index %u [%s], rc = %i",
4534 i, DKEY(&nodekey), rc));
4543 while (low <= high) {
4544 i = (low + high) >> 1;
4546 node = NODEPTR(mp, i);
4547 nodekey.mv_size = NODEKSZ(node);
4548 nodekey.mv_data = NODEKEY(node);
4550 rc = cmp(key, &nodekey);
4553 DPRINTF(("found leaf index %u [%s], rc = %i",
4554 i, DKEY(&nodekey), rc));
4556 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4557 i, DKEY(&nodekey), NODEPGNO(node), rc));
4568 if (rc > 0) { /* Found entry is less than the key. */
4569 i++; /* Skip to get the smallest entry larger than key. */
4571 node = NODEPTR(mp, i);
4574 *exactp = (rc == 0 && nkeys > 0);
4575 /* store the key index */
4576 mc->mc_ki[mc->mc_top] = i;
4578 /* There is no entry larger or equal to the key. */
4581 /* nodeptr is fake for LEAF2 */
4587 mdb_cursor_adjust(MDB_cursor *mc, func)
4591 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4592 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4599 /** Pop a page off the top of the cursor's stack. */
4601 mdb_cursor_pop(MDB_cursor *mc)
4605 MDB_page *top = mc->mc_pg[mc->mc_top];
4611 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4612 DDBI(mc), (void *) mc));
4616 /** Push a page onto the top of the cursor's stack. */
4618 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4620 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4621 DDBI(mc), (void *) mc));
4623 if (mc->mc_snum >= CURSOR_STACK) {
4624 assert(mc->mc_snum < CURSOR_STACK);
4625 return MDB_CURSOR_FULL;
4628 mc->mc_top = mc->mc_snum++;
4629 mc->mc_pg[mc->mc_top] = mp;
4630 mc->mc_ki[mc->mc_top] = 0;
4635 /** Find the address of the page corresponding to a given page number.
4636 * @param[in] txn the transaction for this access.
4637 * @param[in] pgno the page number for the page to retrieve.
4638 * @param[out] ret address of a pointer where the page's address will be stored.
4639 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4640 * @return 0 on success, non-zero on failure.
4643 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4645 MDB_env *env = txn->mt_env;
4649 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4653 MDB_ID2L dl = tx2->mt_u.dirty_list;
4655 /* Spilled pages were dirtied in this txn and flushed
4656 * because the dirty list got full. Bring this page
4657 * back in from the map (but don't unspill it here,
4658 * leave that unless page_touch happens again).
4660 if (tx2->mt_spill_pgs) {
4661 MDB_ID pn = pgno << 1;
4662 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4663 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4664 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4669 unsigned x = mdb_mid2l_search(dl, pgno);
4670 if (x <= dl[0].mid && dl[x].mid == pgno) {
4676 } while ((tx2 = tx2->mt_parent) != NULL);
4679 if (pgno < txn->mt_next_pgno) {
4681 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4683 DPRINTF(("page %"Z"u not found", pgno));
4685 return MDB_PAGE_NOTFOUND;
4695 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4696 * The cursor is at the root page, set up the rest of it.
4699 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4701 MDB_page *mp = mc->mc_pg[mc->mc_top];
4705 while (IS_BRANCH(mp)) {
4709 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4710 assert(NUMKEYS(mp) > 1);
4711 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4713 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4715 if (flags & MDB_PS_LAST)
4716 i = NUMKEYS(mp) - 1;
4719 node = mdb_node_search(mc, key, &exact);
4721 i = NUMKEYS(mp) - 1;
4723 i = mc->mc_ki[mc->mc_top];
4729 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4732 assert(i < NUMKEYS(mp));
4733 node = NODEPTR(mp, i);
4735 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4738 mc->mc_ki[mc->mc_top] = i;
4739 if ((rc = mdb_cursor_push(mc, mp)))
4742 if (flags & MDB_PS_MODIFY) {
4743 if ((rc = mdb_page_touch(mc)) != 0)
4745 mp = mc->mc_pg[mc->mc_top];
4750 DPRINTF(("internal error, index points to a %02X page!?",
4752 return MDB_CORRUPTED;
4755 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4756 key ? DKEY(key) : "null"));
4757 mc->mc_flags |= C_INITIALIZED;
4758 mc->mc_flags &= ~C_EOF;
4763 /** Search for the lowest key under the current branch page.
4764 * This just bypasses a NUMKEYS check in the current page
4765 * before calling mdb_page_search_root(), because the callers
4766 * are all in situations where the current page is known to
4770 mdb_page_search_lowest(MDB_cursor *mc)
4772 MDB_page *mp = mc->mc_pg[mc->mc_top];
4773 MDB_node *node = NODEPTR(mp, 0);
4776 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4779 mc->mc_ki[mc->mc_top] = 0;
4780 if ((rc = mdb_cursor_push(mc, mp)))
4782 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4785 /** Search for the page a given key should be in.
4786 * Push it and its parent pages on the cursor stack.
4787 * @param[in,out] mc the cursor for this operation.
4788 * @param[in] key the key to search for, or NULL for first/last page.
4789 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4790 * are touched (updated with new page numbers).
4791 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4792 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4793 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4794 * @return 0 on success, non-zero on failure.
4797 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4802 /* Make sure the txn is still viable, then find the root from
4803 * the txn's db table and set it as the root of the cursor's stack.
4805 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4806 DPUTS("transaction has failed, must abort");
4809 /* Make sure we're using an up-to-date root */
4810 if (*mc->mc_dbflag & DB_STALE) {
4812 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4813 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4820 MDB_node *leaf = mdb_node_search(&mc2,
4821 &mc->mc_dbx->md_name, &exact);
4823 return MDB_NOTFOUND;
4824 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4827 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4829 /* The txn may not know this DBI, or another process may
4830 * have dropped and recreated the DB with other flags.
4832 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4833 return MDB_INCOMPATIBLE;
4834 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4836 *mc->mc_dbflag &= ~DB_STALE;
4838 root = mc->mc_db->md_root;
4840 if (root == P_INVALID) { /* Tree is empty. */
4841 DPUTS("tree is empty");
4842 return MDB_NOTFOUND;
4847 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4848 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4854 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4855 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4857 if (flags & MDB_PS_MODIFY) {
4858 if ((rc = mdb_page_touch(mc)))
4862 if (flags & MDB_PS_ROOTONLY)
4865 return mdb_page_search_root(mc, key, flags);
4869 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4871 MDB_txn *txn = mc->mc_txn;
4872 pgno_t pg = mp->mp_pgno;
4873 unsigned x = 0, ovpages = mp->mp_pages;
4874 MDB_env *env = txn->mt_env;
4875 MDB_IDL sl = txn->mt_spill_pgs;
4876 MDB_ID pn = pg << 1;
4879 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4880 /* If the page is dirty or on the spill list we just acquired it,
4881 * so we should give it back to our current free list, if any.
4882 * Otherwise put it onto the list of pages we freed in this txn.
4884 * Won't create me_pghead: me_pglast must be inited along with it.
4885 * Unsupported in nested txns: They would need to hide the page
4886 * range in ancestor txns' dirty and spilled lists.
4888 if (env->me_pghead &&
4890 ((mp->mp_flags & P_DIRTY) ||
4891 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4895 MDB_ID2 *dl, ix, iy;
4896 rc = mdb_midl_need(&env->me_pghead, ovpages);
4899 if (!(mp->mp_flags & P_DIRTY)) {
4900 /* This page is no longer spilled */
4907 /* Remove from dirty list */
4908 dl = txn->mt_u.dirty_list;
4910 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4918 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4919 txn->mt_flags |= MDB_TXN_ERROR;
4920 return MDB_CORRUPTED;
4923 if (!(env->me_flags & MDB_WRITEMAP))
4924 mdb_dpage_free(env, mp);
4926 /* Insert in me_pghead */
4927 mop = env->me_pghead;
4928 j = mop[0] + ovpages;
4929 for (i = mop[0]; i && mop[i] < pg; i--)
4935 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4939 mc->mc_db->md_overflow_pages -= ovpages;
4943 /** Return the data associated with a given node.
4944 * @param[in] txn The transaction for this operation.
4945 * @param[in] leaf The node being read.
4946 * @param[out] data Updated to point to the node's data.
4947 * @return 0 on success, non-zero on failure.
4950 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4952 MDB_page *omp; /* overflow page */
4956 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4957 data->mv_size = NODEDSZ(leaf);
4958 data->mv_data = NODEDATA(leaf);
4962 /* Read overflow data.
4964 data->mv_size = NODEDSZ(leaf);
4965 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4966 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4967 DPRINTF(("read overflow page %"Z"u failed", pgno));
4970 data->mv_data = METADATA(omp);
4976 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4977 MDB_val *key, MDB_val *data)
4986 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4988 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4991 if (txn->mt_flags & MDB_TXN_ERROR)
4994 mdb_cursor_init(&mc, txn, dbi, &mx);
4995 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4998 /** Find a sibling for a page.
4999 * Replaces the page at the top of the cursor's stack with the
5000 * specified sibling, if one exists.
5001 * @param[in] mc The cursor for this operation.
5002 * @param[in] move_right Non-zero if the right sibling is requested,
5003 * otherwise the left sibling.
5004 * @return 0 on success, non-zero on failure.
