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>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define MDB_PIDLOCK 0
144 #define pthread_t DWORD
145 #define pthread_mutex_t HANDLE
146 #define pthread_key_t DWORD
147 #define pthread_self() GetCurrentThreadId()
148 #define pthread_key_create(x,y) \
149 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
150 #define pthread_key_delete(x) TlsFree(x)
151 #define pthread_getspecific(x) TlsGetValue(x)
152 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
153 #define pthread_mutex_unlock(x) ReleaseMutex(x)
154 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
155 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
156 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
157 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
158 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
159 #define getpid() GetCurrentProcessId()
160 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
161 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
162 #define ErrCode() GetLastError()
163 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
164 #define close(fd) (CloseHandle(fd) ? 0 : -1)
165 #define munmap(ptr,len) UnmapViewOfFile(ptr)
166 #ifndef PROCESS_QUERY_LIMITED_INFORMATION
167 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
169 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
176 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
177 #define MDB_PIDLOCK 1
179 #ifdef MDB_USE_POSIX_SEM
181 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
182 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
183 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
184 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
187 mdb_sem_wait(sem_t *sem)
190 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
195 /** Lock the reader mutex.
197 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
198 /** Unlock the reader mutex.
200 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
202 /** Lock the writer mutex.
203 * Only a single write transaction is allowed at a time. Other writers
204 * will block waiting for this mutex.
206 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
207 /** Unlock the writer mutex.
209 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
210 #endif /* MDB_USE_POSIX_SEM */
212 /** Get the error code for the last failed system function.
214 #define ErrCode() errno
216 /** An abstraction for a file handle.
217 * On POSIX systems file handles are small integers. On Windows
218 * they're opaque pointers.
222 /** A value for an invalid file handle.
223 * Mainly used to initialize file variables and signify that they are
226 #define INVALID_HANDLE_VALUE (-1)
228 /** Get the size of a memory page for the system.
229 * This is the basic size that the platform's memory manager uses, and is
230 * fundamental to the use of memory-mapped files.
232 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
235 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
238 #define MNAME_LEN (sizeof(pthread_mutex_t))
244 /** A flag for opening a file and requesting synchronous data writes.
245 * This is only used when writing a meta page. It's not strictly needed;
246 * we could just do a normal write and then immediately perform a flush.
247 * But if this flag is available it saves us an extra system call.
249 * @note If O_DSYNC is undefined but exists in /usr/include,
250 * preferably set some compiler flag to get the definition.
251 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
254 # define MDB_DSYNC O_DSYNC
258 /** Function for flushing the data of a file. Define this to fsync
259 * if fdatasync() is not supported.
261 #ifndef MDB_FDATASYNC
262 # define MDB_FDATASYNC fdatasync
266 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
277 /** A page number in the database.
278 * Note that 64 bit page numbers are overkill, since pages themselves
279 * already represent 12-13 bits of addressable memory, and the OS will
280 * always limit applications to a maximum of 63 bits of address space.
282 * @note In the #MDB_node structure, we only store 48 bits of this value,
283 * which thus limits us to only 60 bits of addressable data.
285 typedef MDB_ID pgno_t;
287 /** A transaction ID.
288 * See struct MDB_txn.mt_txnid for details.
290 typedef MDB_ID txnid_t;
292 /** @defgroup debug Debug Macros
296 /** Enable debug output.
297 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
298 * read from and written to the database (used for free space management).
303 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
306 # define DPRINTF (void) /* Vararg macros may be unsupported */
308 static int mdb_debug;
309 static txnid_t mdb_debug_start;
311 /** Print a debug message with printf formatting. */
312 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
313 ((void) ((mdb_debug) && \
314 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
316 # define DPRINTF(fmt, ...) ((void) 0)
317 # define MDB_DEBUG_SKIP
319 /** Print a debug string.
320 * The string is printed literally, with no format processing.
322 #define DPUTS(arg) DPRINTF("%s", arg)
325 /** A default memory page size.
326 * The actual size is platform-dependent, but we use this for
327 * boot-strapping. We probably should not be using this any more.
328 * The #GET_PAGESIZE() macro is used to get the actual size.
330 * Note that we don't currently support Huge pages. On Linux,
331 * regular data files cannot use Huge pages, and in general
332 * Huge pages aren't actually pageable. We rely on the OS
333 * demand-pager to read our data and page it out when memory
334 * pressure from other processes is high. So until OSs have
335 * actual paging support for Huge pages, they're not viable.
337 #define MDB_PAGESIZE 4096
339 /** The minimum number of keys required in a database page.
340 * Setting this to a larger value will place a smaller bound on the
341 * maximum size of a data item. Data items larger than this size will
342 * be pushed into overflow pages instead of being stored directly in
343 * the B-tree node. This value used to default to 4. With a page size
344 * of 4096 bytes that meant that any item larger than 1024 bytes would
345 * go into an overflow page. That also meant that on average 2-3KB of
346 * each overflow page was wasted space. The value cannot be lower than
347 * 2 because then there would no longer be a tree structure. With this
348 * value, items larger than 2KB will go into overflow pages, and on
349 * average only 1KB will be wasted.
351 #define MDB_MINKEYS 2
353 /** A stamp that identifies a file as an MDB file.
354 * There's nothing special about this value other than that it is easily
355 * recognizable, and it will reflect any byte order mismatches.
357 #define MDB_MAGIC 0xBEEFC0DE
359 /** The version number for a database's datafile format. */
360 #define MDB_DATA_VERSION 1
361 /** The version number for a database's lockfile format. */
362 #define MDB_LOCK_VERSION 1
364 /** @brief The maximum size of a key in the database.
366 * The library rejects bigger keys, and cannot deal with records
367 * with bigger keys stored by a library with bigger max keysize.
369 * We require that keys all fit onto a regular page. This limit
370 * could be raised a bit further if needed; to something just
371 * under #MDB_PAGESIZE / #MDB_MINKEYS.
373 * Note that data items in an #MDB_DUPSORT database are actually keys
374 * of a subDB, so they're also limited to this size.
376 #ifndef MDB_MAXKEYSIZE
377 #define MDB_MAXKEYSIZE 511
380 /** @brief The maximum size of a data item.
382 * We only store a 32 bit value for node sizes.
384 #define MAXDATASIZE 0xffffffffUL
389 * This is used for printing a hex dump of a key's contents.
391 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
392 /** Display a key in hex.
394 * Invoke a function to display a key in hex.
396 #define DKEY(x) mdb_dkey(x, kbuf)
398 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
402 /** An invalid page number.
403 * Mainly used to denote an empty tree.
405 #define P_INVALID (~(pgno_t)0)
407 /** Test if the flags \b f are set in a flag word \b w. */
408 #define F_ISSET(w, f) (((w) & (f)) == (f))
410 /** Used for offsets within a single page.
411 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
414 typedef uint16_t indx_t;
416 /** Default size of memory map.
417 * This is certainly too small for any actual applications. Apps should always set
418 * the size explicitly using #mdb_env_set_mapsize().
420 #define DEFAULT_MAPSIZE 1048576
422 /** @defgroup readers Reader Lock Table
423 * Readers don't acquire any locks for their data access. Instead, they
424 * simply record their transaction ID in the reader table. The reader
425 * mutex is needed just to find an empty slot in the reader table. The
426 * slot's address is saved in thread-specific data so that subsequent read
427 * transactions started by the same thread need no further locking to proceed.
429 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
431 * No reader table is used if the database is on a read-only filesystem.
433 * Since the database uses multi-version concurrency control, readers don't
434 * actually need any locking. This table is used to keep track of which
435 * readers are using data from which old transactions, so that we'll know
436 * when a particular old transaction is no longer in use. Old transactions
437 * that have discarded any data pages can then have those pages reclaimed
438 * for use by a later write transaction.
440 * The lock table is constructed such that reader slots are aligned with the
441 * processor's cache line size. Any slot is only ever used by one thread.
442 * This alignment guarantees that there will be no contention or cache
443 * thrashing as threads update their own slot info, and also eliminates
444 * any need for locking when accessing a slot.
446 * A writer thread will scan every slot in the table to determine the oldest
447 * outstanding reader transaction. Any freed pages older than this will be
448 * reclaimed by the writer. The writer doesn't use any locks when scanning
449 * this table. This means that there's no guarantee that the writer will
450 * see the most up-to-date reader info, but that's not required for correct
451 * operation - all we need is to know the upper bound on the oldest reader,
452 * we don't care at all about the newest reader. So the only consequence of
453 * reading stale information here is that old pages might hang around a
454 * while longer before being reclaimed. That's actually good anyway, because
455 * the longer we delay reclaiming old pages, the more likely it is that a
456 * string of contiguous pages can be found after coalescing old pages from
457 * many old transactions together.
460 /** Number of slots in the reader table.
461 * This value was chosen somewhat arbitrarily. 126 readers plus a
462 * couple mutexes fit exactly into 8KB on my development machine.
463 * Applications should set the table size using #mdb_env_set_maxreaders().
465 #define DEFAULT_READERS 126
467 /** The size of a CPU cache line in bytes. We want our lock structures
468 * aligned to this size to avoid false cache line sharing in the
470 * This value works for most CPUs. For Itanium this should be 128.
476 /** The information we store in a single slot of the reader table.
477 * In addition to a transaction ID, we also record the process and
478 * thread ID that owns a slot, so that we can detect stale information,
479 * e.g. threads or processes that went away without cleaning up.
480 * @note We currently don't check for stale records. We simply re-init
481 * the table when we know that we're the only process opening the
484 typedef struct MDB_rxbody {
485 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
486 * Multiple readers that start at the same time will probably have the
487 * same ID here. Again, it's not important to exclude them from
488 * anything; all we need to know is which version of the DB they
489 * started from so we can avoid overwriting any data used in that
490 * particular version.
493 /** The process ID of the process owning this reader txn. */
495 /** The thread ID of the thread owning this txn. */
499 /** The actual reader record, with cacheline padding. */
500 typedef struct MDB_reader {
503 /** shorthand for mrb_txnid */
504 #define mr_txnid mru.mrx.mrb_txnid
505 #define mr_pid mru.mrx.mrb_pid
506 #define mr_tid mru.mrx.mrb_tid
507 /** cache line alignment */
508 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
512 /** The header for the reader table.
513 * The table resides in a memory-mapped file. (This is a different file
514 * than is used for the main database.)
516 * For POSIX the actual mutexes reside in the shared memory of this
517 * mapped file. On Windows, mutexes are named objects allocated by the
518 * kernel; we store the mutex names in this mapped file so that other
519 * processes can grab them. This same approach is also used on
520 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
521 * process-shared POSIX mutexes. For these cases where a named object
522 * is used, the object name is derived from a 64 bit FNV hash of the
523 * environment pathname. As such, naming collisions are extremely
524 * unlikely. If a collision occurs, the results are unpredictable.
526 typedef struct MDB_txbody {
527 /** Stamp identifying this as an MDB file. It must be set
530 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
532 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
533 char mtb_rmname[MNAME_LEN];
535 /** Mutex protecting access to this table.
536 * This is the reader lock that #LOCK_MUTEX_R acquires.
538 pthread_mutex_t mtb_mutex;
540 /** The ID of the last transaction committed to the database.
541 * This is recorded here only for convenience; the value can always
542 * be determined by reading the main database meta pages.
545 /** The number of slots that have been used in the reader table.
546 * This always records the maximum count, it is not decremented
547 * when readers release their slots.
549 unsigned mtb_numreaders;
552 /** The actual reader table definition. */
553 typedef struct MDB_txninfo {
556 #define mti_magic mt1.mtb.mtb_magic
557 #define mti_format mt1.mtb.mtb_format
558 #define mti_mutex mt1.mtb.mtb_mutex
559 #define mti_rmname mt1.mtb.mtb_rmname
560 #define mti_txnid mt1.mtb.mtb_txnid
561 #define mti_numreaders mt1.mtb.mtb_numreaders
562 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
565 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
566 char mt2_wmname[MNAME_LEN];
567 #define mti_wmname mt2.mt2_wmname
569 pthread_mutex_t mt2_wmutex;
570 #define mti_wmutex mt2.mt2_wmutex
572 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
574 MDB_reader mti_readers[1];
577 /** Lockfile format signature: version, features and field layout */
578 #define MDB_LOCK_FORMAT \
580 ((MDB_LOCK_VERSION) \
581 /* Flags which describe functionality */ \
582 + (((MDB_PIDLOCK) != 0) << 16)))
585 /** Common header for all page types.
586 * Overflow records occupy a number of contiguous pages with no
587 * headers on any page after the first.
589 typedef struct MDB_page {
590 #define mp_pgno mp_p.p_pgno
591 #define mp_next mp_p.p_next
593 pgno_t p_pgno; /**< page number */
594 void * p_next; /**< for in-memory list of freed structs */
597 /** @defgroup mdb_page Page Flags
599 * Flags for the page headers.
602 #define P_BRANCH 0x01 /**< branch page */
603 #define P_LEAF 0x02 /**< leaf page */
604 #define P_OVERFLOW 0x04 /**< overflow page */
605 #define P_META 0x08 /**< meta page */
606 #define P_DIRTY 0x10 /**< dirty page */
607 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
608 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
609 #define P_KEEP 0x8000 /**< leave this page alone during spill */
611 uint16_t mp_flags; /**< @ref mdb_page */
612 #define mp_lower mp_pb.pb.pb_lower
613 #define mp_upper mp_pb.pb.pb_upper
614 #define mp_pages mp_pb.pb_pages
617 indx_t pb_lower; /**< lower bound of free space */
618 indx_t pb_upper; /**< upper bound of free space */
620 uint32_t pb_pages; /**< number of overflow pages */
622 indx_t mp_ptrs[1]; /**< dynamic size */
625 /** Size of the page header, excluding dynamic data at the end */
626 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
628 /** Address of first usable data byte in a page, after the header */
629 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
631 /** Number of nodes on a page */
632 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
634 /** The amount of space remaining in the page */
635 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
637 /** The percentage of space used in the page, in tenths of a percent. */
638 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
639 ((env)->me_psize - PAGEHDRSZ))
640 /** The minimum page fill factor, in tenths of a percent.