5007 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5013 if (mc->mc_snum < 2) {
5014 return MDB_NOTFOUND; /* root has no siblings */
5018 DPRINTF(("parent page is page %"Z"u, index %u",
5019 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5021 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5022 : (mc->mc_ki[mc->mc_top] == 0)) {
5023 DPRINTF(("no more keys left, moving to %s sibling",
5024 move_right ? "right" : "left"));
5025 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5026 /* undo cursor_pop before returning */
5033 mc->mc_ki[mc->mc_top]++;
5035 mc->mc_ki[mc->mc_top]--;
5036 DPRINTF(("just moving to %s index key %u",
5037 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5039 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
5041 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5042 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5043 /* mc will be inconsistent if caller does mc_snum++ as above */
5044 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5048 mdb_cursor_push(mc, mp);
5050 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5055 /** Move the cursor to the next data item. */
5057 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5063 if (mc->mc_flags & C_EOF) {
5064 return MDB_NOTFOUND;
5067 assert(mc->mc_flags & C_INITIALIZED);
5069 mp = mc->mc_pg[mc->mc_top];
5071 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5072 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5073 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5074 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5075 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5076 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5077 if (rc == MDB_SUCCESS)
5078 MDB_GET_KEY(leaf, key);
5083 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5084 if (op == MDB_NEXT_DUP)
5085 return MDB_NOTFOUND;
5089 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5090 if (mc->mc_flags & C_DEL)
5093 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5094 DPUTS("=====> move to next sibling page");
5095 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5096 mc->mc_flags |= C_EOF;
5099 mp = mc->mc_pg[mc->mc_top];
5100 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5102 mc->mc_ki[mc->mc_top]++;
5105 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5106 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5109 key->mv_size = mc->mc_db->md_pad;
5110 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5114 assert(IS_LEAF(mp));
5115 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5117 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5118 mdb_xcursor_init1(mc, leaf);
5121 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5124 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5125 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5126 if (rc != MDB_SUCCESS)
5131 MDB_GET_KEY(leaf, key);
5135 /** Move the cursor to the previous data item. */
5137 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5143 assert(mc->mc_flags & C_INITIALIZED);
5145 mp = mc->mc_pg[mc->mc_top];
5147 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5148 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5149 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5150 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5151 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5152 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5153 if (rc == MDB_SUCCESS)
5154 MDB_GET_KEY(leaf, key);
5158 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5159 if (op == MDB_PREV_DUP)
5160 return MDB_NOTFOUND;
5165 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5167 if (mc->mc_ki[mc->mc_top] == 0) {
5168 DPUTS("=====> move to prev sibling page");
5169 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5172 mp = mc->mc_pg[mc->mc_top];
5173 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5174 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5176 mc->mc_ki[mc->mc_top]--;
5178 mc->mc_flags &= ~C_EOF;
5180 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5181 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5184 key->mv_size = mc->mc_db->md_pad;
5185 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5189 assert(IS_LEAF(mp));
5190 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5192 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5193 mdb_xcursor_init1(mc, leaf);
5196 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5199 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5200 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5201 if (rc != MDB_SUCCESS)
5206 MDB_GET_KEY(leaf, key);
5210 /** Set the cursor on a specific data item. */
5212 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5213 MDB_cursor_op op, int *exactp)
5217 MDB_node *leaf = NULL;
5222 if (key->mv_size == 0)
5223 return MDB_BAD_VALSIZE;
5226 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5228 /* See if we're already on the right page */
5229 if (mc->mc_flags & C_INITIALIZED) {
5232 mp = mc->mc_pg[mc->mc_top];
5234 mc->mc_ki[mc->mc_top] = 0;
5235 return MDB_NOTFOUND;
5237 if (mp->mp_flags & P_LEAF2) {
5238 nodekey.mv_size = mc->mc_db->md_pad;
5239 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5241 leaf = NODEPTR(mp, 0);
5242 MDB_GET_KEY2(leaf, nodekey);
5244 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5246 /* Probably happens rarely, but first node on the page
5247 * was the one we wanted.
5249 mc->mc_ki[mc->mc_top] = 0;
5256 unsigned int nkeys = NUMKEYS(mp);
5258 if (mp->mp_flags & P_LEAF2) {
5259 nodekey.mv_data = LEAF2KEY(mp,
5260 nkeys-1, nodekey.mv_size);
5262 leaf = NODEPTR(mp, nkeys-1);
5263 MDB_GET_KEY2(leaf, nodekey);
5265 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5267 /* last node was the one we wanted */
5268 mc->mc_ki[mc->mc_top] = nkeys-1;
5274 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5275 /* This is definitely the right page, skip search_page */
5276 if (mp->mp_flags & P_LEAF2) {
5277 nodekey.mv_data = LEAF2KEY(mp,
5278 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5280 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5281 MDB_GET_KEY2(leaf, nodekey);
5283 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5285 /* current node was the one we wanted */
5295 /* If any parents have right-sibs, search.
5296 * Otherwise, there's nothing further.
5298 for (i=0; i<mc->mc_top; i++)
5300 NUMKEYS(mc->mc_pg[i])-1)
5302 if (i == mc->mc_top) {
5303 /* There are no other pages */
5304 mc->mc_ki[mc->mc_top] = nkeys;
5305 return MDB_NOTFOUND;
5309 /* There are no other pages */
5310 mc->mc_ki[mc->mc_top] = 0;
5311 if (op == MDB_SET_RANGE) {
5315 return MDB_NOTFOUND;
5319 rc = mdb_page_search(mc, key, 0);
5320 if (rc != MDB_SUCCESS)
5323 mp = mc->mc_pg[mc->mc_top];
5324 assert(IS_LEAF(mp));
5327 leaf = mdb_node_search(mc, key, exactp);
5328 if (exactp != NULL && !*exactp) {
5329 /* MDB_SET specified and not an exact match. */
5330 return MDB_NOTFOUND;
5334 DPUTS("===> inexact leaf not found, goto sibling");
5335 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5336 return rc; /* no entries matched */
5337 mp = mc->mc_pg[mc->mc_top];
5338 assert(IS_LEAF(mp));
5339 leaf = NODEPTR(mp, 0);
5343 mc->mc_flags |= C_INITIALIZED;
5344 mc->mc_flags &= ~C_EOF;
5347 key->mv_size = mc->mc_db->md_pad;
5348 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5352 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5353 mdb_xcursor_init1(mc, leaf);
5356 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5357 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5358 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5361 if (op == MDB_GET_BOTH) {
5367 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5368 if (rc != MDB_SUCCESS)
5371 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5373 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5375 rc = mc->mc_dbx->md_dcmp(data, &d2);
5377 if (op == MDB_GET_BOTH || rc > 0)
5378 return MDB_NOTFOUND;
5385 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5386 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5391 /* The key already matches in all other cases */
5392 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5393 MDB_GET_KEY(leaf, key);
5394 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5399 /** Move the cursor to the first item in the database. */
5401 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5407 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5409 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5410 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5411 if (rc != MDB_SUCCESS)
5414 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5416 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5417 mc->mc_flags |= C_INITIALIZED;
5418 mc->mc_flags &= ~C_EOF;
5420 mc->mc_ki[mc->mc_top] = 0;
5422 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5423 key->mv_size = mc->mc_db->md_pad;
5424 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5429 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5430 mdb_xcursor_init1(mc, leaf);
5431 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5435 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5439 MDB_GET_KEY(leaf, key);
5443 /** Move the cursor to the last item in the database. */
5445 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5451 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5453 if (!(mc->mc_flags & C_EOF)) {
5455 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5456 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5457 if (rc != MDB_SUCCESS)
5460 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5463 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5464 mc->mc_flags |= C_INITIALIZED|C_EOF;
5465 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5467 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5468 key->mv_size = mc->mc_db->md_pad;
5469 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5474 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5475 mdb_xcursor_init1(mc, leaf);
5476 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5480 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5485 MDB_GET_KEY(leaf, key);
5490 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5495 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5499 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5503 case MDB_GET_CURRENT:
5504 if (!(mc->mc_flags & C_INITIALIZED)) {
5507 MDB_page *mp = mc->mc_pg[mc->mc_top];
5508 int nkeys = NUMKEYS(mp);
5509 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5510 mc->mc_ki[mc->mc_top] = nkeys;
5516 key->mv_size = mc->mc_db->md_pad;
5517 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5519 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5520 MDB_GET_KEY(leaf, key);
5522 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5523 if (mc->mc_flags & C_DEL)
5524 mdb_xcursor_init1(mc, leaf);
5525 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5527 rc = mdb_node_read(mc->mc_txn, leaf, data);
5534 case MDB_GET_BOTH_RANGE:
5539 if (mc->mc_xcursor == NULL) {
5540 rc = MDB_INCOMPATIBLE;
5550 rc = mdb_cursor_set(mc, key, data, op,
5551 op == MDB_SET_RANGE ? NULL : &exact);
5554 case MDB_GET_MULTIPLE:
5555 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5559 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5560 rc = MDB_INCOMPATIBLE;
5564 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5565 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5568 case MDB_NEXT_MULTIPLE:
5573 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5574 rc = MDB_INCOMPATIBLE;
5577 if (!(mc->mc_flags & C_INITIALIZED))
5578 rc = mdb_cursor_first(mc, key, data);
5580 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5581 if (rc == MDB_SUCCESS) {
5582 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5585 mx = &mc->mc_xcursor->mx_cursor;
5586 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5588 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5589 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5597 case MDB_NEXT_NODUP:
5598 if (!(mc->mc_flags & C_INITIALIZED))
5599 rc = mdb_cursor_first(mc, key, data);
5601 rc = mdb_cursor_next(mc, key, data, op);
5605 case MDB_PREV_NODUP:
5606 if (!(mc->mc_flags & C_INITIALIZED)) {
5607 rc = mdb_cursor_last(mc, key, data);
5610 mc->mc_flags |= C_INITIALIZED;
5611 mc->mc_ki[mc->mc_top]++;
5613 rc = mdb_cursor_prev(mc, key, data, op);
5616 rc = mdb_cursor_first(mc, key, data);
5619 mfunc = mdb_cursor_first;
5621 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5625 if (mc->mc_xcursor == NULL) {
5626 rc = MDB_INCOMPATIBLE;
5629 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5633 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5636 rc = mdb_cursor_last(mc, key, data);
5639 mfunc = mdb_cursor_last;
5642 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5647 if (mc->mc_flags & C_DEL)
5648 mc->mc_flags ^= C_DEL;
5653 /** Touch all the pages in the cursor stack. Set mc_top.
5654 * Makes sure all the pages are writable, before attempting a write operation.
5655 * @param[in] mc The cursor to operate on.