641 * Pages emptier than this are candidates for merging.
643 #define FILL_THRESHOLD 250
645 /** Test if a page is a leaf page */
646 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
647 /** Test if a page is a LEAF2 page */
648 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
649 /** Test if a page is a branch page */
650 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
651 /** Test if a page is an overflow page */
652 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
653 /** Test if a page is a sub page */
654 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
656 /** The number of overflow pages needed to store the given size. */
657 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
659 /** Header for a single key/data pair within a page.
660 * We guarantee 2-byte alignment for nodes.
662 typedef struct MDB_node {
663 /** lo and hi are used for data size on leaf nodes and for
664 * child pgno on branch nodes. On 64 bit platforms, flags
665 * is also used for pgno. (Branch nodes have no flags).
666 * They are in host byte order in case that lets some
667 * accesses be optimized into a 32-bit word access.
669 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
670 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
671 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
672 /** @defgroup mdb_node Node Flags
674 * Flags for node headers.
677 #define F_BIGDATA 0x01 /**< data put on overflow page */
678 #define F_SUBDATA 0x02 /**< data is a sub-database */
679 #define F_DUPDATA 0x04 /**< data has duplicates */
681 /** valid flags for #mdb_node_add() */
682 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
685 unsigned short mn_flags; /**< @ref mdb_node */
686 unsigned short mn_ksize; /**< key size */
687 char mn_data[1]; /**< key and data are appended here */
690 /** Size of the node header, excluding dynamic data at the end */
691 #define NODESIZE offsetof(MDB_node, mn_data)
693 /** Bit position of top word in page number, for shifting mn_flags */
694 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
696 /** Size of a node in a branch page with a given key.
697 * This is just the node header plus the key, there is no data.
699 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
701 /** Size of a node in a leaf page with a given key and data.
702 * This is node header plus key plus data size.
704 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
706 /** Address of node \b i in page \b p */
707 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
709 /** Address of the key for the node */
710 #define NODEKEY(node) (void *)((node)->mn_data)
712 /** Address of the data for a node */
713 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
715 /** Get the page number pointed to by a branch node */
716 #define NODEPGNO(node) \
717 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
718 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
719 /** Set the page number in a branch node */
720 #define SETPGNO(node,pgno) do { \
721 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
722 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
724 /** Get the size of the data in a leaf node */
725 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
726 /** Set the size of the data for a leaf node */
727 #define SETDSZ(node,size) do { \
728 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
729 /** The size of a key in a node */
730 #define NODEKSZ(node) ((node)->mn_ksize)
732 /** Copy a page number from src to dst */
734 #define COPY_PGNO(dst,src) dst = src
736 #if SIZE_MAX > 4294967295UL
737 #define COPY_PGNO(dst,src) do { \
738 unsigned short *s, *d; \
739 s = (unsigned short *)&(src); \
740 d = (unsigned short *)&(dst); \
747 #define COPY_PGNO(dst,src) do { \
748 unsigned short *s, *d; \
749 s = (unsigned short *)&(src); \
750 d = (unsigned short *)&(dst); \
756 /** The address of a key in a LEAF2 page.
757 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
758 * There are no node headers, keys are stored contiguously.
760 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
762 /** Set the \b node's key into \b key, if requested. */
763 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
764 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
766 /** Information about a single database in the environment. */
767 typedef struct MDB_db {
768 uint32_t md_pad; /**< also ksize for LEAF2 pages */
769 uint16_t md_flags; /**< @ref mdb_dbi_open */
770 uint16_t md_depth; /**< depth of this tree */
771 pgno_t md_branch_pages; /**< number of internal pages */
772 pgno_t md_leaf_pages; /**< number of leaf pages */
773 pgno_t md_overflow_pages; /**< number of overflow pages */
774 size_t md_entries; /**< number of data items */
775 pgno_t md_root; /**< the root page of this tree */
778 /** mdb_dbi_open flags */
779 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
780 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
781 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
782 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
784 /** Handle for the DB used to track free pages. */
786 /** Handle for the default DB. */
789 /** Meta page content. */
790 typedef struct MDB_meta {
791 /** Stamp identifying this as an MDB file. It must be set
794 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
796 void *mm_address; /**< address for fixed mapping */
797 size_t mm_mapsize; /**< size of mmap region */
798 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
799 /** The size of pages used in this DB */
800 #define mm_psize mm_dbs[0].md_pad
801 /** Any persistent environment flags. @ref mdb_env */
802 #define mm_flags mm_dbs[0].md_flags
803 pgno_t mm_last_pg; /**< last used page in file */
804 txnid_t mm_txnid; /**< txnid that committed this page */
807 /** Buffer for a stack-allocated dirty page.
808 * The members define size and alignment, and silence type
809 * aliasing warnings. They are not used directly; that could
810 * mean incorrectly using several union members in parallel.
812 typedef union MDB_pagebuf {
813 char mb_raw[MDB_PAGESIZE];
816 char mm_pad[PAGEHDRSZ];
821 /** Auxiliary DB info.
822 * The information here is mostly static/read-only. There is
823 * only a single copy of this record in the environment.
825 typedef struct MDB_dbx {
826 MDB_val md_name; /**< name of the database */
827 MDB_cmp_func *md_cmp; /**< function for comparing keys */
828 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
829 MDB_rel_func *md_rel; /**< user relocate function */
830 void *md_relctx; /**< user-provided context for md_rel */
833 /** A database transaction.
834 * Every operation requires a transaction handle.
837 MDB_txn *mt_parent; /**< parent of a nested txn */
838 MDB_txn *mt_child; /**< nested txn under this txn */
839 pgno_t mt_next_pgno; /**< next unallocated page */
840 /** The ID of this transaction. IDs are integers incrementing from 1.
841 * Only committed write transactions increment the ID. If a transaction
842 * aborts, the ID may be re-used by the next writer.
845 MDB_env *mt_env; /**< the DB environment */
846 /** The list of pages that became unused during this transaction.
849 /** The sorted list of dirty pages we temporarily wrote to disk
850 * because the dirty list was full.
852 MDB_IDL mt_spill_pgs;
854 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
856 /** For read txns: This thread/txn's reader table slot, or NULL. */
859 /** Array of records for each DB known in the environment. */
861 /** Array of MDB_db records for each known DB */
863 /** @defgroup mt_dbflag Transaction DB Flags
867 #define DB_DIRTY 0x01 /**< DB was written in this txn */
868 #define DB_STALE 0x02 /**< DB record is older than txnID */
869 #define DB_NEW 0x04 /**< DB handle opened in this txn */
870 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
872 /** In write txns, array of cursors for each DB */
873 MDB_cursor **mt_cursors;
874 /** Array of flags for each DB */
875 unsigned char *mt_dbflags;
876 /** Number of DB records in use. This number only ever increments;
877 * we don't decrement it when individual DB handles are closed.
881 /** @defgroup mdb_txn Transaction Flags
885 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
886 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
887 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
888 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
890 unsigned int mt_flags; /**< @ref mdb_txn */
891 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
892 unsigned int mt_dirty_room;
893 /** Tracks which of the two meta pages was used at the start
894 * of this transaction.
896 unsigned int mt_toggle;
899 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
900 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
901 * raise this on a 64 bit machine.
903 #define CURSOR_STACK 32
907 /** Cursors are used for all DB operations */
909 /** Next cursor on this DB in this txn */
911 /** Backup of the original cursor if this cursor is a shadow */
912 MDB_cursor *mc_backup;
913 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
914 struct MDB_xcursor *mc_xcursor;
915 /** The transaction that owns this cursor */
917 /** The database handle this cursor operates on */
919 /** The database record for this cursor */
921 /** The database auxiliary record for this cursor */
923 /** The @ref mt_dbflag for this database */
924 unsigned char *mc_dbflag;
925 unsigned short mc_snum; /**< number of pushed pages */
926 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
927 /** @defgroup mdb_cursor Cursor Flags
929 * Cursor state flags.
932 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
933 #define C_EOF 0x02 /**< No more data */
934 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
935 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
936 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
938 unsigned int mc_flags; /**< @ref mdb_cursor */
939 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
940 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
943 /** Context for sorted-dup records.
944 * We could have gone to a fully recursive design, with arbitrarily
945 * deep nesting of sub-databases. But for now we only handle these
946 * levels - main DB, optional sub-DB, sorted-duplicate DB.
948 typedef struct MDB_xcursor {
949 /** A sub-cursor for traversing the Dup DB */
950 MDB_cursor mx_cursor;
951 /** The database record for this Dup DB */
953 /** The auxiliary DB record for this Dup DB */
955 /** The @ref mt_dbflag for this Dup DB */
956 unsigned char mx_dbflag;
959 /** State of FreeDB old pages, stored in the MDB_env */
960 typedef struct MDB_pgstate {
961 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
962 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
965 /** The database environment. */
967 HANDLE me_fd; /**< The main data file */
968 HANDLE me_lfd; /**< The lock file */
969 HANDLE me_mfd; /**< just for writing the meta pages */
970 /** Failed to update the meta page. Probably an I/O error. */
971 #define MDB_FATAL_ERROR 0x80000000U
972 /** Some fields are initialized. */
973 #define MDB_ENV_ACTIVE 0x20000000U
974 /** me_txkey is set */
975 #define MDB_ENV_TXKEY 0x10000000U
976 /** Have liveness lock in reader table */
977 #define MDB_LIVE_READER 0x08000000U
978 uint32_t me_flags; /**< @ref mdb_env */
979 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
980 unsigned int me_maxreaders; /**< size of the reader table */
981 unsigned int me_numreaders; /**< max numreaders set by this env */
982 MDB_dbi me_numdbs; /**< number of DBs opened */
983 MDB_dbi me_maxdbs; /**< size of the DB table */
984 pid_t me_pid; /**< process ID of this env */
985 char *me_path; /**< path to the DB files */
986 char *me_map; /**< the memory map of the data file */
987 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
988 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
989 MDB_txn *me_txn; /**< current write transaction */
990 size_t me_mapsize; /**< size of the data memory map */
991 off_t me_size; /**< current file size */
992 pgno_t me_maxpg; /**< me_mapsize / me_psize */
993 MDB_dbx *me_dbxs; /**< array of static DB info */
994 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
995 pthread_key_t me_txkey; /**< thread-key for readers */
996 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
997 # define me_pglast me_pgstate.mf_pglast
998 # define me_pghead me_pgstate.mf_pghead
999 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1000 /** IDL of pages that became unused in a write txn */
1001 MDB_IDL me_free_pgs;
1002 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1003 MDB_ID2L me_dirty_list;
1004 /** Max number of freelist items that can fit in a single overflow page */
1006 /** Max size of a node on a page */
1007 unsigned int me_nodemax;
1009 int me_pidquery; /**< Used in OpenProcess */
1010 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1012 #elif defined(MDB_USE_POSIX_SEM)
1013 sem_t *me_rmutex; /* Shared mutexes are not supported */
1018 /** Nested transaction */
1019 typedef struct MDB_ntxn {
1020 MDB_txn mnt_txn; /* the transaction */
1021 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
1024 /** max number of pages to commit in one writev() call */
1025 #define MDB_COMMIT_PAGES 64
1026 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1027 #undef MDB_COMMIT_PAGES
1028 #define MDB_COMMIT_PAGES IOV_MAX
1031 /* max bytes to write in one call */
1032 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1034 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1035 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1036 static int mdb_page_touch(MDB_cursor *mc);
1038 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1039 static int mdb_page_search_root(MDB_cursor *mc,
1040 MDB_val *key, int modify);
1041 #define MDB_PS_MODIFY 1
1042 #define MDB_PS_ROOTONLY 2
1043 static int mdb_page_search(MDB_cursor *mc,
1044 MDB_val *key, int flags);
1045 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1047 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1048 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1049 pgno_t newpgno, unsigned int nflags);
1051 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1052 static int mdb_env_pick_meta(const MDB_env *env);
1053 static int mdb_env_write_meta(MDB_txn *txn);
1054 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1055 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1057 static void mdb_env_close0(MDB_env *env, int excl);
1059 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1060 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1061 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1062 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1063 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1064 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1065 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1066 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1067 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1069 static int mdb_rebalance(MDB_cursor *mc);
1070 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1072 static void mdb_cursor_pop(MDB_cursor *mc);
1073 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1075 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1076 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1077 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1078 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1079 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1081 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1082 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1084 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1085 static void mdb_xcursor_init0(MDB_cursor *mc);
1086 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1088 static int mdb_drop0(MDB_cursor *mc, int subs);
1089 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1092 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1096 static SECURITY_DESCRIPTOR mdb_null_sd;
1097 static SECURITY_ATTRIBUTES mdb_all_sa;
1098 static int mdb_sec_inited;
1101 /** Return the library version info. */
1103 mdb_version(int *major, int *minor, int *patch)
1105 if (major) *major = MDB_VERSION_MAJOR;
1106 if (minor) *minor = MDB_VERSION_MINOR;
1107 if (patch) *patch = MDB_VERSION_PATCH;
1108 return MDB_VERSION_STRING;
1111 /** Table of descriptions for MDB @ref errors */
1112 static char *const mdb_errstr[] = {
1113 "MDB_KEYEXIST: Key/data pair already exists",
1114 "MDB_NOTFOUND: No matching key/data pair found",
1115 "MDB_PAGE_NOTFOUND: Requested page not found",
1116 "MDB_CORRUPTED: Located page was wrong type",
1117 "MDB_PANIC: Update of meta page failed",
1118 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1119 "MDB_INVALID: File is not an MDB file",
1120 "MDB_MAP_FULL: Environment mapsize limit reached",
1121 "MDB_DBS_FULL: Environment maxdbs limit reached",
1122 "MDB_READERS_FULL: Environment maxreaders limit reached",
1123 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1124 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1125 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1126 "MDB_PAGE_FULL: Internal error - page has no more space",
1127 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1128 "MDB_INCOMPATIBLE: Database flags changed or would change",
1129 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1133 mdb_strerror(int err)
1137 return ("Successful return: 0");
1139 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1140 i = err - MDB_KEYEXIST;
1141 return mdb_errstr[i];
1144 return strerror(err);
1148 /** Display a key in hexadecimal and return the address of the result.
1149 * @param[in] key the key to display
1150 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1151 * @return The key in hexadecimal form.
1154 mdb_dkey(MDB_val *key, char *buf)
1157 unsigned char *c = key->mv_data;
1163 if (key->mv_size > MDB_MAXKEYSIZE)
1164 return "MDB_MAXKEYSIZE";
1165 /* may want to make this a dynamic check: if the key is mostly
1166 * printable characters, print it as-is instead of converting to hex.
1170 for (i=0; i<key->mv_size; i++)
1171 ptr += sprintf(ptr, "%02x", *c++);
1173 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1178 /** Display all the keys in the page. */
1180 mdb_page_list(MDB_page *mp)
1183 unsigned int i, nkeys, nsize;
1187 nkeys = NUMKEYS(mp);
1188 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1189 for (i=0; i<nkeys; i++) {
1190 node = NODEPTR(mp, i);
1191 key.mv_size = node->mn_ksize;
1192 key.mv_data = node->mn_data;
1193 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1194 if (IS_BRANCH(mp)) {
1195 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1198 if (F_ISSET(node->mn_flags, F_BIGDATA))
1199 nsize += sizeof(pgno_t);
1201 nsize += NODEDSZ(node);
1202 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1208 mdb_cursor_chk(MDB_cursor *mc)
1214 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1215 for (i=0; i<mc->mc_top; i++) {
1217 node = NODEPTR(mp, mc->mc_ki[i]);
1218 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1221 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1227 /** Count all the pages in each DB and in the freelist
1228 * and make sure it matches the actual number of pages
1231 static void mdb_audit(MDB_txn *txn)
1235 MDB_ID freecount, count;
1240 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1241 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1242 freecount += *(MDB_ID *)data.mv_data;
1245 for (i = 0; i<txn->mt_numdbs; i++) {
1247 mdb_cursor_init(&mc, txn, i, &mx);
1248 if (txn->mt_dbs[i].md_root == P_INVALID)
1250 count += txn->mt_dbs[i].md_branch_pages +
1251 txn->mt_dbs[i].md_leaf_pages +
1252 txn->mt_dbs[i].md_overflow_pages;
1253 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1254 mdb_page_search(&mc, NULL, 0);
1258 mp = mc.mc_pg[mc.mc_top];
1259 for (j=0; j<NUMKEYS(mp); j++) {
1260 MDB_node *leaf = NODEPTR(mp, j);
1261 if (leaf->mn_flags & F_SUBDATA) {
1263 memcpy(&db, NODEDATA(leaf), sizeof(db));
1264 count += db.md_branch_pages + db.md_leaf_pages +
1265 db.md_overflow_pages;
1269 while (mdb_cursor_sibling(&mc, 1) == 0);
1272 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1273 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1274 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1280 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1282 return txn->mt_dbxs[dbi].md_cmp(a, b);
1286 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1288 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1291 /** Allocate memory for a page.
1292 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1295 mdb_page_malloc(MDB_txn *txn, unsigned num)
1297 MDB_env *env = txn->mt_env;
1298 MDB_page *ret = env->me_dpages;
1299 size_t sz = env->me_psize;
1302 VGMEMP_ALLOC(env, ret, sz);
1303 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1304 env->me_dpages = ret->mp_next;
1310 if ((ret = malloc(sz)) != NULL) {
1311 VGMEMP_ALLOC(env, ret, sz);
1316 /** Free a single page.
1317 * Saves single pages to a list, for future reuse.
1318 * (This is not used for multi-page overflow pages.)
1321 mdb_page_free(MDB_env *env, MDB_page *mp)
1323 mp->mp_next = env->me_dpages;
1324 VGMEMP_FREE(env, mp);
1325 env->me_dpages = mp;
1328 /* Free a dirty page */
1330 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1332 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1333 mdb_page_free(env, dp);
1335 /* large pages just get freed directly */
1336 VGMEMP_FREE(env, dp);
1341 /** Return all dirty pages to dpage list */
1343 mdb_dlist_free(MDB_txn *txn)
1345 MDB_env *env = txn->mt_env;
1346 MDB_ID2L dl = txn->mt_u.dirty_list;
1347 unsigned i, n = dl[0].mid;
1349 for (i = 1; i <= n; i++) {
1350 mdb_dpage_free(env, dl[i].mptr);
1355 /* Set or clear P_KEEP in non-overflow, non-sub pages in known cursors.
1356 * When clearing, only consider backup cursors (from parent txns) since
1357 * other P_KEEP flags have already been cleared.
1358 * @param[in] mc A cursor handle for the current operation.
1359 * @param[in] pflags Flags of the pages to update:
1360 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1363 mdb_cursorpages_mark(MDB_cursor *mc, unsigned pflags)
1365 MDB_txn *txn = mc->mc_txn;
1366 MDB_cursor *m2, *m3;
1370 if (mc->mc_flags & C_UNTRACK)
1371 mc = NULL; /* will find mc in mt_cursors */
1372 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1373 for (; mc; mc=mc->mc_next) {
1374 m2 = pflags == P_DIRTY ? mc : mc->mc_backup;
1375 for (; m2; m2 = m2->mc_backup) {
1376 for (m3=m2; m3->mc_flags & C_INITIALIZED; m3=&mx->mx_cursor) {
1377 for (j=0; j<m3->mc_snum; j++)
1378 if ((m3->mc_pg[j]->mp_flags & (P_SUBP|P_DIRTY|P_KEEP))
1380 m3->mc_pg[j]->mp_flags ^= P_KEEP;
1381 if (!(m3->mc_db->md_flags & MDB_DUPSORT))
1383 /* Cursor backups have mx malloced at the end of m2 */
1384 mx = (m3 == mc ? m3->mc_xcursor : (MDB_xcursor *)(m3+1));
1393 static int mdb_page_flush(MDB_txn *txn);
1395 /** Spill pages from the dirty list back to disk.
1396 * This is intended to prevent running into #MDB_TXN_FULL situations,
1397 * but note that they may still occur in a few cases:
1398 * 1) pages in #MDB_DUPSORT sub-DBs are never spilled, so if there
1399 * are too many of these dirtied in one txn, the txn may still get
1401 * 2) child txns may run out of space if their parents dirtied a
1402 * lot of pages and never spilled them. TODO: we probably should do
1403 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1404 * the parent's dirty_room is below a given threshold.
1405 * 3) our estimate of the txn size could be too small. At the
1406 * moment this seems unlikely.
1408 * Otherwise, if not using nested txns, it is expected that apps will
1409 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1410 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1411 * If the txn never references them again, they can be left alone.
1412 * If the txn only reads them, they can be used without any fuss.
1413 * If the txn writes them again, they can be dirtied immediately without
1414 * going thru all of the work of #mdb_page_touch(). Such references are
1415 * handled by #mdb_page_unspill().
1417 * Also note, we never spill DB root pages, nor pages of active cursors,
1418 * because we'll need these back again soon anyway. And in nested txns,
1419 * we can't spill a page in a child txn if it was already spilled in a
1420 * parent txn. That would alter the parent txns' data even though
1421 * the child hasn't committed yet, and we'd have no way to undo it if
1422 * the child aborted.
1424 * @param[in] m0 cursor A cursor handle identifying the transaction and
1425 * database for which we are checking space.
1426 * @param[in] key For a put operation, the key being stored.
1427 * @param[in] data For a put operation, the data being stored.
1428 * @return 0 on success, non-zero on failure.
1431 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1433 MDB_txn *txn = m0->mc_txn;
1435 MDB_ID2L dl = txn->mt_u.dirty_list;
1439 if (m0->mc_flags & C_SUB)
1442 /* Estimate how much space this op will take */
1443 i = m0->mc_db->md_depth;
1444 /* Named DBs also dirty the main DB */
1445 if (m0->mc_dbi > MAIN_DBI)
1446 i += txn->mt_dbs[MAIN_DBI].md_depth;
1447 /* For puts, roughly factor in the key+data size */
1449 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1450 i += i; /* double it for good measure */
1452 if (txn->mt_dirty_room > i)
1455 if (!txn->mt_spill_pgs) {
1456 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1457 if (!txn->mt_spill_pgs)
1461 /* Mark all the dirty root pages we want to preserve */
1462 for (i=0; i<txn->mt_numdbs; i++) {
1463 if (txn->mt_dbflags[i] & DB_DIRTY) {
1464 pgno_t pgno = txn->mt_dbs[i].md_root;
1465 if (pgno == P_INVALID)
1467 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1469 if ((dp->mp_flags & P_DIRTY) && level <= 1)
1470 dp->mp_flags |= P_KEEP;
1474 /* Preserve pages used by cursors */
1475 mdb_cursorpages_mark(m0, P_DIRTY);
1477 /* Save the page IDs of all the pages we're flushing */
1478 for (i=1; i<=dl[0].mid; i++) {
1480 if (dp->mp_flags & P_KEEP)
1482 /* Can't spill twice, make sure it's not already in a parent's
1485 if (txn->mt_parent) {
1487 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1488 if (tx2->mt_spill_pgs) {
1489 j = mdb_midl_search(tx2->mt_spill_pgs, dl[i].mid);
1490 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == dl[i].mid) {
1491 dp->mp_flags |= P_KEEP;
1499 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, dl[i].mid)))
1502 mdb_midl_sort(txn->mt_spill_pgs);
1504 rc = mdb_page_flush(txn);
1506 mdb_cursorpages_mark(m0, P_DIRTY|P_KEEP);
1510 if (txn->mt_parent) {
1512 pgno_t pgno = dl[i].mid;
1513 txn->mt_dirty_room = txn->mt_parent->mt_dirty_room - dl[0].mid;
1514 /* dirty pages that are dirty in an ancestor don't
1515 * count against this txn's dirty_room.
1517 for (i=1; i<=dl[0].mid; i++) {
1518 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1519 j = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1520 if (j <= tx2->mt_u.dirty_list[0].mid &&
1521 tx2->mt_u.dirty_list[j].mid == pgno) {
1522 txn->mt_dirty_room++;
1528 txn->mt_dirty_room = MDB_IDL_UM_MAX - dl[0].mid;
1530 txn->mt_flags |= MDB_TXN_SPILLS;
1532 txn->mt_flags |= MDB_TXN_ERROR;
1537 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1539 mdb_find_oldest(MDB_txn *txn)
1542 txnid_t mr, oldest = txn->mt_txnid - 1;
1543 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1544 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1554 /** Add a page to the txn's dirty list */
1556 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1559 int (*insert)(MDB_ID2L, MDB_ID2 *);
1561 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1562 insert = mdb_mid2l_append;
1564 insert = mdb_mid2l_insert;
1566 mid.mid = mp->mp_pgno;
1568 insert(txn->mt_u.dirty_list, &mid);
1569 txn->mt_dirty_room--;
1572 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1573 * me_pghead and mt_next_pgno.
1575 * If there are free pages available from older transactions, they
1576 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1577 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1578 * and move me_pglast to say which records were consumed. Only this
1579 * function can create me_pghead and move me_pglast/mt_next_pgno.
1580 * @param[in] mc cursor A cursor handle identifying the transaction and
1581 * database for which we are allocating.
1582 * @param[in] num the number of pages to allocate.
1583 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1584 * will always be satisfied by a single contiguous chunk of memory.
1585 * @return 0 on success, non-zero on failure.
1588 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1590 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1591 /* Get at most <Max_retries> more freeDB records once me_pghead
1592 * has enough pages. If not enough, use new pages from the map.
1593 * If <Paranoid> and mc is updating the freeDB, only get new
1594 * records if me_pghead is empty. Then the freelist cannot play
1595 * catch-up with itself by growing while trying to save it.
1597 enum { Paranoid = 1, Max_retries = 500 };
1599 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1601 int rc, n2 = num-1, retry = Max_retries;
1602 MDB_txn *txn = mc->mc_txn;
1603 MDB_env *env = txn->mt_env;
1604 pgno_t pgno, *mop = env->me_pghead;
1605 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1607 txnid_t oldest = 0, last;
1613 /* If our dirty list is already full, we can't do anything */
1614 if (txn->mt_dirty_room == 0)
1615 return MDB_TXN_FULL;
1617 for (op = MDB_FIRST;; op = MDB_NEXT) {
1620 pgno_t *idl, old_id, new_id;
1622 /* Seek a big enough contiguous page range. Prefer
1623 * pages at the tail, just truncating the list.