5658 mdb_cursor_touch(MDB_cursor *mc)
5660 int rc = MDB_SUCCESS;
5662 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5665 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5666 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5669 *mc->mc_dbflag |= DB_DIRTY;
5674 rc = mdb_page_touch(mc);
5675 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5676 mc->mc_top = mc->mc_snum-1;
5681 /** Do not spill pages to disk if txn is getting full, may fail instead */
5682 #define MDB_NOSPILL 0x8000
5685 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5688 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5689 MDB_env *env = mc->mc_txn->mt_env;
5690 MDB_node *leaf = NULL;
5693 MDB_val xdata, *rdata, dkey, olddata;
5695 int do_sub = 0, insert;
5696 unsigned int mcount = 0, dcount = 0, nospill;
5699 unsigned int nflags;
5702 /* Check this first so counter will always be zero on any
5705 if (flags & MDB_MULTIPLE) {
5706 dcount = data[1].mv_size;
5707 data[1].mv_size = 0;
5708 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5709 return MDB_INCOMPATIBLE;
5712 nospill = flags & MDB_NOSPILL;
5713 flags &= ~MDB_NOSPILL;
5715 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5716 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5718 if (flags != MDB_CURRENT && key->mv_size-1 >= ENV_MAXKEY(env))
5719 return MDB_BAD_VALSIZE;
5721 #if SIZE_MAX > MAXDATASIZE
5722 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5723 return MDB_BAD_VALSIZE;
5725 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5726 return MDB_BAD_VALSIZE;
5729 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5730 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5734 if (flags == MDB_CURRENT) {
5735 if (!(mc->mc_flags & C_INITIALIZED))
5738 } else if (mc->mc_db->md_root == P_INVALID) {
5739 /* new database, cursor has nothing to point to */
5742 mc->mc_flags &= ~C_INITIALIZED;
5747 if (flags & MDB_APPEND) {
5749 rc = mdb_cursor_last(mc, &k2, &d2);
5751 rc = mc->mc_dbx->md_cmp(key, &k2);
5754 mc->mc_ki[mc->mc_top]++;
5756 /* new key is <= last key */
5761 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5763 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5764 DPRINTF(("duplicate key [%s]", DKEY(key)));
5766 return MDB_KEYEXIST;
5768 if (rc && rc != MDB_NOTFOUND)
5772 if (mc->mc_flags & C_DEL)
5773 mc->mc_flags ^= C_DEL;
5775 /* Cursor is positioned, check for room in the dirty list */
5777 if (flags & MDB_MULTIPLE) {
5779 xdata.mv_size = data->mv_size * dcount;
5783 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5787 if (rc == MDB_NO_ROOT) {
5789 /* new database, write a root leaf page */
5790 DPUTS("allocating new root leaf page");
5791 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5794 mdb_cursor_push(mc, np);
5795 mc->mc_db->md_root = np->mp_pgno;
5796 mc->mc_db->md_depth++;
5797 *mc->mc_dbflag |= DB_DIRTY;
5798 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5800 np->mp_flags |= P_LEAF2;
5801 mc->mc_flags |= C_INITIALIZED;
5803 /* make sure all cursor pages are writable */
5804 rc2 = mdb_cursor_touch(mc);
5811 /* The key does not exist */
5812 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5813 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5814 LEAFSIZE(key, data) > env->me_nodemax)
5816 /* Too big for a node, insert in sub-DB */
5817 fp_flags = P_LEAF|P_DIRTY;
5819 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5820 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
5824 /* there's only a key anyway, so this is a no-op */
5825 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5826 unsigned int ksize = mc->mc_db->md_pad;
5827 if (key->mv_size != ksize)
5828 return MDB_BAD_VALSIZE;
5829 if (flags == MDB_CURRENT) {
5830 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5831 memcpy(ptr, key->mv_data, ksize);
5837 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5838 olddata.mv_size = NODEDSZ(leaf);
5839 olddata.mv_data = NODEDATA(leaf);
5842 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5843 /* Prepare (sub-)page/sub-DB to accept the new item,
5844 * if needed. fp: old sub-page or a header faking
5845 * it. mp: new (sub-)page. offset: growth in page
5846 * size. xdata: node data with new page or DB.
5848 ssize_t i, offset = 0;
5849 mp = fp = xdata.mv_data = env->me_pbuf;
5850 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5852 /* Was a single item before, must convert now */
5853 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5854 /* Just overwrite the current item */
5855 if (flags == MDB_CURRENT)
5858 #if UINT_MAX < SIZE_MAX
5859 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5860 #ifdef MISALIGNED_OK
5861 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5863 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5866 /* if data matches, skip it */
5867 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5868 if (flags & MDB_NODUPDATA)
5870 else if (flags & MDB_MULTIPLE)
5877 /* Back up original data item */
5878 dkey.mv_size = olddata.mv_size;
5879 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5881 /* Make sub-page header for the dup items, with dummy body */
5882 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5883 fp->mp_lower = PAGEHDRSZ;
5884 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5885 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5886 fp->mp_flags |= P_LEAF2;
5887 fp->mp_pad = data->mv_size;
5888 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
5890 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
5891 (dkey.mv_size & 1) + (data->mv_size & 1);
5893 fp->mp_upper = xdata.mv_size;
5894 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
5895 } else if (leaf->mn_flags & F_SUBDATA) {
5896 /* Data is on sub-DB, just store it */
5897 flags |= F_DUPDATA|F_SUBDATA;
5900 /* Data is on sub-page */
5901 fp = olddata.mv_data;
5904 i = -(ssize_t)SIZELEFT(fp);
5905 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5906 offset = i += (ssize_t) EVEN(
5907 sizeof(indx_t) + NODESIZE + data->mv_size);
5909 i += offset = fp->mp_pad;
5910 offset *= 4; /* space for 4 more */
5914 /* FALLTHRU: Sub-page is big enough */
5916 fp->mp_flags |= P_DIRTY;
5917 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
5918 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5922 xdata.mv_size = olddata.mv_size + offset;
5925 fp_flags = fp->mp_flags;
5926 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
5927 /* Too big for a sub-page, convert to sub-DB */
5928 fp_flags &= ~P_SUBP;
5930 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5931 fp_flags |= P_LEAF2;
5932 dummy.md_pad = fp->mp_pad;
5933 dummy.md_flags = MDB_DUPFIXED;
5934 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5935 dummy.md_flags |= MDB_INTEGERKEY;
5941 dummy.md_branch_pages = 0;
5942 dummy.md_leaf_pages = 1;
5943 dummy.md_overflow_pages = 0;
5944 dummy.md_entries = NUMKEYS(fp);
5945 xdata.mv_size = sizeof(MDB_db);
5946 xdata.mv_data = &dummy;
5947 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5949 offset = env->me_psize - olddata.mv_size;
5950 flags |= F_DUPDATA|F_SUBDATA;
5951 dummy.md_root = mp->mp_pgno;
5954 mp->mp_flags = fp_flags | P_DIRTY;
5955 mp->mp_pad = fp->mp_pad;
5956 mp->mp_lower = fp->mp_lower;
5957 mp->mp_upper = fp->mp_upper + offset;
5958 if (fp_flags & P_LEAF2) {
5959 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5961 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
5962 olddata.mv_size - fp->mp_upper);
5963 for (i = NUMKEYS(fp); --i >= 0; )
5964 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5972 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5976 /* overflow page overwrites need special handling */
5977 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5980 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
5982 memcpy(&pg, olddata.mv_data, sizeof(pg));
5983 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5985 ovpages = omp->mp_pages;
5987 /* Is the ov page large enough? */
5988 if (ovpages >= dpages) {
5989 if (!(omp->mp_flags & P_DIRTY) &&
5990 (level || (env->me_flags & MDB_WRITEMAP)))
5992 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5995 level = 0; /* dirty in this txn or clean */
5998 if (omp->mp_flags & P_DIRTY) {
5999 /* yes, overwrite it. Note in this case we don't
6000 * bother to try shrinking the page if the new data
6001 * is smaller than the overflow threshold.
6004 /* It is writable only in a parent txn */
6005 size_t sz = (size_t) env->me_psize * ovpages, off;
6006 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6012 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6013 if (!(flags & MDB_RESERVE)) {
6014 /* Copy end of page, adjusting alignment so
6015 * compiler may copy words instead of bytes.
6017 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6018 memcpy((size_t *)((char *)np + off),
6019 (size_t *)((char *)omp + off), sz - off);
6022 memcpy(np, omp, sz); /* Copy beginning of page */
6025 SETDSZ(leaf, data->mv_size);
6026 if (F_ISSET(flags, MDB_RESERVE))
6027 data->mv_data = METADATA(omp);
6029 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6033 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6035 } else if (data->mv_size == olddata.mv_size) {
6036 /* same size, just replace it. Note that we could
6037 * also reuse this node if the new data is smaller,
6038 * but instead we opt to shrink the node in that case.
6040 if (F_ISSET(flags, MDB_RESERVE))
6041 data->mv_data = olddata.mv_data;
6042 else if (data->mv_size)
6043 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6045 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6048 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6049 mc->mc_db->md_entries--;
6055 nflags = flags & NODE_ADD_FLAGS;
6056 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6057 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6058 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6059 nflags &= ~MDB_APPEND;
6061 nflags |= MDB_SPLIT_REPLACE;
6062 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6064 /* There is room already in this leaf page. */
6065 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6066 if (rc == 0 && !do_sub && insert) {
6067 /* Adjust other cursors pointing to mp */
6068 MDB_cursor *m2, *m3;
6069 MDB_dbi dbi = mc->mc_dbi;
6070 unsigned i = mc->mc_top;
6071 MDB_page *mp = mc->mc_pg[i];
6073 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6074 if (mc->mc_flags & C_SUB)
6075 m3 = &m2->mc_xcursor->mx_cursor;
6078 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6079 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6086 if (rc != MDB_SUCCESS)
6087 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6089 /* Now store the actual data in the child DB. Note that we're
6090 * storing the user data in the keys field, so there are strict
6091 * size limits on dupdata. The actual data fields of the child
6092 * DB are all zero size.
6099 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6100 if (flags & MDB_CURRENT) {
6101 xflags = MDB_CURRENT|MDB_NOSPILL;
6103 mdb_xcursor_init1(mc, leaf);
6104 xflags = (flags & MDB_NODUPDATA) ?