1625 if (mop_len >= (unsigned)num) {
1629 if (mop[i-n2] == pgno+n2)
1631 } while (--i >= (unsigned)num);
1632 if (Max_retries < INT_MAX && --retry < 0)
1636 if (op == MDB_FIRST) { /* 1st iteration */
1637 /* Prepare to fetch more and coalesce */
1638 oldest = mdb_find_oldest(txn);
1639 last = env->me_pglast;
1640 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1643 key.mv_data = &last; /* will look up last+1 */
1644 key.mv_size = sizeof(last);
1646 if (Paranoid && mc->mc_dbi == FREE_DBI)
1649 if (Paranoid && retry < 0 && mop_len)
1653 /* Do not fetch more if the record will be too recent */
1656 rc = mdb_cursor_get(&m2, &key, NULL, op);
1658 if (rc == MDB_NOTFOUND)
1662 last = *(txnid_t*)key.mv_data;
1665 np = m2.mc_pg[m2.mc_top];
1666 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1667 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1670 idl = (MDB_ID *) data.mv_data;
1673 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1676 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1678 mop = env->me_pghead;
1680 env->me_pglast = last;
1682 DPRINTF("IDL read txn %"Z"u root %"Z"u num %u",
1683 last, txn->mt_dbs[FREE_DBI].md_root, i);
1685 DPRINTF("IDL %"Z"u", idl[k]);
1687 /* Merge in descending sorted order */
1690 mop[0] = (pgno_t)-1;
1694 for (; old_id < new_id; old_id = mop[--j])
1701 /* Use new pages from the map when nothing suitable in the freeDB */
1703 pgno = txn->mt_next_pgno;
1704 if (pgno + num >= env->me_maxpg) {
1705 DPUTS("DB size maxed out");
1706 return MDB_MAP_FULL;
1710 if (env->me_flags & MDB_WRITEMAP) {
1711 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1713 if (!(np = mdb_page_malloc(txn, num)))
1717 mop[0] = mop_len -= num;
1718 /* Move any stragglers down */
1719 for (j = i-num; j < mop_len; )
1720 mop[++j] = mop[++i];
1722 txn->mt_next_pgno = pgno + num;
1725 mdb_page_dirty(txn, np);
1731 /** Copy the used portions of a non-overflow page.
1732 * @param[in] dst page to copy into
1733 * @param[in] src page to copy from
1734 * @param[in] psize size of a page
1737 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1739 enum { Align = sizeof(pgno_t) };
1740 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1742 /* If page isn't full, just copy the used portion. Adjust
1743 * alignment so memcpy may copy words instead of bytes.
1745 if ((unused &= -Align) && !IS_LEAF2(src)) {
1747 memcpy(dst, src, (lower + (Align-1)) & -Align);
1748 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1751 memcpy(dst, src, psize - unused);
1755 /** Pull a page off the txn's spill list, if present.
1756 * If a page being referenced was spilled to disk in this txn, bring
1757 * it back and make it dirty/writable again.
1758 * @param[in] tx0 the transaction handle.
1759 * @param[in] mp the page being referenced.
1760 * @param[out] ret the writable page, if any. ret is unchanged if
1761 * mp wasn't spilled.
1764 mdb_page_unspill(MDB_txn *tx0, MDB_page *mp, MDB_page **ret)
1766 MDB_env *env = tx0->mt_env;
1769 pgno_t pgno = mp->mp_pgno;
1771 for (txn = tx0; txn; txn=txn->mt_parent) {
1772 if (!txn->mt_spill_pgs)
1774 x = mdb_midl_search(txn->mt_spill_pgs, pgno);
1775 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pgno) {
1778 if (IS_OVERFLOW(mp))
1782 if (env->me_flags & MDB_WRITEMAP) {
1785 np = mdb_page_malloc(txn, num);
1789 memcpy(np, mp, num * env->me_psize);
1791 mdb_page_copy(np, mp, env->me_psize);
1794 /* If in current txn, this page is no longer spilled */
1795 for (; x < txn->mt_spill_pgs[0]; x++)
1796 txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
1797 txn->mt_spill_pgs[0]--;
1798 } /* otherwise, if belonging to a parent txn, the
1799 * page remains spilled until child commits
1802 if (txn->mt_parent) {
1804 /* If this page is also in a parent's dirty list, then
1805 * it's already accounted in dirty_room, and we need to
1806 * cancel out the decrement that mdb_page_dirty does.
1808 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1809 x = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1810 if (x <= tx2->mt_u.dirty_list[0].mid &&
1811 tx2->mt_u.dirty_list[x].mid == pgno) {
1812 txn->mt_dirty_room++;
1817 mdb_page_dirty(tx0, np);
1818 np->mp_flags |= P_DIRTY;
1826 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1827 * @param[in] mc cursor pointing to the page to be touched
1828 * @return 0 on success, non-zero on failure.
1831 mdb_page_touch(MDB_cursor *mc)
1833 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1834 MDB_txn *txn = mc->mc_txn;
1835 MDB_cursor *m2, *m3;
1840 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1841 if (txn->mt_flags & MDB_TXN_SPILLS) {
1843 rc = mdb_page_unspill(txn, mp, &np);
1849 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1850 (rc = mdb_page_alloc(mc, 1, &np)))
1853 DPRINTF("touched db %u page %"Z"u -> %"Z"u", mc->mc_dbi,mp->mp_pgno,pgno);
1854 assert(mp->mp_pgno != pgno);
1855 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1856 /* Update the parent page, if any, to point to the new page */
1858 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1859 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1860 SETPGNO(node, pgno);
1862 mc->mc_db->md_root = pgno;
1864 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1865 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1867 /* If txn has a parent, make sure the page is in our
1871 unsigned x = mdb_mid2l_search(dl, pgno);
1872 if (x <= dl[0].mid && dl[x].mid == pgno) {
1873 if (mp != dl[x].mptr) { /* bad cursor? */
1874 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1875 return MDB_CORRUPTED;
1880 assert(dl[0].mid < MDB_IDL_UM_MAX);
1882 np = mdb_page_malloc(txn, 1);
1887 mdb_mid2l_insert(dl, &mid);
1892 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1894 np->mp_flags |= P_DIRTY;
1897 /* Adjust cursors pointing to mp */
1898 mc->mc_pg[mc->mc_top] = np;
1900 if (mc->mc_flags & C_SUB) {
1902 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1903 m3 = &m2->mc_xcursor->mx_cursor;
1904 if (m3->mc_snum < mc->mc_snum) continue;
1905 if (m3->mc_pg[mc->mc_top] == mp)
1906 m3->mc_pg[mc->mc_top] = np;
1909 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1910 if (m2->mc_snum < mc->mc_snum) continue;
1911 if (m2->mc_pg[mc->mc_top] == mp) {
1912 m2->mc_pg[mc->mc_top] = np;
1913 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1914 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1916 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1917 if (!(leaf->mn_flags & F_SUBDATA))
1918 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1927 mdb_env_sync(MDB_env *env, int force)
1930 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1931 if (env->me_flags & MDB_WRITEMAP) {
1932 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1933 ? MS_ASYNC : MS_SYNC;
1934 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1937 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1941 if (MDB_FDATASYNC(env->me_fd))
1948 /** Back up parent txn's cursors, then grab the originals for tracking */
1950 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1952 MDB_cursor *mc, *bk;
1957 for (i = src->mt_numdbs; --i >= 0; ) {
1958 if ((mc = src->mt_cursors[i]) != NULL) {
1959 size = sizeof(MDB_cursor);
1961 size += sizeof(MDB_xcursor);
1962 for (; mc; mc = bk->mc_next) {
1968 mc->mc_db = &dst->mt_dbs[i];
1969 /* Kill pointers into src - and dst to reduce abuse: The
1970 * user may not use mc until dst ends. Otherwise we'd...
1972 mc->mc_txn = NULL; /* ...set this to dst */
1973 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
1974 if ((mx = mc->mc_xcursor) != NULL) {
1975 *(MDB_xcursor *)(bk+1) = *mx;
1976 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
1978 mc->mc_next = dst->mt_cursors[i];
1979 dst->mt_cursors[i] = mc;
1986 /** Close this write txn's cursors, give parent txn's cursors back to parent.
1987 * @param[in] txn the transaction handle.
1988 * @param[in] merge true to keep changes to parent cursors, false to revert.
1989 * @return 0 on success, non-zero on failure.
1992 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1994 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
1998 for (i = txn->mt_numdbs; --i >= 0; ) {
1999 for (mc = cursors[i]; mc; mc = next) {
2001 if ((bk = mc->mc_backup) != NULL) {
2003 /* Commit changes to parent txn */
2004 mc->mc_next = bk->mc_next;
2005 mc->mc_backup = bk->mc_backup;
2006 mc->mc_txn = bk->mc_txn;
2007 mc->mc_db = bk->mc_db;
2008 mc->mc_dbflag = bk->mc_dbflag;
2009 if ((mx = mc->mc_xcursor) != NULL)
2010 mx->mx_cursor.mc_txn = bk->mc_txn;
2012 /* Abort nested txn */
2014 if ((mx = mc->mc_xcursor) != NULL)
2015 *mx = *(MDB_xcursor *)(bk+1);
2019 /* Only malloced cursors are permanently tracked. */
2026 #ifdef MDB_DEBUG_SKIP
2027 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2030 mdb_txn_reset0(MDB_txn *txn, const char *act);
2032 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2038 Pidset = F_SETLK, Pidcheck = F_GETLK
2042 /** Set or check a pid lock. Set returns 0 on success.
2043 * Check returns 0 if the process is certainly dead, nonzero if it may
2044 * be alive (the lock exists or an error happened so we do not know).
2046 * On Windows Pidset is a no-op, we merely check for the existence
2047 * of the process with the given pid. On POSIX we use a single byte
2048 * lock on the lockfile, set at an offset equal to the pid.
2051 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2053 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2056 if (op == Pidcheck) {
2057 h = OpenProcess(env->me_pidquery, FALSE, pid);
2058 /* No documented "no such process" code, but other program use this: */
2060 return ErrCode() != ERROR_INVALID_PARAMETER;
2061 /* A process exists until all handles to it close. Has it exited? */
2062 ret = WaitForSingleObject(h, 0) != 0;
2069 struct flock lock_info;
2070 memset(&lock_info, 0, sizeof(lock_info));
2071 lock_info.l_type = F_WRLCK;
2072 lock_info.l_whence = SEEK_SET;
2073 lock_info.l_start = pid;
2074 lock_info.l_len = 1;
2075 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2076 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2078 } else if ((rc = ErrCode()) == EINTR) {
2086 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2087 * @param[in] txn the transaction handle to initialize
2088 * @return 0 on success, non-zero on failure.
2091 mdb_txn_renew0(MDB_txn *txn)
2093 MDB_env *env = txn->mt_env;
2096 int rc, new_notls = 0;
2099 txn->mt_numdbs = env->me_numdbs;
2100 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2102 if (txn->mt_flags & MDB_TXN_RDONLY) {
2103 if (!env->me_txns) {
2104 i = mdb_env_pick_meta(env);
2105 txn->mt_txnid = env->me_metas[i]->mm_txnid;
2106 txn->mt_u.reader = NULL;
2108 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2109 pthread_getspecific(env->me_txkey);
2111 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2112 return MDB_BAD_RSLOT;
2114 pid_t pid = env->me_pid;
2115 pthread_t tid = pthread_self();
2117 if (!(env->me_flags & MDB_LIVE_READER)) {
2118 rc = mdb_reader_pid(env, Pidset, pid);
2120 UNLOCK_MUTEX_R(env);
2123 env->me_flags |= MDB_LIVE_READER;
2127 for (i=0; i<env->me_txns->mti_numreaders; i++)
2128 if (env->me_txns->mti_readers[i].mr_pid == 0)
2130 if (i == env->me_maxreaders) {
2131 UNLOCK_MUTEX_R(env);
2132 return MDB_READERS_FULL;
2134 env->me_txns->mti_readers[i].mr_pid = pid;
2135 env->me_txns->mti_readers[i].mr_tid = tid;
2136 if (i >= env->me_txns->mti_numreaders)
2137 env->me_txns->mti_numreaders = i+1;
2138 /* Save numreaders for un-mutexed mdb_env_close() */
2139 env->me_numreaders = env->me_txns->mti_numreaders;
2140 UNLOCK_MUTEX_R(env);
2141 r = &env->me_txns->mti_readers[i];
2142 new_notls = (env->me_flags & MDB_NOTLS);
2143 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2148 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2149 txn->mt_u.reader = r;
2151 txn->mt_toggle = txn->mt_txnid & 1;
2155 txn->mt_txnid = env->me_txns->mti_txnid;
2156 txn->mt_toggle = txn->mt_txnid & 1;
2159 if (txn->mt_txnid == mdb_debug_start)
2162 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2163 txn->mt_u.dirty_list = env->me_dirty_list;
2164 txn->mt_u.dirty_list[0].mid = 0;
2165 txn->mt_free_pgs = env->me_free_pgs;
2166 txn->mt_free_pgs[0] = 0;
2167 txn->mt_spill_pgs = NULL;
2171 /* Copy the DB info and flags */
2172 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
2174 /* Moved to here to avoid a data race in read TXNs */
2175 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
2177 for (i=2; i<txn->mt_numdbs; i++) {
2178 x = env->me_dbflags[i];
2179 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2180 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2182 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2184 if (env->me_maxpg < txn->mt_next_pgno) {
2185 mdb_txn_reset0(txn, "renew0-mapfail");
2187 txn->mt_u.reader->mr_pid = 0;
2188 txn->mt_u.reader = NULL;
2190 return MDB_MAP_RESIZED;
2197 mdb_txn_renew(MDB_txn *txn)
2201 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2204 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2205 DPUTS("environment had fatal error, must shutdown!");
2209 rc = mdb_txn_renew0(txn);
2210 if (rc == MDB_SUCCESS) {
2211 DPRINTF("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2212 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2213 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2219 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2223 int rc, size, tsize = sizeof(MDB_txn);
2225 if (env->me_flags & MDB_FATAL_ERROR) {
2226 DPUTS("environment had fatal error, must shutdown!");
2229 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2232 /* Nested transactions: Max 1 child, write txns only, no writemap */
2233 if (parent->mt_child ||
2234 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
2235 (env->me_flags & MDB_WRITEMAP))
2239 tsize = sizeof(MDB_ntxn);
2241 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2242 if (!(flags & MDB_RDONLY))
2243 size += env->me_maxdbs * sizeof(MDB_cursor *);
2245 if ((txn = calloc(1, size)) == NULL) {
2246 DPRINTF("calloc: %s", strerror(ErrCode()));
2249 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2250 if (flags & MDB_RDONLY) {
2251 txn->mt_flags |= MDB_TXN_RDONLY;
2252 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2254 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2255 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2261 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2262 if (!txn->mt_u.dirty_list ||
2263 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2265 free(txn->mt_u.dirty_list);
2269 txn->mt_txnid = parent->mt_txnid;
2270 txn->mt_toggle = parent->mt_toggle;
2271 txn->mt_dirty_room = parent->mt_dirty_room;
2272 txn->mt_u.dirty_list[0].mid = 0;
2273 txn->mt_spill_pgs = NULL;
2274 txn->mt_next_pgno = parent->mt_next_pgno;
2275 parent->mt_child = txn;
2276 txn->mt_parent = parent;
2277 txn->mt_numdbs = parent->mt_numdbs;
2278 txn->mt_flags = parent->mt_flags;
2279 txn->mt_dbxs = parent->mt_dbxs;
2280 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2281 /* Copy parent's mt_dbflags, but clear DB_NEW */
2282 for (i=0; i<txn->mt_numdbs; i++)
2283 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2285 ntxn = (MDB_ntxn *)txn;
2286 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2287 if (env->me_pghead) {
2288 size = MDB_IDL_SIZEOF(env->me_pghead);
2289 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2291 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2296 rc = mdb_cursor_shadow(parent, txn);
2298 mdb_txn_reset0(txn, "beginchild-fail");
2300 rc = mdb_txn_renew0(txn);
2306 DPRINTF("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2307 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2308 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
2314 /** Export or close DBI handles opened in this txn. */
2316 mdb_dbis_update(MDB_txn *txn, int keep)
2319 MDB_dbi n = txn->mt_numdbs;
2320 MDB_env *env = txn->mt_env;
2321 unsigned char *tdbflags = txn->mt_dbflags;
2323 for (i = n; --i >= 2;) {
2324 if (tdbflags[i] & DB_NEW) {
2326 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2328 char *ptr = env->me_dbxs[i].md_name.mv_data;
2329 env->me_dbxs[i].md_name.mv_data = NULL;
2330 env->me_dbxs[i].md_name.mv_size = 0;
2331 env->me_dbflags[i] = 0;
2336 if (keep && env->me_numdbs < n)
2340 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2341 * May be called twice for readonly txns: First reset it, then abort.