6105 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6107 /* converted, write the original data first */
6109 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6113 /* Adjust other cursors pointing to mp */
6115 unsigned i = mc->mc_top;
6116 MDB_page *mp = mc->mc_pg[i];
6118 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6119 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6120 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6121 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6122 mdb_xcursor_init1(m2, leaf);
6126 /* we've done our job */
6129 if (flags & MDB_APPENDDUP)
6130 xflags |= MDB_APPEND;
6131 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6132 if (flags & F_SUBDATA) {
6133 void *db = NODEDATA(leaf);
6134 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6137 /* sub-writes might have failed so check rc again.
6138 * Don't increment count if we just replaced an existing item.
6140 if (!rc && !(flags & MDB_CURRENT))
6141 mc->mc_db->md_entries++;
6142 if (flags & MDB_MULTIPLE) {
6146 /* let caller know how many succeeded, if any */
6147 data[1].mv_size = mcount;
6148 if (mcount < dcount) {
6149 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6156 /* If we succeeded and the key didn't exist before, make sure
6157 * the cursor is marked valid.
6160 mc->mc_flags |= C_INITIALIZED;
6165 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6171 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6172 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6174 if (!(mc->mc_flags & C_INITIALIZED))
6177 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6178 return MDB_NOTFOUND;
6180 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6183 rc = mdb_cursor_touch(mc);
6187 mp = mc->mc_pg[mc->mc_top];
6188 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6190 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6191 if (!(flags & MDB_NODUPDATA)) {
6192 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6193 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6195 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6196 /* If sub-DB still has entries, we're done */
6197 if (mc->mc_xcursor->mx_db.md_entries) {
6198 if (leaf->mn_flags & F_SUBDATA) {
6199 /* update subDB info */
6200 void *db = NODEDATA(leaf);
6201 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6204 /* shrink fake page */
6205 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6206 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6207 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6208 /* fix other sub-DB cursors pointed at this fake page */
6209 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6210 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6211 if (m2->mc_pg[mc->mc_top] == mp &&
6212 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6213 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6216 mc->mc_db->md_entries--;
6217 mc->mc_flags |= C_DEL;
6220 /* otherwise fall thru and delete the sub-DB */
6223 if (leaf->mn_flags & F_SUBDATA) {
6224 /* add all the child DB's pages to the free list */
6225 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6226 if (rc == MDB_SUCCESS) {
6227 mc->mc_db->md_entries -=
6228 mc->mc_xcursor->mx_db.md_entries;
6233 return mdb_cursor_del0(mc, leaf);
6236 /** Allocate and initialize new pages for a database.
6237 * @param[in] mc a cursor on the database being added to.
6238 * @param[in] flags flags defining what type of page is being allocated.
6239 * @param[in] num the number of pages to allocate. This is usually 1,
6240 * unless allocating overflow pages for a large record.
6241 * @param[out] mp Address of a page, or NULL on failure.
6242 * @return 0 on success, non-zero on failure.
6245 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6250 if ((rc = mdb_page_alloc(mc, num, &np)))
6252 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6253 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6254 np->mp_flags = flags | P_DIRTY;
6255 np->mp_lower = PAGEHDRSZ;
6256 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6259 mc->mc_db->md_branch_pages++;
6260 else if (IS_LEAF(np))
6261 mc->mc_db->md_leaf_pages++;
6262 else if (IS_OVERFLOW(np)) {
6263 mc->mc_db->md_overflow_pages += num;
6271 /** Calculate the size of a leaf node.
6272 * The size depends on the environment's page size; if a data item
6273 * is too large it will be put onto an overflow page and the node
6274 * size will only include the key and not the data. Sizes are always
6275 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6276 * of the #MDB_node headers.
6277 * @param[in] env The environment handle.
6278 * @param[in] key The key for the node.
6279 * @param[in] data The data for the node.
6280 * @return The number of bytes needed to store the node.
6283 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6287 sz = LEAFSIZE(key, data);
6288 if (sz > env->me_nodemax) {
6289 /* put on overflow page */
6290 sz -= data->mv_size - sizeof(pgno_t);
6293 return EVEN(sz + sizeof(indx_t));
6296 /** Calculate the size of a branch node.
6297 * The size should depend on the environment's page size but since
6298 * we currently don't support spilling large keys onto overflow
6299 * pages, it's simply the size of the #MDB_node header plus the
6300 * size of the key. Sizes are always rounded up to an even number
6301 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6302 * @param[in] env The environment handle.
6303 * @param[in] key The key for the node.
6304 * @return The number of bytes needed to store the node.
6307 mdb_branch_size(MDB_env *env, MDB_val *key)
6312 if (sz > env->me_nodemax) {
6313 /* put on overflow page */
6314 /* not implemented */
6315 /* sz -= key->size - sizeof(pgno_t); */
6318 return sz + sizeof(indx_t);
6321 /** Add a node to the page pointed to by the cursor.
6322 * @param[in] mc The cursor for this operation.
6323 * @param[in] indx The index on the page where the new node should be added.
6324 * @param[in] key The key for the new node.
6325 * @param[in] data The data for the new node, if any.
6326 * @param[in] pgno The page number, if adding a branch node.
6327 * @param[in] flags Flags for the node.
6328 * @return 0 on success, non-zero on failure. Possible errors are:
6330 * <li>ENOMEM - failed to allocate overflow pages for the node.
6331 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6332 * should never happen since all callers already calculate the
6333 * page's free space before calling this function.
6337 mdb_node_add(MDB_cursor *mc, indx_t indx,
6338 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6341 size_t node_size = NODESIZE;
6345 MDB_page *mp = mc->mc_pg[mc->mc_top];
6346 MDB_page *ofp = NULL; /* overflow page */
6349 assert(mp->mp_upper >= mp->mp_lower);
6351 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6352 IS_LEAF(mp) ? "leaf" : "branch",
6353 IS_SUBP(mp) ? "sub-" : "",
6354 mp->mp_pgno, indx, data ? data->mv_size : 0,
6355 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6358 /* Move higher keys up one slot. */
6359 int ksize = mc->mc_db->md_pad, dif;
6360 char *ptr = LEAF2KEY(mp, indx, ksize);
6361 dif = NUMKEYS(mp) - indx;
6363 memmove(ptr+ksize, ptr, dif*ksize);
6364 /* insert new key */
6365 memcpy(ptr, key->mv_data, ksize);
6367 /* Just using these for counting */
6368 mp->mp_lower += sizeof(indx_t);
6369 mp->mp_upper -= ksize - sizeof(indx_t);
6373 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6375 node_size += key->mv_size;
6378 if (F_ISSET(flags, F_BIGDATA)) {
6379 /* Data already on overflow page. */
6380 node_size += sizeof(pgno_t);
6381 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6382 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6384 /* Put data on overflow page. */
6385 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6386 data->mv_size, node_size+data->mv_size));
6387 node_size = EVEN(node_size + sizeof(pgno_t));
6388 if ((ssize_t)node_size > room)
6390 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6392 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6396 node_size += data->mv_size;
6399 node_size = EVEN(node_size);
6400 if ((ssize_t)node_size > room)
6404 /* Move higher pointers up one slot. */
6405 for (i = NUMKEYS(mp); i > indx; i--)
6406 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6408 /* Adjust free space offsets. */
6409 ofs = mp->mp_upper - node_size;
6410 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6411 mp->mp_ptrs[indx] = ofs;
6413 mp->mp_lower += sizeof(indx_t);
6415 /* Write the node data. */
6416 node = NODEPTR(mp, indx);
6417 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6418 node->mn_flags = flags;
6420 SETDSZ(node,data->mv_size);
6425 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6430 if (F_ISSET(flags, F_BIGDATA))
6431 memcpy(node->mn_data + key->mv_size, data->mv_data,
6433 else if (F_ISSET(flags, MDB_RESERVE))
6434 data->mv_data = node->mn_data + key->mv_size;
6436 memcpy(node->mn_data + key->mv_size, data->mv_data,
6439 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6441 if (F_ISSET(flags, MDB_RESERVE))
6442 data->mv_data = METADATA(ofp);
6444 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6451 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6452 mp->mp_pgno, NUMKEYS(mp)));
6453 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6454 DPRINTF(("node size = %"Z"u", node_size));
6455 return MDB_PAGE_FULL;
6458 /** Delete the specified node from a page.
6459 * @param[in] mp The page to operate on.
6460 * @param[in] indx The index of the node to delete.
6461 * @param[in] ksize The size of a node. Only used if the page is
6462 * part of a #MDB_DUPFIXED database.
6465 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6468 indx_t i, j, numkeys, ptr;
6475 COPY_PGNO(pgno, mp->mp_pgno);
6476 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6477 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6480 assert(indx < NUMKEYS(mp));
6483 int x = NUMKEYS(mp) - 1 - indx;
6484 base = LEAF2KEY(mp, indx, ksize);
6486 memmove(base, base + ksize, x * ksize);
6487 mp->mp_lower -= sizeof(indx_t);
6488 mp->mp_upper += ksize - sizeof(indx_t);
6492 node = NODEPTR(mp, indx);
6493 sz = NODESIZE + node->mn_ksize;
6495 if (F_ISSET(node->mn_flags, F_BIGDATA))
6496 sz += sizeof(pgno_t);
6498 sz += NODEDSZ(node);
6502 ptr = mp->mp_ptrs[indx];
6503 numkeys = NUMKEYS(mp);
6504 for (i = j = 0; i < numkeys; i++) {
6506 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6507 if (mp->mp_ptrs[i] < ptr)
6508 mp->mp_ptrs[j] += sz;
6513 base = (char *)mp + mp->mp_upper;
6514 memmove(base + sz, base, ptr - mp->mp_upper);
6516 mp->mp_lower -= sizeof(indx_t);
6520 /** Compact the main page after deleting a node on a subpage.
6521 * @param[in] mp The main page to operate on.
6522 * @param[in] indx The index of the subpage on the main page.
6525 mdb_node_shrink(MDB_page *mp, indx_t indx)
6531 indx_t i, numkeys, ptr;
6533 node = NODEPTR(mp, indx);
6534 sp = (MDB_page *)NODEDATA(node);
6535 delta = SIZELEFT(sp);
6536 xp = (MDB_page *)((char *)sp + delta);
6538 /* shift subpage upward */
6540 nsize = NUMKEYS(sp) * sp->mp_pad;
6542 return; /* do not make the node uneven-sized */
6543 memmove(METADATA(xp), METADATA(sp), nsize);
6546 numkeys = NUMKEYS(sp);
6547 for (i=numkeys-1; i>=0; i--)
6548 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6550 xp->mp_upper = sp->mp_lower;
6551 xp->mp_lower = sp->mp_lower;
6552 xp->mp_flags = sp->mp_flags;
6553 xp->mp_pad = sp->mp_pad;
6554 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6556 nsize = NODEDSZ(node) - delta;
6557 SETDSZ(node, nsize);
6559 /* shift lower nodes upward */
6560 ptr = mp->mp_ptrs[indx];
6561 numkeys = NUMKEYS(mp);
6562 for (i = 0; i < numkeys; i++) {
6563 if (mp->mp_ptrs[i] <= ptr)
6564 mp->mp_ptrs[i] += delta;
6567 base = (char *)mp + mp->mp_upper;
6568 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6569 mp->mp_upper += delta;
6572 /** Initial setup of a sorted-dups cursor.
6573 * Sorted duplicates are implemented as a sub-database for the given key.
6574 * The duplicate data items are actually keys of the sub-database.
6575 * Operations on the duplicate data items are performed using a sub-cursor
6576 * initialized when the sub-database is first accessed. This function does
6577 * the preliminary setup of the sub-cursor, filling in the fields that
6578 * depend only on the parent DB.
6579 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6582 mdb_xcursor_init0(MDB_cursor *mc)
6584 MDB_xcursor *mx = mc->mc_xcursor;
6586 mx->mx_cursor.mc_xcursor = NULL;
6587 mx->mx_cursor.mc_txn = mc->mc_txn;
6588 mx->mx_cursor.mc_db = &mx->mx_db;
6589 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6590 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6591 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6592 mx->mx_cursor.mc_snum = 0;
6593 mx->mx_cursor.mc_top = 0;
6594 mx->mx_cursor.mc_flags = C_SUB;
6595 mx->mx_dbx.md_name.mv_size = 0;
6596 mx->mx_dbx.md_name.mv_data = NULL;
6597 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6598 mx->mx_dbx.md_dcmp = NULL;
6599 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6602 /** Final setup of a sorted-dups cursor.
6603 * Sets up the fields that depend on the data from the main cursor.
6604 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6605 * @param[in] node The data containing the #MDB_db record for the
6606 * sorted-dup database.
6609 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6611 MDB_xcursor *mx = mc->mc_xcursor;
6613 if (node->mn_flags & F_SUBDATA) {
6614 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6615 mx->mx_cursor.mc_pg[0] = 0;
6616 mx->mx_cursor.mc_snum = 0;
6617 mx->mx_cursor.mc_top = 0;
6618 mx->mx_cursor.mc_flags = C_SUB;
6620 MDB_page *fp = NODEDATA(node);
6621 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6622 mx->mx_db.md_flags = 0;
6623 mx->mx_db.md_depth = 1;
6624 mx->mx_db.md_branch_pages = 0;
6625 mx->mx_db.md_leaf_pages = 1;
6626 mx->mx_db.md_overflow_pages = 0;
6627 mx->mx_db.md_entries = NUMKEYS(fp);
6628 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6629 mx->mx_cursor.mc_snum = 1;
6630 mx->mx_cursor.mc_top = 0;
6631 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6632 mx->mx_cursor.mc_pg[0] = fp;
6633 mx->mx_cursor.mc_ki[0] = 0;
6634 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6635 mx->mx_db.md_flags = MDB_DUPFIXED;
6636 mx->mx_db.md_pad = fp->mp_pad;
6637 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6638 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6641 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6642 mx->mx_db.md_root));
6643 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6644 #if UINT_MAX < SIZE_MAX
6645 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6646 #ifdef MISALIGNED_OK
6647 mx->mx_dbx.md_cmp = mdb_cmp_long;
6649 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6654 /** Initialize a cursor for a given transaction and database. */
6656 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6659 mc->mc_backup = NULL;
6662 mc->mc_db = &txn->mt_dbs[dbi];
6663 mc->mc_dbx = &txn->mt_dbxs[dbi];
6664 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6669 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6671 mc->mc_xcursor = mx;
6672 mdb_xcursor_init0(mc);
6674 mc->mc_xcursor = NULL;
6676 if (*mc->mc_dbflag & DB_STALE) {
6677 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6682 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6685 size_t size = sizeof(MDB_cursor);
6687 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6690 if (txn->mt_flags & MDB_TXN_ERROR)
6693 /* Allow read access to the freelist */
6694 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6697 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6698 size += sizeof(MDB_xcursor);
6700 if ((mc = malloc(size)) != NULL) {
6701 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6702 if (txn->mt_cursors) {
6703 mc->mc_next = txn->mt_cursors[dbi];
6704 txn->mt_cursors[dbi] = mc;
6705 mc->mc_flags |= C_UNTRACK;
6717 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6719 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6722 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6725 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6729 /* Return the count of duplicate data items for the current key */
6731 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6735 if (mc == NULL || countp == NULL)
6738 if (mc->mc_xcursor == NULL)
6739 return MDB_INCOMPATIBLE;
6741 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6742 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6745 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6748 *countp = mc->mc_xcursor->mx_db.md_entries;
6754 mdb_cursor_close(MDB_cursor *mc)
6756 if (mc && !mc->mc_backup) {
6757 /* remove from txn, if tracked */
6758 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6759 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6760 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6762 *prev = mc->mc_next;
6769 mdb_cursor_txn(MDB_cursor *mc)
6771 if (!mc) return NULL;
6776 mdb_cursor_dbi(MDB_cursor *mc)
6782 /** Replace the key for a branch node with a new key.
6783 * @param[in] mc Cursor pointing to the node to operate on.
6784 * @param[in] key The new key to use.
6785 * @return 0 on success, non-zero on failure.
6788 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6794 int delta, ksize, oksize;
6795 indx_t ptr, i, numkeys, indx;
6798 indx = mc->mc_ki[mc->mc_top];
6799 mp = mc->mc_pg[mc->mc_top];
6800 node = NODEPTR(mp, indx);
6801 ptr = mp->mp_ptrs[indx];
6805 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6806 k2.mv_data = NODEKEY(node);
6807 k2.mv_size = node->mn_ksize;
6808 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6810 mdb_dkey(&k2, kbuf2),
6816 /* Sizes must be 2-byte aligned. */
6817 ksize = EVEN(key->mv_size);
6818 oksize = EVEN(node->mn_ksize);
6819 delta = ksize - oksize;
6821 /* Shift node contents if EVEN(key length) changed. */
6823 if (delta > 0 && SIZELEFT(mp) < delta) {
6825 /* not enough space left, do a delete and split */
6826 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6827 pgno = NODEPGNO(node);
6828 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6829 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6832 numkeys = NUMKEYS(mp);
6833 for (i = 0; i < numkeys; i++) {
6834 if (mp->mp_ptrs[i] <= ptr)
6835 mp->mp_ptrs[i] -= delta;
6838 base = (char *)mp + mp->mp_upper;
6839 len = ptr - mp->mp_upper + NODESIZE;
6840 memmove(base - delta, base, len);
6841 mp->mp_upper -= delta;
6843 node = NODEPTR(mp, indx);
6846 /* But even if no shift was needed, update ksize */
6847 if (node->mn_ksize != key->mv_size)
6848 node->mn_ksize = key->mv_size;
6851 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6857 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6859 /** Move a node from csrc to cdst.
6862 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6869 unsigned short flags;
6873 /* Mark src and dst as dirty. */
6874 if ((rc = mdb_page_touch(csrc)) ||
6875 (rc = mdb_page_touch(cdst)))
6878 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6879 key.mv_size = csrc->mc_db->md_pad;
6880 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6882 data.mv_data = NULL;
6886 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6887 assert(!((size_t)srcnode&1));
6888 srcpg = NODEPGNO(srcnode);
6889 flags = srcnode->mn_flags;
6890 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6891 unsigned int snum = csrc->mc_snum;
6893 /* must find the lowest key below src */
6894 mdb_page_search_lowest(csrc);
6895 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6896 key.mv_size = csrc->mc_db->md_pad;
6897 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6899 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6900 key.mv_size = NODEKSZ(s2);
6901 key.mv_data = NODEKEY(s2);
6903 csrc->mc_snum = snum--;
6904 csrc->mc_top = snum;
6906 key.mv_size = NODEKSZ(srcnode);
6907 key.mv_data = NODEKEY(srcnode);
6909 data.mv_size = NODEDSZ(srcnode);
6910 data.mv_data = NODEDATA(srcnode);
6912 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6913 unsigned int snum = cdst->mc_snum;
6916 /* must find the lowest key below dst */
6917 mdb_page_search_lowest(cdst);
6918 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6919 bkey.mv_size = cdst->mc_db->md_pad;
6920 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6922 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6923 bkey.mv_size = NODEKSZ(s2);
6924 bkey.mv_data = NODEKEY(s2);
6926 cdst->mc_snum = snum--;
6927 cdst->mc_top = snum;
6928 mdb_cursor_copy(cdst, &mn);
6930 rc = mdb_update_key(&mn, &bkey);
6935 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6936 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6937 csrc->mc_ki[csrc->mc_top],
6939 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6940 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6942 /* Add the node to the destination page.
6944 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6945 if (rc != MDB_SUCCESS)
6948 /* Delete the node from the source page.
6950 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6953 /* Adjust other cursors pointing to mp */
6954 MDB_cursor *m2, *m3;
6955 MDB_dbi dbi = csrc->mc_dbi;
6956 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6958 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6959 if (csrc->mc_flags & C_SUB)
6960 m3 = &m2->mc_xcursor->mx_cursor;
6963 if (m3 == csrc) continue;
6964 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6965 csrc->mc_ki[csrc->mc_top]) {
6966 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6967 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6972 /* Update the parent separators.