2342 * @param[in] txn the transaction handle to reset
2343 * @param[in] act why the transaction is being reset
2346 mdb_txn_reset0(MDB_txn *txn, const char *act)
2348 MDB_env *env = txn->mt_env;
2350 /* Close any DBI handles opened in this txn */
2351 mdb_dbis_update(txn, 0);
2353 DPRINTF("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2354 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2355 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2357 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2358 if (txn->mt_u.reader) {
2359 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2360 if (!(env->me_flags & MDB_NOTLS))
2361 txn->mt_u.reader = NULL; /* txn does not own reader */
2363 txn->mt_numdbs = 0; /* close nothing if called again */
2364 txn->mt_dbxs = NULL; /* mark txn as reset */
2366 mdb_cursors_close(txn, 0);
2368 if (!(env->me_flags & MDB_WRITEMAP)) {
2369 mdb_dlist_free(txn);
2371 mdb_midl_free(env->me_pghead);
2373 if (txn->mt_parent) {
2374 txn->mt_parent->mt_child = NULL;
2375 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2376 mdb_midl_free(txn->mt_free_pgs);
2377 mdb_midl_free(txn->mt_spill_pgs);
2378 free(txn->mt_u.dirty_list);
2382 if (mdb_midl_shrink(&txn->mt_free_pgs))
2383 env->me_free_pgs = txn->mt_free_pgs;
2384 env->me_pghead = NULL;
2388 /* The writer mutex was locked in mdb_txn_begin. */
2389 UNLOCK_MUTEX_W(env);
2394 mdb_txn_reset(MDB_txn *txn)
2399 /* This call is only valid for read-only txns */
2400 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2403 mdb_txn_reset0(txn, "reset");
2407 mdb_txn_abort(MDB_txn *txn)
2413 mdb_txn_abort(txn->mt_child);
2415 mdb_txn_reset0(txn, "abort");
2416 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2417 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2418 txn->mt_u.reader->mr_pid = 0;
2423 /** Save the freelist as of this transaction to the freeDB.
2424 * This changes the freelist. Keep trying until it stabilizes.
2427 mdb_freelist_save(MDB_txn *txn)
2429 /* env->me_pghead[] can grow and shrink during this call.
2430 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2431 * Page numbers cannot disappear from txn->mt_free_pgs[].
2434 MDB_env *env = txn->mt_env;
2435 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2436 txnid_t pglast = 0, head_id = 0;
2437 pgno_t freecnt = 0, *free_pgs, *mop;
2438 ssize_t head_room = 0, total_room = 0, mop_len;
2440 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2442 if (env->me_pghead) {
2443 /* Make sure first page of freeDB is touched and on freelist */
2444 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2445 if (rc && rc != MDB_NOTFOUND)
2450 /* Come back here after each Put() in case freelist changed */
2453 /* If using records from freeDB which we have not yet
2454 * deleted, delete them and any we reserved for me_pghead.
2456 while (pglast < env->me_pglast) {
2457 rc = mdb_cursor_first(&mc, &key, NULL);
2460 pglast = head_id = *(txnid_t *)key.mv_data;
2461 total_room = head_room = 0;
2462 assert(pglast <= env->me_pglast);
2463 rc = mdb_cursor_del(&mc, 0);
2468 /* Save the IDL of pages freed by this txn, to a single record */
2469 if (freecnt < txn->mt_free_pgs[0]) {
2471 /* Make sure last page of freeDB is touched and on freelist */
2472 key.mv_size = MDB_MAXKEYSIZE+1;
2474 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2475 if (rc && rc != MDB_NOTFOUND)
2478 free_pgs = txn->mt_free_pgs;
2479 /* Write to last page of freeDB */
2480 key.mv_size = sizeof(txn->mt_txnid);
2481 key.mv_data = &txn->mt_txnid;
2483 freecnt = free_pgs[0];
2484 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2485 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2488 /* Retry if mt_free_pgs[] grew during the Put() */
2489 free_pgs = txn->mt_free_pgs;
2490 } while (freecnt < free_pgs[0]);
2491 mdb_midl_sort(free_pgs);
2492 memcpy(data.mv_data, free_pgs, data.mv_size);
2495 unsigned int i = free_pgs[0];
2496 DPRINTF("IDL write txn %"Z"u root %"Z"u num %u",
2497 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2499 DPRINTF("IDL %"Z"u", free_pgs[i]);
2505 mop = env->me_pghead;
2506 mop_len = mop ? mop[0] : 0;
2508 /* Reserve records for me_pghead[]. Split it if multi-page,
2509 * to avoid searching freeDB for a page range. Use keys in
2510 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2512 if (total_room >= mop_len) {
2513 if (total_room == mop_len || --more < 0)
2515 } else if (head_room >= maxfree_1pg && head_id > 1) {
2516 /* Keep current record (overflow page), add a new one */
2520 /* (Re)write {key = head_id, IDL length = head_room} */
2521 total_room -= head_room;
2522 head_room = mop_len - total_room;
2523 if (head_room > maxfree_1pg && head_id > 1) {
2524 /* Overflow multi-page for part of me_pghead */
2525 head_room /= head_id; /* amortize page sizes */
2526 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2527 } else if (head_room < 0) {
2528 /* Rare case, not bothering to delete this record */
2531 key.mv_size = sizeof(head_id);
2532 key.mv_data = &head_id;
2533 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2534 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2537 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2538 total_room += head_room;
2541 /* Fill in the reserved, touched me_pghead records */
2547 rc = mdb_cursor_first(&mc, &key, &data);
2548 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2549 unsigned flags = MDB_CURRENT;
2550 txnid_t id = *(txnid_t *)key.mv_data;
2551 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2554 assert(len >= 0 && id <= env->me_pglast);
2556 if (len > mop_len) {
2558 data.mv_size = (len + 1) * sizeof(MDB_ID);
2561 data.mv_data = mop -= len;
2564 rc = mdb_cursor_put(&mc, &key, &data, flags);
2566 if (rc || !(mop_len -= len))
2573 /** Flush dirty pages to the map, after clearing their dirty flag.
2576 mdb_page_flush(MDB_txn *txn)
2578 MDB_env *env = txn->mt_env;
2579 MDB_ID2L dl = txn->mt_u.dirty_list;
2580 unsigned psize = env->me_psize, j;
2581 int i, pagecount = dl[0].mid, rc;
2582 size_t size = 0, pos = 0;
2584 MDB_page *dp = NULL;
2588 struct iovec iov[MDB_COMMIT_PAGES];
2589 ssize_t wpos = 0, wsize = 0, wres;
2590 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2595 if (env->me_flags & MDB_WRITEMAP) {
2596 /* Clear dirty flags */
2597 for (i=1; i<=pagecount; i++) {
2599 /* Don't flush this page yet */
2600 if (dp->mp_flags & P_KEEP) {
2601 dp->mp_flags ^= P_KEEP;
2605 dp->mp_flags &= ~P_DIRTY;
2611 /* Write the pages */
2613 if (i <= pagecount) {
2615 /* Don't flush this page yet */
2616 if (dp->mp_flags & P_KEEP) {
2617 dp->mp_flags ^= P_KEEP;
2622 /* clear dirty flag */
2623 dp->mp_flags &= ~P_DIRTY;
2626 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2631 /* Windows actually supports scatter/gather I/O, but only on
2632 * unbuffered file handles. Since we're relying on the OS page
2633 * cache for all our data, that's self-defeating. So we just
2634 * write pages one at a time. We use the ov structure to set
2635 * the write offset, to at least save the overhead of a Seek
2638 DPRINTF("committing page %"Z"u", pgno);
2639 memset(&ov, 0, sizeof(ov));
2640 ov.Offset = pos & 0xffffffff;
2641 ov.OffsetHigh = pos >> 16 >> 16;
2642 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2644 DPRINTF("WriteFile: %d", rc);
2648 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2649 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2651 /* Write previous page(s) */
2652 #ifdef MDB_USE_PWRITEV
2653 wres = pwritev(env->me_fd, iov, n, wpos);
2656 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2658 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2660 DPRINTF("lseek: %s", strerror(rc));
2663 wres = writev(env->me_fd, iov, n);
2666 if (wres != wsize) {
2669 DPRINTF("Write error: %s", strerror(rc));
2671 rc = EIO; /* TODO: Use which error code? */
2672 DPUTS("short write, filesystem full?");
2683 DPRINTF("committing page %"Z"u", pgno);
2684 next_pos = pos + size;
2685 iov[n].iov_len = size;
2686 iov[n].iov_base = (char *)dp;
2693 for (i=1; i<=pagecount; i++) {
2695 /* This is a page we skipped above */
2698 dl[j].mid = dp->mp_pgno;
2701 mdb_dpage_free(env, dp);
2709 mdb_txn_commit(MDB_txn *txn)
2715 assert(txn != NULL);
2716 assert(txn->mt_env != NULL);
2718 if (txn->mt_child) {
2719 rc = mdb_txn_commit(txn->mt_child);
2720 txn->mt_child = NULL;
2727 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2728 mdb_dbis_update(txn, 1);
2729 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2734 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2735 DPUTS("error flag is set, can't commit");
2737 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2742 if (txn->mt_parent) {
2743 MDB_txn *parent = txn->mt_parent;
2747 /* Append our free list to parent's */
2748 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2751 mdb_midl_free(txn->mt_free_pgs);
2753 parent->mt_next_pgno = txn->mt_next_pgno;
2754 parent->mt_flags = txn->mt_flags;
2756 /* Merge our cursors into parent's and close them */
2757 mdb_cursors_close(txn, 1);
2759 /* Update parent's DB table. */
2760 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2761 parent->mt_numdbs = txn->mt_numdbs;
2762 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2763 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2764 for (i=2; i<txn->mt_numdbs; i++) {
2765 /* preserve parent's DB_NEW status */
2766 x = parent->mt_dbflags[i] & DB_NEW;
2767 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2770 dst = parent->mt_u.dirty_list;
2771 src = txn->mt_u.dirty_list;
2772 /* Remove anything in our dirty list from parent's spill list */
2773 if (parent->mt_spill_pgs) {
2774 x = parent->mt_spill_pgs[0];
2776 /* zero out our dirty pages in parent spill list */
2777 for (i=1; i<=src[0].mid; i++) {
2778 if (src[i].mid < parent->mt_spill_pgs[x])
2780 if (src[i].mid > parent->mt_spill_pgs[x]) {
2786 parent->mt_spill_pgs[x] = 0;
2789 /* OK, we had a few hits, squash zeros from the spill list */
2790 if (len < parent->mt_spill_pgs[0]) {
2792 for (y=1; y<=parent->mt_spill_pgs[0]; y++) {
2793 if (parent->mt_spill_pgs[y]) {
2795 parent->mt_spill_pgs[x] = parent->mt_spill_pgs[y];
2800 parent->mt_spill_pgs[0] = len;
2803 /* Find len = length of merging our dirty list with parent's */
2805 dst[0].mid = 0; /* simplify loops */
2806 if (parent->mt_parent) {
2807 len = x + src[0].mid;
2808 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2809 for (i = x; y && i; y--) {
2810 pgno_t yp = src[y].mid;
2811 while (yp < dst[i].mid)
2813 if (yp == dst[i].mid) {
2818 } else { /* Simplify the above for single-ancestor case */
2819 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2821 /* Merge our dirty list with parent's */
2823 for (i = len; y; dst[i--] = src[y--]) {
2824 pgno_t yp = src[y].mid;
2825 while (yp < dst[x].mid)
2826 dst[i--] = dst[x--];
2827 if (yp == dst[x].mid)
2828 free(dst[x--].mptr);
2832 free(txn->mt_u.dirty_list);
2833 parent->mt_dirty_room = txn->mt_dirty_room;
2834 if (txn->mt_spill_pgs) {
2835 if (parent->mt_spill_pgs) {
2836 mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2837 mdb_midl_free(txn->mt_spill_pgs);
2838 mdb_midl_sort(parent->mt_spill_pgs);
2840 parent->mt_spill_pgs = txn->mt_spill_pgs;
2844 parent->mt_child = NULL;
2845 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2850 if (txn != env->me_txn) {
2851 DPUTS("attempt to commit unknown transaction");
2856 mdb_cursors_close(txn, 0);
2858 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2861 DPRINTF("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2862 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2864 /* Update DB root pointers */
2865 if (txn->mt_numdbs > 2) {
2869 data.mv_size = sizeof(MDB_db);
2871 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2872 for (i = 2; i < txn->mt_numdbs; i++) {
2873 if (txn->mt_dbflags[i] & DB_DIRTY) {
2874 data.mv_data = &txn->mt_dbs[i];
2875 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2882 rc = mdb_freelist_save(txn);
2886 mdb_midl_free(env->me_pghead);
2887 env->me_pghead = NULL;
2888 if (mdb_midl_shrink(&txn->mt_free_pgs))
2889 env->me_free_pgs = txn->mt_free_pgs;
2895 if ((rc = mdb_page_flush(txn)) ||
2896 (rc = mdb_env_sync(env, 0)) ||
2897 (rc = mdb_env_write_meta(txn)))
2903 mdb_dbis_update(txn, 1);
2905 UNLOCK_MUTEX_W(env);
2915 /** Read the environment parameters of a DB environment before
2916 * mapping it into memory.