6974 if (csrc->mc_ki[csrc->mc_top] == 0) {
6975 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6976 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6977 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6979 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6980 key.mv_size = NODEKSZ(srcnode);
6981 key.mv_data = NODEKEY(srcnode);
6983 DPRINTF(("update separator for source page %"Z"u to [%s]",
6984 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6985 mdb_cursor_copy(csrc, &mn);
6988 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6991 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6993 indx_t ix = csrc->mc_ki[csrc->mc_top];
6994 nullkey.mv_size = 0;
6995 csrc->mc_ki[csrc->mc_top] = 0;
6996 rc = mdb_update_key(csrc, &nullkey);
6997 csrc->mc_ki[csrc->mc_top] = ix;
6998 assert(rc == MDB_SUCCESS);
7002 if (cdst->mc_ki[cdst->mc_top] == 0) {
7003 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7004 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7005 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7007 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7008 key.mv_size = NODEKSZ(srcnode);
7009 key.mv_data = NODEKEY(srcnode);
7011 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7012 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7013 mdb_cursor_copy(cdst, &mn);
7016 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7019 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7021 indx_t ix = cdst->mc_ki[cdst->mc_top];
7022 nullkey.mv_size = 0;
7023 cdst->mc_ki[cdst->mc_top] = 0;
7024 rc = mdb_update_key(cdst, &nullkey);
7025 cdst->mc_ki[cdst->mc_top] = ix;
7026 assert(rc == MDB_SUCCESS);
7033 /** Merge one page into another.
7034 * The nodes from the page pointed to by \b csrc will
7035 * be copied to the page pointed to by \b cdst and then
7036 * the \b csrc page will be freed.
7037 * @param[in] csrc Cursor pointing to the source page.
7038 * @param[in] cdst Cursor pointing to the destination page.
7041 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7049 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
7050 cdst->mc_pg[cdst->mc_top]->mp_pgno));
7052 assert(csrc->mc_snum > 1); /* can't merge root page */
7053 assert(cdst->mc_snum > 1);
7055 /* Mark dst as dirty. */
7056 if ((rc = mdb_page_touch(cdst)))
7059 /* Move all nodes from src to dst.
7061 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
7062 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7063 key.mv_size = csrc->mc_db->md_pad;
7064 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
7065 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7066 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7067 if (rc != MDB_SUCCESS)
7069 key.mv_data = (char *)key.mv_data + key.mv_size;
7072 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7073 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7074 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7075 unsigned int snum = csrc->mc_snum;
7077 /* must find the lowest key below src */
7078 mdb_page_search_lowest(csrc);
7079 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7080 key.mv_size = csrc->mc_db->md_pad;
7081 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7083 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7084 key.mv_size = NODEKSZ(s2);
7085 key.mv_data = NODEKEY(s2);
7087 csrc->mc_snum = snum--;
7088 csrc->mc_top = snum;
7090 key.mv_size = srcnode->mn_ksize;
7091 key.mv_data = NODEKEY(srcnode);
7094 data.mv_size = NODEDSZ(srcnode);
7095 data.mv_data = NODEDATA(srcnode);
7096 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7097 if (rc != MDB_SUCCESS)
7102 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7103 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7104 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7106 /* Unlink the src page from parent and add to free list.
7108 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
7109 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
7112 rc = mdb_update_key(csrc, &key);
7118 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7119 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7122 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7123 csrc->mc_db->md_leaf_pages--;
7125 csrc->mc_db->md_branch_pages--;
7127 /* Adjust other cursors pointing to mp */
7128 MDB_cursor *m2, *m3;
7129 MDB_dbi dbi = csrc->mc_dbi;
7130 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7132 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7133 if (csrc->mc_flags & C_SUB)
7134 m3 = &m2->mc_xcursor->mx_cursor;
7137 if (m3 == csrc) continue;
7138 if (m3->mc_snum < csrc->mc_snum) continue;
7139 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7140 m3->mc_pg[csrc->mc_top] = mp;
7141 m3->mc_ki[csrc->mc_top] += nkeys;
7145 mdb_cursor_pop(csrc);
7147 return mdb_rebalance(csrc);
7150 /** Copy the contents of a cursor.
7151 * @param[in] csrc The cursor to copy from.
7152 * @param[out] cdst The cursor to copy to.
7155 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7159 cdst->mc_txn = csrc->mc_txn;
7160 cdst->mc_dbi = csrc->mc_dbi;
7161 cdst->mc_db = csrc->mc_db;
7162 cdst->mc_dbx = csrc->mc_dbx;
7163 cdst->mc_snum = csrc->mc_snum;
7164 cdst->mc_top = csrc->mc_top;
7165 cdst->mc_flags = csrc->mc_flags;
7167 for (i=0; i<csrc->mc_snum; i++) {
7168 cdst->mc_pg[i] = csrc->mc_pg[i];
7169 cdst->mc_ki[i] = csrc->mc_ki[i];
7173 /** Rebalance the tree after a delete operation.
7174 * @param[in] mc Cursor pointing to the page where rebalancing
7176 * @return 0 on success, non-zero on failure.
7179 mdb_rebalance(MDB_cursor *mc)
7183 unsigned int ptop, minkeys;
7186 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7190 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7191 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7192 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7193 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7194 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7198 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7199 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7202 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7203 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7209 if (mc->mc_snum < 2) {
7210 MDB_page *mp = mc->mc_pg[0];
7212 DPUTS("Can't rebalance a subpage, ignoring");
7215 if (NUMKEYS(mp) == 0) {
7216 DPUTS("tree is completely empty");
7217 mc->mc_db->md_root = P_INVALID;
7218 mc->mc_db->md_depth = 0;
7219 mc->mc_db->md_leaf_pages = 0;
7220 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7223 /* Adjust cursors pointing to mp */
7226 mc->mc_flags &= ~C_INITIALIZED;
7228 MDB_cursor *m2, *m3;
7229 MDB_dbi dbi = mc->mc_dbi;
7231 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7232 if (mc->mc_flags & C_SUB)
7233 m3 = &m2->mc_xcursor->mx_cursor;
7236 if (m3->mc_snum < mc->mc_snum) continue;
7237 if (m3->mc_pg[0] == mp) {
7240 m3->mc_flags &= ~C_INITIALIZED;
7244 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7245 DPUTS("collapsing root page!");
7246 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7249 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7250 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7253 mc->mc_db->md_depth--;
7254 mc->mc_db->md_branch_pages--;
7255 mc->mc_ki[0] = mc->mc_ki[1];
7257 /* Adjust other cursors pointing to mp */
7258 MDB_cursor *m2, *m3;
7259 MDB_dbi dbi = mc->mc_dbi;
7261 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7262 if (mc->mc_flags & C_SUB)
7263 m3 = &m2->mc_xcursor->mx_cursor;
7266 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7267 if (m3->mc_pg[0] == mp) {
7271 for (i=0; i<m3->mc_snum; i++) {
7272 m3->mc_pg[i] = m3->mc_pg[i+1];
7273 m3->mc_ki[i] = m3->mc_ki[i+1];
7279 DPUTS("root page doesn't need rebalancing");
7283 /* The parent (branch page) must have at least 2 pointers,
7284 * otherwise the tree is invalid.
7286 ptop = mc->mc_top-1;
7287 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7289 /* Leaf page fill factor is below the threshold.
7290 * Try to move keys from left or right neighbor, or
7291 * merge with a neighbor page.
7296 mdb_cursor_copy(mc, &mn);
7297 mn.mc_xcursor = NULL;
7299 if (mc->mc_ki[ptop] == 0) {
7300 /* We're the leftmost leaf in our parent.
7302 DPUTS("reading right neighbor");
7304 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7305 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7308 mn.mc_ki[mn.mc_top] = 0;
7309 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7311 /* There is at least one neighbor to the left.
7313 DPUTS("reading left neighbor");
7315 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7316 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7319 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7320 mc->mc_ki[mc->mc_top] = 0;
7323 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7324 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7325 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7327 /* If the neighbor page is above threshold and has enough keys,
7328 * move one key from it. Otherwise we should try to merge them.
7329 * (A branch page must never have less than 2 keys.)
7331 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7332 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7333 return mdb_node_move(&mn, mc);
7335 if (mc->mc_ki[ptop] == 0)
7336 rc = mdb_page_merge(&mn, mc);
7338 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7339 rc = mdb_page_merge(mc, &mn);
7340 mdb_cursor_copy(&mn, mc);
7342 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7347 /** Complete a delete operation started by #mdb_cursor_del(). */
7349 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7356 mp = mc->mc_pg[mc->mc_top];
7357 ki = mc->mc_ki[mc->mc_top];
7359 /* add overflow pages to free list */
7360 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7364 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7365 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7366 (rc = mdb_ovpage_free(mc, omp)))
7369 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7370 mc->mc_db->md_entries--;
7371 rc = mdb_rebalance(mc);
7372 if (rc != MDB_SUCCESS)
7373 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7376 MDB_dbi dbi = mc->mc_dbi;
7378 mp = mc->mc_pg[mc->mc_top];
7379 nkeys = NUMKEYS(mp);
7381 /* if mc points past last node in page, find next sibling */
7382 if (mc->mc_ki[mc->mc_top] >= nkeys)
7383 mdb_cursor_sibling(mc, 1);
7385 /* Adjust other cursors pointing to mp */
7386 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7387 if (m2 == mc || m2->mc_snum < mc->mc_snum)
7389 if (!(m2->mc_flags & C_INITIALIZED))
7391 if (m2->mc_pg[mc->mc_top] == mp) {
7392 if (m2->mc_ki[mc->mc_top] >= ki) {
7393 m2->mc_flags |= C_DEL;
7394 if (m2->mc_ki[mc->mc_top] > ki)
7395 m2->mc_ki[mc->mc_top]--;
7397 if (m2->mc_ki[mc->mc_top] >= nkeys)
7398 mdb_cursor_sibling(m2, 1);
7401 mc->mc_flags |= C_DEL;
7408 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7409 MDB_val *key, MDB_val *data)
7414 MDB_val rdata, *xdata;
7418 assert(key != NULL);
7420 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7422 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7425 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7426 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7428 mdb_cursor_init(&mc, txn, dbi, &mx);
7431 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7432 /* must ignore any data */
7443 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7445 /* let mdb_page_split know about this cursor if needed:
7446 * delete will trigger a rebalance; if it needs to move
7447 * a node from one page to another, it will have to
7448 * update the parent's separator key(s). If the new sepkey
7449 * is larger than the current one, the parent page may
7450 * run out of space, triggering a split. We need this
7451 * cursor to be consistent until the end of the rebalance.