2917 * @param[in] env the environment handle
2918 * @param[out] meta address of where to store the meta information
2919 * @return 0 on success, non-zero on failure.
2922 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2929 /* We don't know the page size yet, so use a minimum value.
2930 * Read both meta pages so we can use the latest one.
2933 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2937 memset(&ov, 0, sizeof(ov));
2939 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2940 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2943 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2945 if (rc != MDB_PAGESIZE) {
2946 if (rc == 0 && off == 0)
2948 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2949 DPRINTF("read: %s", mdb_strerror(rc));
2953 p = (MDB_page *)&pbuf;
2955 if (!F_ISSET(p->mp_flags, P_META)) {
2956 DPRINTF("page %"Z"u not a meta page", p->mp_pgno);
2961 if (m->mm_magic != MDB_MAGIC) {
2962 DPUTS("meta has invalid magic");
2966 if (m->mm_version != MDB_DATA_VERSION) {
2967 DPRINTF("database is version %u, expected version %u",
2968 m->mm_version, MDB_DATA_VERSION);
2969 return MDB_VERSION_MISMATCH;
2972 if (off == 0 || m->mm_txnid > meta->mm_txnid)
2978 /** Write the environment parameters of a freshly created DB environment.
2979 * @param[in] env the environment handle
2980 * @param[out] meta address of where to store the meta information
2981 * @return 0 on success, non-zero on failure.
2984 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2992 memset(&ov, 0, sizeof(ov));
2993 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
2995 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
2998 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
2999 len = pwrite(fd, ptr, size, pos); \
3000 rc = (len >= 0); } while(0)
3003 DPUTS("writing new meta page");
3005 GET_PAGESIZE(psize);
3007 meta->mm_magic = MDB_MAGIC;
3008 meta->mm_version = MDB_DATA_VERSION;
3009 meta->mm_mapsize = env->me_mapsize;
3010 meta->mm_psize = psize;
3011 meta->mm_last_pg = 1;
3012 meta->mm_flags = env->me_flags & 0xffff;
3013 meta->mm_flags |= MDB_INTEGERKEY;
3014 meta->mm_dbs[0].md_root = P_INVALID;
3015 meta->mm_dbs[1].md_root = P_INVALID;
3017 p = calloc(2, psize);
3019 p->mp_flags = P_META;
3020 *(MDB_meta *)METADATA(p) = *meta;
3022 q = (MDB_page *)((char *)p + psize);
3024 q->mp_flags = P_META;
3025 *(MDB_meta *)METADATA(q) = *meta;
3027 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3030 else if ((unsigned) len == psize * 2)
3038 /** Update the environment info to commit a transaction.
3039 * @param[in] txn the transaction that's being committed
3040 * @return 0 on success, non-zero on failure.
3043 mdb_env_write_meta(MDB_txn *txn)
3046 MDB_meta meta, metab, *mp;
3048 int rc, len, toggle;
3057 assert(txn != NULL);
3058 assert(txn->mt_env != NULL);
3060 toggle = !txn->mt_toggle;
3061 DPRINTF("writing meta page %d for root page %"Z"u",
3062 toggle, txn->mt_dbs[MAIN_DBI].md_root);
3065 mp = env->me_metas[toggle];
3067 if (env->me_flags & MDB_WRITEMAP) {
3068 /* Persist any increases of mapsize config */
3069 if (env->me_mapsize > mp->mm_mapsize)
3070 mp->mm_mapsize = env->me_mapsize;
3071 mp->mm_dbs[0] = txn->mt_dbs[0];
3072 mp->mm_dbs[1] = txn->mt_dbs[1];
3073 mp->mm_last_pg = txn->mt_next_pgno - 1;
3074 mp->mm_txnid = txn->mt_txnid;
3075 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3076 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3079 ptr += env->me_psize;
3080 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3087 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3088 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3090 ptr = (char *)&meta;
3091 if (env->me_mapsize > mp->mm_mapsize) {
3092 /* Persist any increases of mapsize config */
3093 meta.mm_mapsize = env->me_mapsize;
3094 off = offsetof(MDB_meta, mm_mapsize);
3096 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3098 len = sizeof(MDB_meta) - off;
3101 meta.mm_dbs[0] = txn->mt_dbs[0];
3102 meta.mm_dbs[1] = txn->mt_dbs[1];
3103 meta.mm_last_pg = txn->mt_next_pgno - 1;
3104 meta.mm_txnid = txn->mt_txnid;
3107 off += env->me_psize;
3110 /* Write to the SYNC fd */
3111 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3112 env->me_fd : env->me_mfd;
3115 memset(&ov, 0, sizeof(ov));
3117 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3121 rc = pwrite(mfd, ptr, len, off);
3124 rc = rc < 0 ? ErrCode() : EIO;
3125 DPUTS("write failed, disk error?");
3126 /* On a failure, the pagecache still contains the new data.
3127 * Write some old data back, to prevent it from being used.
3128 * Use the non-SYNC fd; we know it will fail anyway.
3130 meta.mm_last_pg = metab.mm_last_pg;
3131 meta.mm_txnid = metab.mm_txnid;
3133 memset(&ov, 0, sizeof(ov));
3135 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3137 r2 = pwrite(env->me_fd, ptr, len, off);
3140 env->me_flags |= MDB_FATAL_ERROR;
3144 /* Memory ordering issues are irrelevant; since the entire writer
3145 * is wrapped by wmutex, all of these changes will become visible
3146 * after the wmutex is unlocked. Since the DB is multi-version,
3147 * readers will get consistent data regardless of how fresh or
3148 * how stale their view of these values is.
3150 env->me_txns->mti_txnid = txn->mt_txnid;
3155 /** Check both meta pages to see which one is newer.
3156 * @param[in] env the environment handle
3157 * @return meta toggle (0 or 1).
3160 mdb_env_pick_meta(const MDB_env *env)
3162 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3166 mdb_env_create(MDB_env **env)
3170 e = calloc(1, sizeof(MDB_env));
3174 e->me_maxreaders = DEFAULT_READERS;
3175 e->me_maxdbs = e->me_numdbs = 2;
3176 e->me_fd = INVALID_HANDLE_VALUE;
3177 e->me_lfd = INVALID_HANDLE_VALUE;
3178 e->me_mfd = INVALID_HANDLE_VALUE;
3179 #ifdef MDB_USE_POSIX_SEM
3180 e->me_rmutex = SEM_FAILED;
3181 e->me_wmutex = SEM_FAILED;
3183 e->me_pid = getpid();
3184 VGMEMP_CREATE(e,0,0);
3190 mdb_env_set_mapsize(MDB_env *env, size_t size)
3194 env->me_mapsize = size;
3196 env->me_maxpg = env->me_mapsize / env->me_psize;
3201 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3205 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3210 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3212 if (env->me_map || readers < 1)
3214 env->me_maxreaders = readers;
3219 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3221 if (!env || !readers)
3223 *readers = env->me_maxreaders;
3227 /** Further setup required for opening an MDB environment
3230 mdb_env_open2(MDB_env *env)
3232 unsigned int flags = env->me_flags;
3240 memset(&meta, 0, sizeof(meta));
3242 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3245 DPUTS("new mdbenv");
3249 /* Was a mapsize configured? */
3250 if (!env->me_mapsize) {
3251 /* If this is a new environment, take the default,
3252 * else use the size recorded in the existing env.
3254 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3255 } else if (env->me_mapsize < meta.mm_mapsize) {
3256 /* If the configured size is smaller, make sure it's
3257 * still big enough. Silently round up to minimum if not.
3259 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3260 if (env->me_mapsize < minsize)
3261 env->me_mapsize = minsize;
3268 LONG sizelo, sizehi;
3269 sizelo = env->me_mapsize & 0xffffffff;
3270 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3272 /* See if we should use QueryLimited */
3274 if ((rc & 0xff) > 5)
3275 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3277 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3279 /* Windows won't create mappings for zero length files.
3280 * Just allocate the maxsize right now.
3283 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3284 || !SetEndOfFile(env->me_fd)
3285 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3288 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3289 PAGE_READWRITE : PAGE_READONLY,
3290 sizehi, sizelo, NULL);
3293 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3294 FILE_MAP_WRITE : FILE_MAP_READ,
3295 0, 0, env->me_mapsize, meta.mm_address);
3296 rc = env->me_map ? 0 : ErrCode();
3304 if (flags & MDB_WRITEMAP) {
3306 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3309 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
3311 if (env->me_map == MAP_FAILED) {
3315 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3317 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3319 #ifdef POSIX_MADV_RANDOM
3320 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3321 #endif /* POSIX_MADV_RANDOM */
3322 #endif /* MADV_RANDOM */
3326 if (flags & MDB_FIXEDMAP)
3327 meta.mm_address = env->me_map;
3328 i = mdb_env_init_meta(env, &meta);
3329 if (i != MDB_SUCCESS) {
3332 } else if (meta.mm_address && env->me_map != meta.mm_address) {
3333 /* Can happen because the address argument to mmap() is just a
3334 * hint. mmap() can pick another, e.g. if the range is in use.
3335 * The MAP_FIXED flag would prevent that, but then mmap could
3336 * instead unmap existing pages to make room for the new map.
3338 return EBUSY; /* TODO: Make a new MDB_* error code? */
3340 env->me_psize = meta.mm_psize;
3341 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3342 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3344 env->me_maxpg = env->me_mapsize / env->me_psize;
3346 p = (MDB_page *)env->me_map;
3347 env->me_metas[0] = METADATA(p);
3348 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
3352 int toggle = mdb_env_pick_meta(env);
3353 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3355 DPRINTF("opened database version %u, pagesize %u",
3356 env->me_metas[0]->mm_version, env->me_psize);
3357 DPRINTF("using meta page %d", toggle);
3358 DPRINTF("depth: %u", db->md_depth);
3359 DPRINTF("entries: %"Z"u", db->md_entries);
3360 DPRINTF("branch pages: %"Z"u", db->md_branch_pages);
3361 DPRINTF("leaf pages: %"Z"u", db->md_leaf_pages);
3362 DPRINTF("overflow pages: %"Z"u", db->md_overflow_pages);
3363 DPRINTF("root: %"Z"u", db->md_root);
3371 /** Release a reader thread's slot in the reader lock table.
3372 * This function is called automatically when a thread exits.
3373 * @param[in] ptr This points to the slot in the reader lock table.
3376 mdb_env_reader_dest(void *ptr)
3378 MDB_reader *reader = ptr;
3384 /** Junk for arranging thread-specific callbacks on Windows. This is
3385 * necessarily platform and compiler-specific. Windows supports up
3386 * to 1088 keys. Let's assume nobody opens more than 64 environments
3387 * in a single process, for now. They can override this if needed.
3389 #ifndef MAX_TLS_KEYS
3390 #define MAX_TLS_KEYS 64
3392 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3393 static int mdb_tls_nkeys;
3395 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3399 case DLL_PROCESS_ATTACH: break;
3400 case DLL_THREAD_ATTACH: break;
3401 case DLL_THREAD_DETACH:
3402 for (i=0; i<mdb_tls_nkeys; i++) {
3403 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3404 mdb_env_reader_dest(r);
3407 case DLL_PROCESS_DETACH: break;
3412 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3414 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3418 /* Force some symbol references.
3419 * _tls_used forces the linker to create the TLS directory if not already done
3420 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3422 #pragma comment(linker, "/INCLUDE:_tls_used")
3423 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3424 #pragma const_seg(".CRT$XLB")
3425 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3426 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3429 #pragma comment(linker, "/INCLUDE:__tls_used")
3430 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3431 #pragma data_seg(".CRT$XLB")
3432 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3434 #endif /* WIN 32/64 */
3435 #endif /* !__GNUC__ */
3438 /** Downgrade the exclusive lock on the region back to shared */
3440 mdb_env_share_locks(MDB_env *env, int *excl)
3442 int rc = 0, toggle = mdb_env_pick_meta(env);
3444 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3449 /* First acquire a shared lock. The Unlock will
3450 * then release the existing exclusive lock.
3452 memset(&ov, 0, sizeof(ov));
3453 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3456 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3462 struct flock lock_info;
3463 /* The shared lock replaces the existing lock */
3464 memset((void *)&lock_info, 0, sizeof(lock_info));
3465 lock_info.l_type = F_RDLCK;
3466 lock_info.l_whence = SEEK_SET;
3467 lock_info.l_start = 0;
3468 lock_info.l_len = 1;
3469 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3470 (rc = ErrCode()) == EINTR) ;
3471 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3478 /** Try to get exlusive lock, otherwise shared.