7453 mc.mc_flags |= C_UNTRACK;
7454 mc.mc_next = txn->mt_cursors[dbi];
7455 txn->mt_cursors[dbi] = &mc;
7456 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7457 txn->mt_cursors[dbi] = mc.mc_next;
7462 /** Split a page and insert a new node.
7463 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7464 * The cursor will be updated to point to the actual page and index where
7465 * the node got inserted after the split.
7466 * @param[in] newkey The key for the newly inserted node.
7467 * @param[in] newdata The data for the newly inserted node.
7468 * @param[in] newpgno The page number, if the new node is a branch node.
7469 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7470 * @return 0 on success, non-zero on failure.
7473 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7474 unsigned int nflags)
7477 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7480 int i, j, split_indx, nkeys, pmax;
7481 MDB_env *env = mc->mc_txn->mt_env;
7483 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7484 MDB_page *copy = NULL;
7485 MDB_page *mp, *rp, *pp;
7490 mp = mc->mc_pg[mc->mc_top];
7491 newindx = mc->mc_ki[mc->mc_top];
7492 nkeys = NUMKEYS(mp);
7494 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7495 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7496 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7498 /* Create a right sibling. */
7499 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7501 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7503 if (mc->mc_snum < 2) {
7504 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7506 /* shift current top to make room for new parent */
7507 mc->mc_pg[1] = mc->mc_pg[0];
7508 mc->mc_ki[1] = mc->mc_ki[0];
7511 mc->mc_db->md_root = pp->mp_pgno;
7512 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7513 mc->mc_db->md_depth++;
7516 /* Add left (implicit) pointer. */
7517 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7518 /* undo the pre-push */
7519 mc->mc_pg[0] = mc->mc_pg[1];
7520 mc->mc_ki[0] = mc->mc_ki[1];
7521 mc->mc_db->md_root = mp->mp_pgno;
7522 mc->mc_db->md_depth--;
7529 ptop = mc->mc_top-1;
7530 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7533 mc->mc_flags |= C_SPLITTING;
7534 mdb_cursor_copy(mc, &mn);
7535 mn.mc_pg[mn.mc_top] = rp;
7536 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7538 if (nflags & MDB_APPEND) {
7539 mn.mc_ki[mn.mc_top] = 0;
7541 split_indx = newindx;
7545 split_indx = (nkeys+1) / 2;
7550 unsigned int lsize, rsize, ksize;
7551 /* Move half of the keys to the right sibling */
7553 x = mc->mc_ki[mc->mc_top] - split_indx;
7554 ksize = mc->mc_db->md_pad;
7555 split = LEAF2KEY(mp, split_indx, ksize);
7556 rsize = (nkeys - split_indx) * ksize;
7557 lsize = (nkeys - split_indx) * sizeof(indx_t);
7558 mp->mp_lower -= lsize;
7559 rp->mp_lower += lsize;
7560 mp->mp_upper += rsize - lsize;
7561 rp->mp_upper -= rsize - lsize;
7562 sepkey.mv_size = ksize;
7563 if (newindx == split_indx) {
7564 sepkey.mv_data = newkey->mv_data;
7566 sepkey.mv_data = split;
7569 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7570 memcpy(rp->mp_ptrs, split, rsize);
7571 sepkey.mv_data = rp->mp_ptrs;
7572 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7573 memcpy(ins, newkey->mv_data, ksize);
7574 mp->mp_lower += sizeof(indx_t);
7575 mp->mp_upper -= ksize - sizeof(indx_t);
7578 memcpy(rp->mp_ptrs, split, x * ksize);
7579 ins = LEAF2KEY(rp, x, ksize);
7580 memcpy(ins, newkey->mv_data, ksize);
7581 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7582 rp->mp_lower += sizeof(indx_t);
7583 rp->mp_upper -= ksize - sizeof(indx_t);
7584 mc->mc_ki[mc->mc_top] = x;
7585 mc->mc_pg[mc->mc_top] = rp;
7588 int psize, nsize, k;
7589 /* Maximum free space in an empty page */
7590 pmax = env->me_psize - PAGEHDRSZ;
7592 nsize = mdb_leaf_size(env, newkey, newdata);
7594 nsize = mdb_branch_size(env, newkey);
7595 nsize = EVEN(nsize);
7597 /* grab a page to hold a temporary copy */
7598 copy = mdb_page_malloc(mc->mc_txn, 1);
7601 copy->mp_pgno = mp->mp_pgno;
7602 copy->mp_flags = mp->mp_flags;
7603 copy->mp_lower = PAGEHDRSZ;
7604 copy->mp_upper = env->me_psize;
7606 /* prepare to insert */
7607 for (i=0, j=0; i<nkeys; i++) {
7609 copy->mp_ptrs[j++] = 0;
7611 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7614 /* When items are relatively large the split point needs
7615 * to be checked, because being off-by-one will make the
7616 * difference between success or failure in mdb_node_add.
7618 * It's also relevant if a page happens to be laid out
7619 * such that one half of its nodes are all "small" and
7620 * the other half of its nodes are "large." If the new
7621 * item is also "large" and falls on the half with
7622 * "large" nodes, it also may not fit.
7624 * As a final tweak, if the new item goes on the last
7625 * spot on the page (and thus, onto the new page), bias
7626 * the split so the new page is emptier than the old page.
7627 * This yields better packing during sequential inserts.
7629 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7630 /* Find split point */
7632 if (newindx <= split_indx || newindx >= nkeys) {
7634 k = newindx >= nkeys ? nkeys : split_indx+2;
7639 for (; i!=k; i+=j) {
7644 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7645 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7647 if (F_ISSET(node->mn_flags, F_BIGDATA))
7648 psize += sizeof(pgno_t);
7650 psize += NODEDSZ(node);
7652 psize = EVEN(psize);
7654 if (psize > pmax || i == k-j) {
7655 split_indx = i + (j<0);
7660 if (split_indx == newindx) {
7661 sepkey.mv_size = newkey->mv_size;
7662 sepkey.mv_data = newkey->mv_data;
7664 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7665 sepkey.mv_size = node->mn_ksize;
7666 sepkey.mv_data = NODEKEY(node);
7671 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7673 /* Copy separator key to the parent.
7675 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7679 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7682 if (mn.mc_snum == mc->mc_snum) {
7683 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7684 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7685 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7686 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7691 /* Right page might now have changed parent.
7692 * Check if left page also changed parent.
7694 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7695 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7696 for (i=0; i<ptop; i++) {
7697 mc->mc_pg[i] = mn.mc_pg[i];
7698 mc->mc_ki[i] = mn.mc_ki[i];
7700 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7701 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7705 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7708 mc->mc_flags ^= C_SPLITTING;
7709 if (rc != MDB_SUCCESS) {
7712 if (nflags & MDB_APPEND) {
7713 mc->mc_pg[mc->mc_top] = rp;
7714 mc->mc_ki[mc->mc_top] = 0;
7715 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7718 for (i=0; i<mc->mc_top; i++)
7719 mc->mc_ki[i] = mn.mc_ki[i];
7720 } else if (!IS_LEAF2(mp)) {
7722 mc->mc_pg[mc->mc_top] = rp;
7727 rkey.mv_data = newkey->mv_data;
7728 rkey.mv_size = newkey->mv_size;
7734 /* Update index for the new key. */
7735 mc->mc_ki[mc->mc_top] = j;
7737 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7738 rkey.mv_data = NODEKEY(node);
7739 rkey.mv_size = node->mn_ksize;
7741 xdata.mv_data = NODEDATA(node);
7742 xdata.mv_size = NODEDSZ(node);
7745 pgno = NODEPGNO(node);
7746 flags = node->mn_flags;
7749 if (!IS_LEAF(mp) && j == 0) {
7750 /* First branch index doesn't need key data. */
7754 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7756 /* return tmp page to freelist */
7757 mdb_page_free(env, copy);
7763 mc->mc_pg[mc->mc_top] = copy;
7768 } while (i != split_indx);
7770 nkeys = NUMKEYS(copy);
7771 for (i=0; i<nkeys; i++)
7772 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7773 mp->mp_lower = copy->mp_lower;
7774 mp->mp_upper = copy->mp_upper;
7775 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7776 env->me_psize - copy->mp_upper);
7778 /* reset back to original page */
7779 if (newindx < split_indx) {
7780 mc->mc_pg[mc->mc_top] = mp;
7781 if (nflags & MDB_RESERVE) {
7782 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7783 if (!(node->mn_flags & F_BIGDATA))
7784 newdata->mv_data = NODEDATA(node);
7787 mc->mc_pg[mc->mc_top] = rp;
7789 /* Make sure mc_ki is still valid.
7791 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7792 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7793 for (i=0; i<ptop; i++) {
7794 mc->mc_pg[i] = mn.mc_pg[i];
7795 mc->mc_ki[i] = mn.mc_ki[i];
7797 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7798 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7801 /* return tmp page to freelist */
7802 mdb_page_free(env, copy);
7806 /* Adjust other cursors pointing to mp */
7807 MDB_cursor *m2, *m3;
7808 MDB_dbi dbi = mc->mc_dbi;
7809 int fixup = NUMKEYS(mp);
7811 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7812 if (mc->mc_flags & C_SUB)
7813 m3 = &m2->mc_xcursor->mx_cursor;
7818 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7820 if (m3->mc_flags & C_SPLITTING)
7825 for (k=m3->mc_top; k>=0; k--) {
7826 m3->mc_ki[k+1] = m3->mc_ki[k];
7827 m3->mc_pg[k+1] = m3->mc_pg[k];
7829 if (m3->mc_ki[0] >= split_indx) {
7834 m3->mc_pg[0] = mc->mc_pg[0];
7838 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7839 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7840 m3->mc_ki[mc->mc_top]++;
7841 if (m3->mc_ki[mc->mc_top] >= fixup) {
7842 m3->mc_pg[mc->mc_top] = rp;
7843 m3->mc_ki[mc->mc_top] -= fixup;
7844 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7846 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7847 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7852 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
7857 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7858 MDB_val *key, MDB_val *data, unsigned int flags)
7863 assert(key != NULL);
7864 assert(data != NULL);
7866 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7869 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7872 mdb_cursor_init(&mc, txn, dbi, &mx);
7873 return mdb_cursor_put(&mc, key, data, flags);
7877 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7879 if ((flag & CHANGEABLE) != flag)
7882 env->me_flags |= flag;
7884 env->me_flags &= ~flag;
7889 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7894 *arg = env->me_flags;
7899 mdb_env_get_path(MDB_env *env, const char **arg)
7904 *arg = env->me_path;
7909 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
7918 /** Common code for #mdb_stat() and #mdb_env_stat().
7919 * @param[in] env the environment to operate in.
7920 * @param[in] db the #MDB_db record containing the stats to return.
7921 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7922 * @return 0, this function always succeeds.
7925 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7927 arg->ms_psize = env->me_psize;
7928 arg->ms_depth = db->md_depth;
7929 arg->ms_branch_pages = db->md_branch_pages;
7930 arg->ms_leaf_pages = db->md_leaf_pages;
7931 arg->ms_overflow_pages = db->md_overflow_pages;
7932 arg->ms_entries = db->md_entries;
7937 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7941 if (env == NULL || arg == NULL)
7944 toggle = mdb_env_pick_meta(env);
7946 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7950 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7954 if (env == NULL || arg == NULL)
7957 toggle = mdb_env_pick_meta(env);
7958 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7959 arg->me_mapsize = env->me_mapsize;
7960 arg->me_maxreaders = env->me_maxreaders;
7962 /* me_numreaders may be zero if this process never used any readers. Use
7963 * the shared numreader count if it exists.
7965 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7967 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7968 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7972 /** Set the default comparison functions for a database.
7973 * Called immediately after a database is opened to set the defaults.
7974 * The user can then override them with #mdb_set_compare() or
7975 * #mdb_set_dupsort().
7976 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7977 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7980 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7982 uint16_t f = txn->mt_dbs[dbi].md_flags;
7984 txn->mt_dbxs[dbi].md_cmp =
7985 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7986 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7988 txn->mt_dbxs[dbi].md_dcmp =
7989 !(f & MDB_DUPSORT) ? 0 :
7990 ((f & MDB_INTEGERDUP)
7991 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7992 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7995 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8000 int rc, dbflag, exact;
8001 unsigned int unused = 0;
8004 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8005 mdb_default_cmp(txn, FREE_DBI);
8008 if ((flags & VALID_FLAGS) != flags)
8010 if (txn->mt_flags & MDB_TXN_ERROR)
8016 if (flags & PERSISTENT_FLAGS) {
8017 uint16_t f2 = flags & PERSISTENT_FLAGS;
8018 /* make sure flag changes get committed */
8019 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8020 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8021 txn->mt_flags |= MDB_TXN_DIRTY;
8024 mdb_default_cmp(txn, MAIN_DBI);
8028 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8029 mdb_default_cmp(txn, MAIN_DBI);
8032 /* Is the DB already open? */
8034 for (i=2; i<txn->mt_numdbs; i++) {
8035 if (!txn->mt_dbxs[i].md_name.mv_size) {
8036 /* Remember this free slot */
8037 if (!unused) unused = i;
8040 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8041 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8047 /* If no free slot and max hit, fail */
8048 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8049 return MDB_DBS_FULL;
8051 /* Cannot mix named databases with some mainDB flags */
8052 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8053 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8055 /* Find the DB info */
8056 dbflag = DB_NEW|DB_VALID;
8059 key.mv_data = (void *)name;
8060 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8061 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8062 if (rc == MDB_SUCCESS) {
8063 /* make sure this is actually a DB */
8064 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8065 if (!(node->mn_flags & F_SUBDATA))
8066 return MDB_INCOMPATIBLE;
8067 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8068 /* Create if requested */
8070 data.mv_size = sizeof(MDB_db);
8071 data.mv_data = &dummy;
8072 memset(&dummy, 0, sizeof(dummy));
8073 dummy.md_root = P_INVALID;
8074 dummy.md_flags = flags & PERSISTENT_FLAGS;
8075 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8079 /* OK, got info, add to table */
8080 if (rc == MDB_SUCCESS) {
8081 unsigned int slot = unused ? unused : txn->mt_numdbs;
8082 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8083 txn->mt_dbxs[slot].md_name.mv_size = len;
8084 txn->mt_dbxs[slot].md_rel = NULL;
8085 txn->mt_dbflags[slot] = dbflag;
8086 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8088 mdb_default_cmp(txn, slot);
8097 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8099 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8102 if (txn->mt_dbflags[dbi] & DB_STALE) {
8105 /* Stale, must read the DB's root. cursor_init does it for us. */
8106 mdb_cursor_init(&mc, txn, dbi, &mx);
8108 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8111 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8114 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8116 ptr = env->me_dbxs[dbi].md_name.mv_data;
8117 env->me_dbxs[dbi].md_name.mv_data = NULL;
8118 env->me_dbxs[dbi].md_name.mv_size = 0;
8119 env->me_dbflags[dbi] = 0;
8123 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8125 /* We could return the flags for the FREE_DBI too but what's the point? */
8126 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8128 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8132 /** Add all the DB's pages to the free list.
8133 * @param[in] mc Cursor on the DB to free.
8134 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8135 * @return 0 on success, non-zero on failure.
8138 mdb_drop0(MDB_cursor *mc, int subs)
8142 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8143 if (rc == MDB_SUCCESS) {
8144 MDB_txn *txn = mc->mc_txn;
8149 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8150 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8153 mdb_cursor_copy(mc, &mx);
8154 while (mc->mc_snum > 0) {
8155 MDB_page *mp = mc->mc_pg[mc->mc_top];
8156 unsigned n = NUMKEYS(mp);
8158 for (i=0; i<n; i++) {
8159 ni = NODEPTR(mp, i);
8160 if (ni->mn_flags & F_BIGDATA) {
8163 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8164 rc = mdb_page_get(txn, pg, &omp, NULL);
8167 assert(IS_OVERFLOW(omp));
8168 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8172 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8173 mdb_xcursor_init1(mc, ni);
8174 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8180 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8182 for (i=0; i<n; i++) {
8184 ni = NODEPTR(mp, i);
8187 mdb_midl_xappend(txn->mt_free_pgs, pg);
8192 mc->mc_ki[mc->mc_top] = i;
8193 rc = mdb_cursor_sibling(mc, 1);
8195 /* no more siblings, go back to beginning
8196 * of previous level.
8200 for (i=1; i<mc->mc_snum; i++) {
8202 mc->mc_pg[i] = mx.mc_pg[i];
8207 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8208 } else if (rc == MDB_NOTFOUND) {
8214 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8216 MDB_cursor *mc, *m2;
8219 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8222 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8225 rc = mdb_cursor_open(txn, dbi, &mc);
8229 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8230 /* Invalidate the dropped DB's cursors */
8231 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8232 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8236 /* Can't delete the main DB */
8237 if (del && dbi > MAIN_DBI) {
8238 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8240 txn->mt_dbflags[dbi] = DB_STALE;
8241 mdb_dbi_close(txn->mt_env, dbi);
8244 /* reset the DB record, mark it dirty */
8245 txn->mt_dbflags[dbi] |= DB_DIRTY;
8246 txn->mt_dbs[dbi].md_depth = 0;
8247 txn->mt_dbs[dbi].md_branch_pages = 0;
8248 txn->mt_dbs[dbi].md_leaf_pages = 0;
8249 txn->mt_dbs[dbi].md_overflow_pages = 0;
8250 txn->mt_dbs[dbi].md_entries = 0;
8251 txn->mt_dbs[dbi].md_root = P_INVALID;
8253 txn->mt_flags |= MDB_TXN_DIRTY;
8256 mdb_cursor_close(mc);
8260 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8262 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8265 txn->mt_dbxs[dbi].md_cmp = cmp;
8269 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8271 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8274 txn->mt_dbxs[dbi].md_dcmp = cmp;
8278 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8280 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8283 txn->mt_dbxs[dbi].md_rel = rel;
8287 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8289 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8292 txn->mt_dbxs[dbi].md_relctx = ctx;
8296 int mdb_env_get_maxkeysize(MDB_env *env)
8298 return ENV_MAXKEY(env);
8301 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8303 unsigned int i, rdrs;
8306 int rc = 0, first = 1;
8310 if (!env->me_txns) {
8311 return func("(no reader locks)\n", ctx);
8313 rdrs = env->me_txns->mti_numreaders;
8314 mr = env->me_txns->mti_readers;
8315 for (i=0; i<rdrs; i++) {
8319 if (mr[i].mr_txnid == (txnid_t)-1) {
8320 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8322 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8326 rc = func(" pid thread txnid\n", ctx);
8330 rc = func(buf, ctx);
8336 rc = func("(no active readers)\n", ctx);
8341 /** Insert pid into list if not already present.
8342 * return -1 if already present.
8344 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
8346 /* binary search of pid in list */
8348 unsigned cursor = 1;
8350 unsigned n = ids[0];
8353 unsigned pivot = n >> 1;
8354 cursor = base + pivot + 1;
8355 val = pid - ids[cursor];
8360 } else if ( val > 0 ) {
8365 /* found, so it's a duplicate */
8374 for (n = ids[0]; n > cursor; n--)
8380 int mdb_reader_check(MDB_env *env, int *dead)
8382 unsigned int i, j, rdrs;
8384 MDB_PID_T *pids, pid;
8393 rdrs = env->me_txns->mti_numreaders;
8394 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
8398 mr = env->me_txns->mti_readers;
8399 for (i=0; i<rdrs; i++) {
8400 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8402 if (mdb_pid_insert(pids, pid) == 0) {
8403 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8405 /* Recheck, a new process may have reused pid */
8406 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8407 for (j=i; j<rdrs; j++)
8408 if (mr[j].mr_pid == pid) {
8409 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8410 (unsigned) pid, mr[j].mr_txnid));
8415 UNLOCK_MUTEX_R(env);