3479 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3482 mdb_env_excl_lock(MDB_env *env, int *excl)
3486 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3490 memset(&ov, 0, sizeof(ov));
3491 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3498 struct flock lock_info;
3499 memset((void *)&lock_info, 0, sizeof(lock_info));
3500 lock_info.l_type = F_WRLCK;
3501 lock_info.l_whence = SEEK_SET;
3502 lock_info.l_start = 0;
3503 lock_info.l_len = 1;
3504 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3505 (rc = ErrCode()) == EINTR) ;
3509 # ifdef MDB_USE_POSIX_SEM
3510 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3513 lock_info.l_type = F_RDLCK;
3514 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3515 (rc = ErrCode()) == EINTR) ;
3523 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3525 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3527 * @(#) $Revision: 5.1 $
3528 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3529 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3531 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3535 * Please do not copyright this code. This code is in the public domain.
3537 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3538 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3539 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3540 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3541 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3542 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3543 * PERFORMANCE OF THIS SOFTWARE.
3546 * chongo <Landon Curt Noll> /\oo/\
3547 * http://www.isthe.com/chongo/
3549 * Share and Enjoy! :-)
3552 typedef unsigned long long mdb_hash_t;
3553 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3555 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3556 * @param[in] val value to hash
3557 * @param[in] hval initial value for hash
3558 * @return 64 bit hash
3560 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3561 * hval arg on the first call.
3564 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3566 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3567 unsigned char *end = s + val->mv_size;
3569 * FNV-1a hash each octet of the string
3572 /* xor the bottom with the current octet */
3573 hval ^= (mdb_hash_t)*s++;
3575 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3576 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3577 (hval << 7) + (hval << 8) + (hval << 40);
3579 /* return our new hash value */
3583 /** Hash the string and output the encoded hash.
3584 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3585 * very short name limits. We don't care about the encoding being reversible,
3586 * we just want to preserve as many bits of the input as possible in a
3587 * small printable string.
3588 * @param[in] str string to hash
3589 * @param[out] encbuf an array of 11 chars to hold the hash
3591 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3594 mdb_pack85(unsigned long l, char *out)
3598 for (i=0; i<5; i++) {
3599 *out++ = mdb_a85[l % 85];
3605 mdb_hash_enc(MDB_val *val, char *encbuf)
3607 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3608 unsigned long *l = (unsigned long *)&h;
3610 mdb_pack85(l[0], encbuf);
3611 mdb_pack85(l[1], encbuf+5);
3616 /** Open and/or initialize the lock region for the environment.
3617 * @param[in] env The MDB environment.
3618 * @param[in] lpath The pathname of the file used for the lock region.
3619 * @param[in] mode The Unix permissions for the file, if we create it.
3620 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3621 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3622 * @return 0 on success, non-zero on failure.
3625 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3628 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3630 # define MDB_ERRCODE_ROFS EROFS
3631 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3632 # define MDB_CLOEXEC O_CLOEXEC
3635 # define MDB_CLOEXEC 0
3642 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3643 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3644 FILE_ATTRIBUTE_NORMAL, NULL);
3646 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3648 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3650 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3655 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3656 /* Lose record locks when exec*() */
3657 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3658 fcntl(env->me_lfd, F_SETFD, fdflags);
3661 if (!(env->me_flags & MDB_NOTLS)) {
3662 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3665 env->me_flags |= MDB_ENV_TXKEY;
3667 /* Windows TLS callbacks need help finding their TLS info. */
3668 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3672 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3676 /* Try to get exclusive lock. If we succeed, then
3677 * nobody is using the lock region and we should initialize it.
3679 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3682 size = GetFileSize(env->me_lfd, NULL);
3684 size = lseek(env->me_lfd, 0, SEEK_END);
3685 if (size == -1) goto fail_errno;
3687 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3688 if (size < rsize && *excl > 0) {
3690 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3691 || !SetEndOfFile(env->me_lfd))
3694 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3698 size = rsize - sizeof(MDB_txninfo);
3699 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3704 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3706 if (!mh) goto fail_errno;
3707 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3709 if (!env->me_txns) goto fail_errno;
3711 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3713 if (m == MAP_FAILED) goto fail_errno;
3719 BY_HANDLE_FILE_INFORMATION stbuf;
3728 if (!mdb_sec_inited) {
3729 InitializeSecurityDescriptor(&mdb_null_sd,
3730 SECURITY_DESCRIPTOR_REVISION);
3731 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3732 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3733 mdb_all_sa.bInheritHandle = FALSE;
3734 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3737 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3738 idbuf.volume = stbuf.dwVolumeSerialNumber;
3739 idbuf.nhigh = stbuf.nFileIndexHigh;
3740 idbuf.nlow = stbuf.nFileIndexLow;
3741 val.mv_data = &idbuf;
3742 val.mv_size = sizeof(idbuf);
3743 mdb_hash_enc(&val, encbuf);
3744 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3745 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3746 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3747 if (!env->me_rmutex) goto fail_errno;
3748 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3749 if (!env->me_wmutex) goto fail_errno;
3750 #elif defined(MDB_USE_POSIX_SEM)
3759 #if defined(__NetBSD__)
3760 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3762 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3763 idbuf.dev = stbuf.st_dev;
3764 idbuf.ino = stbuf.st_ino;
3765 val.mv_data = &idbuf;
3766 val.mv_size = sizeof(idbuf);
3767 mdb_hash_enc(&val, encbuf);
3768 #ifdef MDB_SHORT_SEMNAMES
3769 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3771 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3772 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3773 /* Clean up after a previous run, if needed: Try to
3774 * remove both semaphores before doing anything else.
3776 sem_unlink(env->me_txns->mti_rmname);
3777 sem_unlink(env->me_txns->mti_wmname);
3778 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3779 O_CREAT|O_EXCL, mode, 1);
3780 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3781 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3782 O_CREAT|O_EXCL, mode, 1);
3783 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3784 #else /* MDB_USE_POSIX_SEM */
3785 pthread_mutexattr_t mattr;
3787 if ((rc = pthread_mutexattr_init(&mattr))
3788 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3789 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3790 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3792 pthread_mutexattr_destroy(&mattr);
3793 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3795 env->me_txns->mti_magic = MDB_MAGIC;
3796 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3797 env->me_txns->mti_txnid = 0;
3798 env->me_txns->mti_numreaders = 0;
3801 if (env->me_txns->mti_magic != MDB_MAGIC) {
3802 DPUTS("lock region has invalid magic");
3806 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3807 DPRINTF("lock region has format+version 0x%x, expected 0x%x",
3808 env->me_txns->mti_format, MDB_LOCK_FORMAT);
3809 rc = MDB_VERSION_MISMATCH;
3813 if (rc && rc != EACCES && rc != EAGAIN) {
3817 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3818 if (!env->me_rmutex) goto fail_errno;
3819 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3820 if (!env->me_wmutex) goto fail_errno;
3821 #elif defined(MDB_USE_POSIX_SEM)
3822 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3823 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3824 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3825 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3836 /** The name of the lock file in the DB environment */
3837 #define LOCKNAME "/lock.mdb"
3838 /** The name of the data file in the DB environment */
3839 #define DATANAME "/data.mdb"
3840 /** The suffix of the lock file when no subdir is used */
3841 #define LOCKSUFF "-lock"
3842 /** Only a subset of the @ref mdb_env flags can be changed
3843 * at runtime. Changing other flags requires closing the
3844 * environment and re-opening it with the new flags.
3846 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3847 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3850 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3852 int oflags, rc, len, excl = -1;
3853 char *lpath, *dpath;
3855 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3859 if (flags & MDB_NOSUBDIR) {
3860 rc = len + sizeof(LOCKSUFF) + len + 1;
3862 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3867 if (flags & MDB_NOSUBDIR) {
3868 dpath = lpath + len + sizeof(LOCKSUFF);
3869 sprintf(lpath, "%s" LOCKSUFF, path);
3870 strcpy(dpath, path);
3872 dpath = lpath + len + sizeof(LOCKNAME);
3873 sprintf(lpath, "%s" LOCKNAME, path);
3874 sprintf(dpath, "%s" DATANAME, path);
3878 flags |= env->me_flags;
3879 if (flags & MDB_RDONLY) {
3880 /* silently ignore WRITEMAP when we're only getting read access */
3881 flags &= ~MDB_WRITEMAP;
3883 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3884 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3887 env->me_flags = flags |= MDB_ENV_ACTIVE;
3891 env->me_path = strdup(path);
3892 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3893 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3894 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3899 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3904 if (F_ISSET(flags, MDB_RDONLY)) {
3905 oflags = GENERIC_READ;
3906 len = OPEN_EXISTING;
3908 oflags = GENERIC_READ|GENERIC_WRITE;
3911 mode = FILE_ATTRIBUTE_NORMAL;
3912 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3913 NULL, len, mode, NULL);
3915 if (F_ISSET(flags, MDB_RDONLY))
3918 oflags = O_RDWR | O_CREAT;
3920 env->me_fd = open(dpath, oflags, mode);
3922 if (env->me_fd == INVALID_HANDLE_VALUE) {
3927 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3928 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3929 env->me_mfd = env->me_fd;
3931 /* Synchronous fd for meta writes. Needed even with
3932 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3935 env->me_mfd = CreateFile(dpath, oflags,
3936 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3937 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3939 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3941 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3946 DPRINTF("opened dbenv %p", (void *) env);
3948 rc = mdb_env_share_locks(env, &excl);
3954 mdb_env_close0(env, excl);
3960 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3962 mdb_env_close0(MDB_env *env, int excl)
3966 if (!(env->me_flags & MDB_ENV_ACTIVE))
3969 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3970 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3971 free(env->me_dbxs[i].md_name.mv_data);
3973 free(env->me_dbflags);
3976 free(env->me_dirty_list);
3977 mdb_midl_free(env->me_free_pgs);
3979 if (env->me_flags & MDB_ENV_TXKEY) {
3980 pthread_key_delete(env->me_txkey);
3982 /* Delete our key from the global list */
3983 for (i=0; i<mdb_tls_nkeys; i++)
3984 if (mdb_tls_keys[i] == env->me_txkey) {
3985 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3993 munmap(env->me_map, env->me_mapsize);
3995 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3996 (void) close(env->me_mfd);
3997 if (env->me_fd != INVALID_HANDLE_VALUE)
3998 (void) close(env->me_fd);
4000 pid_t pid = env->me_pid;
4001 /* Clearing readers is done in this function because
4002 * me_txkey with its destructor must be disabled first.
4004 for (i = env->me_numreaders; --i >= 0; )
4005 if (env->me_txns->mti_readers[i].mr_pid == pid)
4006 env->me_txns->mti_readers[i].mr_pid = 0;
4008 if (env->me_rmutex) {
4009 CloseHandle(env->me_rmutex);
4010 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4012 /* Windows automatically destroys the mutexes when
4013 * the last handle closes.
4015 #elif defined(MDB_USE_POSIX_SEM)
4016 if (env->me_rmutex != SEM_FAILED) {
4017 sem_close(env->me_rmutex);
4018 if (env->me_wmutex != SEM_FAILED)
4019 sem_close(env->me_wmutex);
4020 /* If we have the filelock: If we are the
4021 * only remaining user, clean up semaphores.
4024 mdb_env_excl_lock(env, &excl);
4026 sem_unlink(env->me_txns->mti_rmname);
4027 sem_unlink(env->me_txns->mti_wmname);
4031 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4033 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4036 /* Unlock the lockfile. Windows would have unlocked it
4037 * after closing anyway, but not necessarily at once.
4039 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4042 (void) close(env->me_lfd);
4045 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4049 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4051 MDB_txn *txn = NULL;
4057 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4061 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4064 /* Do the lock/unlock of the reader mutex before starting the
4065 * write txn. Otherwise other read txns could block writers.
4067 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4072 /* We must start the actual read txn after blocking writers */
4073 mdb_txn_reset0(txn, "reset-stage1");
4075 /* Temporarily block writers until we snapshot the meta pages */
4078 rc = mdb_txn_renew0(txn);
4080 UNLOCK_MUTEX_W(env);
4085 wsize = env->me_psize * 2;
4089 DO_WRITE(rc, fd, ptr, w2, len);
4093 } else if (len > 0) {
4099 /* Non-blocking or async handles are not supported */
4105 UNLOCK_MUTEX_W(env);
4110 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4112 if (wsize > MAX_WRITE)
4116 DO_WRITE(rc, fd, ptr, w2, len);
4120 } else if (len > 0) {
4137 mdb_env_copy(MDB_env *env, const char *path)
4141 HANDLE newfd = INVALID_HANDLE_VALUE;
4143 if (env->me_flags & MDB_NOSUBDIR) {
4144 lpath = (char *)path;
4147 len += sizeof(DATANAME);
4148 lpath = malloc(len);
4151 sprintf(lpath, "%s" DATANAME, path);
4154 /* The destination path must exist, but the destination file must not.
4155 * We don't want the OS to cache the writes, since the source data is
4156 * already in the OS cache.
4159 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4160 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4162 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
4168 if (newfd == INVALID_HANDLE_VALUE) {
4173 #ifdef F_NOCACHE /* __APPLE__ */
4174 rc = fcntl(newfd, F_NOCACHE, 1);
4181 rc = mdb_env_copyfd(env, newfd);
4184 if (!(env->me_flags & MDB_NOSUBDIR))
4186 if (newfd != INVALID_HANDLE_VALUE)
4187 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4194 mdb_env_close(MDB_env *env)
4201 VGMEMP_DESTROY(env);
4202 while ((dp = env->me_dpages) != NULL) {
4203 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4204 env->me_dpages = dp->mp_next;
4208 mdb_env_close0(env, 0);
4212 /** Compare two items pointing at aligned size_t's */
4214 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4216 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4217 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4220 /** Compare two items pointing at aligned int's */
4222 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4224 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4225 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4228 /** Compare two items pointing at ints of unknown alignment.
4229 * Nodes and keys are guaranteed to be 2-byte aligned.
4232 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4234 #if BYTE_ORDER == LITTLE_ENDIAN
4235 unsigned short *u, *c;
4238 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4239 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4242 } while(!x && u > (unsigned short *)a->mv_data);
4245 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4249 /** Compare two items lexically */
4251 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4258 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4264 diff = memcmp(a->mv_data, b->mv_data, len);
4265 return diff ? diff : len_diff<0 ? -1 : len_diff;
4268 /** Compare two items in reverse byte order */
4270 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4272 const unsigned char *p1, *p2, *p1_lim;
4276 p1_lim = (const unsigned char *)a->mv_data;
4277 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4278 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4280 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4286 while (p1 > p1_lim) {
4287 diff = *--p1 - *--p2;
4291 return len_diff<0 ? -1 : len_diff;
4294 /** Search for key within a page, using binary search.
4295 * Returns the smallest entry larger or equal to the key.
4296 * If exactp is non-null, stores whether the found entry was an exact match
4297 * in *exactp (1 or 0).
4298 * Updates the cursor index with the index of the found entry.
4299 * If no entry larger or equal to the key is found, returns NULL.
4302 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4304 unsigned int i = 0, nkeys;
4307 MDB_page *mp = mc->mc_pg[mc->mc_top];
4308 MDB_node *node = NULL;
4313 nkeys = NUMKEYS(mp);
4318 COPY_PGNO(pgno, mp->mp_pgno);
4319 DPRINTF("searching %u keys in %s %spage %"Z"u",
4320 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4327 low = IS_LEAF(mp) ? 0 : 1;
4329 cmp = mc->mc_dbx->md_cmp;
4331 /* Branch pages have no data, so if using integer keys,
4332 * alignment is guaranteed. Use faster mdb_cmp_int.
4334 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4335 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4342 nodekey.mv_size = mc->mc_db->md_pad;
4343 node = NODEPTR(mp, 0); /* fake */
4344 while (low <= high) {
4345 i = (low + high) >> 1;
4346 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4347 rc = cmp(key, &nodekey);
4348 DPRINTF("found leaf index %u [%s], rc = %i",
4349 i, DKEY(&nodekey), rc);
4358 while (low <= high) {
4359 i = (low + high) >> 1;
4361 node = NODEPTR(mp, i);
4362 nodekey.mv_size = NODEKSZ(node);
4363 nodekey.mv_data = NODEKEY(node);
4365 rc = cmp(key, &nodekey);
4368 DPRINTF("found leaf index %u [%s], rc = %i",
4369 i, DKEY(&nodekey), rc);
4371 DPRINTF("found branch index %u [%s -> %"Z"u], rc = %i",
4372 i, DKEY(&nodekey), NODEPGNO(node), rc);
4383 if (rc > 0) { /* Found entry is less than the key. */
4384 i++; /* Skip to get the smallest entry larger than key. */
4386 node = NODEPTR(mp, i);
4389 *exactp = (rc == 0);
4390 /* store the key index */
4391 mc->mc_ki[mc->mc_top] = i;
4393 /* There is no entry larger or equal to the key. */
4396 /* nodeptr is fake for LEAF2 */
4402 mdb_cursor_adjust(MDB_cursor *mc, func)
4406 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4407 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4414 /** Pop a page off the top of the cursor's stack. */
4416 mdb_cursor_pop(MDB_cursor *mc)
4419 #ifndef MDB_DEBUG_SKIP
4420 MDB_page *top = mc->mc_pg[mc->mc_top];
4426 DPRINTF("popped page %"Z"u off db %u cursor %p", top->mp_pgno,
4427 mc->mc_dbi, (void *) mc);
4431 /** Push a page onto the top of the cursor's stack. */
4433 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4435 DPRINTF("pushing page %"Z"u on db %u cursor %p", mp->mp_pgno,
4436 mc->mc_dbi, (void *) mc);
4438 if (mc->mc_snum >= CURSOR_STACK) {
4439 assert(mc->mc_snum < CURSOR_STACK);
4440 return MDB_CURSOR_FULL;
4443 mc->mc_top = mc->mc_snum++;
4444 mc->mc_pg[mc->mc_top] = mp;
4445 mc->mc_ki[mc->mc_top] = 0;
4450 /** Find the address of the page corresponding to a given page number.
4451 * @param[in] txn the transaction for this access.
4452 * @param[in] pgno the page number for the page to retrieve.
4453 * @param[out] ret address of a pointer where the page's address will be stored.
4454 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4455 * @return 0 on success, non-zero on failure.
4458 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4460 MDB_env *env = txn->mt_env;
4464 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4468 MDB_ID2L dl = tx2->mt_u.dirty_list;
4470 /* Spilled pages were dirtied in this txn and flushed
4471 * because the dirty list got full. Bring this page
4472 * back in from the map (but don't unspill it here,
4473 * leave that unless page_touch happens again).
4475 if (tx2->mt_spill_pgs) {
4476 x = mdb_midl_search(tx2->mt_spill_pgs, pgno);
4477 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pgno) {
4478 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4483 unsigned x = mdb_mid2l_search(dl, pgno);
4484 if (x <= dl[0].mid && dl[x].mid == pgno) {
4490 } while ((tx2 = tx2->mt_parent) != NULL);
4493 if (pgno < txn->mt_next_pgno) {
4495 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4497 DPRINTF("page %"Z"u not found", pgno);
4499 return MDB_PAGE_NOTFOUND;
4509 /** Search for the page a given key should be in.
4510 * Pushes parent pages on the cursor stack. This function continues a
4511 * search on a cursor that has already been initialized. (Usually by
4512 * #mdb_page_search() but also by #mdb_node_move().)
4513 * @param[in,out] mc the cursor for this operation.
4514 * @param[in] key the key to search for. If NULL, search for the lowest
4515 * page. (This is used by #mdb_cursor_first().)
4516 * @param[in] modify If true, visited pages are updated with new page numbers.
4517 * @return 0 on success, non-zero on failure.
4520 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4522 MDB_page *mp = mc->mc_pg[mc->mc_top];
4527 while (IS_BRANCH(mp)) {
4531 DPRINTF("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp));
4532 assert(NUMKEYS(mp) > 1);
4533 DPRINTF("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0)));
4535 if (key == NULL) /* Initialize cursor to first page. */
4537 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4538 /* cursor to last page */
4542 node = mdb_node_search(mc, key, &exact);
4544 i = NUMKEYS(mp) - 1;
4546 i = mc->mc_ki[mc->mc_top];
4555 DPRINTF("following index %u for key [%s]",
4557 assert(i < NUMKEYS(mp));
4558 node = NODEPTR(mp, i);
4560 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4563 mc->mc_ki[mc->mc_top] = i;
4564 if ((rc = mdb_cursor_push(mc, mp)))
4568 if ((rc = mdb_page_touch(mc)) != 0)
4570 mp = mc->mc_pg[mc->mc_top];
4575 DPRINTF("internal error, index points to a %02X page!?",
4577 return MDB_CORRUPTED;
4580 DPRINTF("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4581 key ? DKEY(key) : NULL);
4582 mc->mc_flags |= C_INITIALIZED;
4583 mc->mc_flags &= ~C_EOF;
4588 /** Search for the lowest key under the current branch page.
4589 * This just bypasses a NUMKEYS check in the current page
4590 * before calling mdb_page_search_root(), because the callers
4591 * are all in situations where the current page is known to
4595 mdb_page_search_lowest(MDB_cursor *mc)
4597 MDB_page *mp = mc->mc_pg[mc->mc_top];
4598 MDB_node *node = NODEPTR(mp, 0);
4601 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4604 mc->mc_ki[mc->mc_top] = 0;
4605 if ((rc = mdb_cursor_push(mc, mp)))
4607 return mdb_page_search_root(mc, NULL, 0);
4610 /** Search for the page a given key should be in.
4611 * Pushes parent pages on the cursor stack. This function just sets up
4612 * the search; it finds the root page for \b mc's database and sets this
4613 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4614 * called to complete the search.
4615 * @param[in,out] mc the cursor for this operation.
4616 * @param[in] key the key to search for. If NULL, search for the lowest
4617 * page. (This is used by #mdb_cursor_first().)
4618 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4619 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4620 * @return 0 on success, non-zero on failure.
4623 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4628 /* Make sure the txn is still viable, then find the root from
4629 * the txn's db table.
4631 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4632 DPUTS("transaction has failed, must abort");
4635 /* Make sure we're using an up-to-date root */
4636 if (mc->mc_dbi > MAIN_DBI) {
4637 if ((*mc->mc_dbflag & DB_STALE) ||
4638 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4640 unsigned char dbflag = 0;
4641 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4642 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4645 if (*mc->mc_dbflag & DB_STALE) {
4649 MDB_node *leaf = mdb_node_search(&mc2,
4650 &mc->mc_dbx->md_name, &exact);
4652 return MDB_NOTFOUND;
4653 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4656 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4658 /* The txn may not know this DBI, or another process may
4659 * have dropped and recreated the DB with other flags.
4661 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4662 return MDB_INCOMPATIBLE;
4663 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4665 if (flags & MDB_PS_MODIFY)
4667 *mc->mc_dbflag &= ~DB_STALE;
4668 *mc->mc_dbflag |= dbflag;
4671 root = mc->mc_db->md_root;
4673 if (root == P_INVALID) { /* Tree is empty. */
4674 DPUTS("tree is empty");
4675 return MDB_NOTFOUND;
4680 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4681 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4687 DPRINTF("db %u root page %"Z"u has flags 0x%X",
4688 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4690 if (flags & MDB_PS_MODIFY) {
4691 if ((rc = mdb_page_touch(mc)))
4695 if (flags & MDB_PS_ROOTONLY)
4698 return mdb_page_search_root(mc, key, flags);
4702 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4704 MDB_txn *txn = mc->mc_txn;
4705 pgno_t pg = mp->mp_pgno;
4706 unsigned i, ovpages = mp->mp_pages;
4707 MDB_env *env = txn->mt_env;
4710 DPRINTF("free ov page %"Z"u (%d)", pg, ovpages);
4711 /* If the page is dirty or on the spill list we just acquired it,
4712 * so we should give it back to our current free list, if any.
4713 * Not currently supported in nested txns.
4714 * Otherwise put it onto the list of pages we freed in this txn.
4716 if (!(mp->mp_flags & P_DIRTY) && txn->mt_spill_pgs) {
4717 unsigned x = mdb_midl_search(txn->mt_spill_pgs, pg);
4718 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pg) {
4719 /* This page is no longer spilled */
4720 for (; x < txn->mt_spill_pgs[0]; x++)
4721 txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
4722 txn->mt_spill_pgs[0]--;
4726 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && env->me_pghead) {
4729 MDB_ID2 *dl, ix, iy;
4730 rc = mdb_midl_need(&env->me_pghead, ovpages);
4733 /* Remove from dirty list */
4734 dl = txn->mt_u.dirty_list;
4736 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4744 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4745 txn->mt_flags |= MDB_TXN_ERROR;
4746 return MDB_CORRUPTED;
4749 if (!(env->me_flags & MDB_WRITEMAP))
4750 mdb_dpage_free(env, mp);
4752 /* Insert in me_pghead */
4753 mop = env->me_pghead;
4754 j = mop[0] + ovpages;
4755 for (i = mop[0]; i && mop[i] < pg; i--)
4761 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4765 mc->mc_db->md_overflow_pages -= ovpages;
4769 /** Return the data associated with a given node.
4770 * @param[in] txn The transaction for this operation.
4771 * @param[in] leaf The node being read.
4772 * @param[out] data Updated to point to the node's data.
4773 * @return 0 on success, non-zero on failure.
4776 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4778 MDB_page *omp; /* overflow page */
4782 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4783 data->mv_size = NODEDSZ(leaf);
4784 data->mv_data = NODEDATA(leaf);
4788 /* Read overflow data.
4790 data->mv_size = NODEDSZ(leaf);
4791 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4792 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4793 DPRINTF("read overflow page %"Z"u failed", pgno);
4796 data->mv_data = METADATA(omp);
4802 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4803 MDB_val *key, MDB_val *data)
4812 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4814 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4817 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4821 mdb_cursor_init(&mc, txn, dbi, &mx);
4822 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4825 /** Find a sibling for a page.
4826 * Replaces the page at the top of the cursor's stack with the
4827 * specified sibling, if one exists.
4828 * @param[in] mc The cursor for this operation.
4829 * @param[in] move_right Non-zero if the right sibling is requested,
4830 * otherwise the left sibling.
4831 * @return 0 on success, non-zero on failure.
4834 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4840 if (mc->mc_snum < 2) {
4841 return MDB_NOTFOUND; /* root has no siblings */
4845 DPRINTF("parent page is page %"Z"u, index %u",
4846 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4848 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4849 : (mc->mc_ki[mc->mc_top] == 0)) {
4850 DPRINTF("no more keys left, moving to %s sibling",
4851 move_right ? "right" : "left");
4852 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4853 /* undo cursor_pop before returning */
4860 mc->mc_ki[mc->mc_top]++;
4862 mc->mc_ki[mc->mc_top]--;
4863 DPRINTF("just moving to %s index key %u",
4864 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4866 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4868 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4869 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4872 mdb_cursor_push(mc, mp);
4874 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4879 /** Move the cursor to the next data item. */
4881 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4887 if (mc->mc_flags & C_EOF) {
4888 return MDB_NOTFOUND;
4891 assert(mc->mc_flags & C_INITIALIZED);
4893 mp = mc->mc_pg[mc->mc_top];
4895 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4896 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4897 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4898 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4899 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4900 if (op != MDB_NEXT || rc != MDB_NOTFOUND)
4904 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4905 if (op == MDB_NEXT_DUP)
4906 return MDB_NOTFOUND;
4910 DPRINTF("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc);
4912 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4913 DPUTS("=====> move to next sibling page");
4914 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4915 mc->mc_flags |= C_EOF;
4918 mp = mc->mc_pg[mc->mc_top];
4919 DPRINTF("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4921 mc->mc_ki[mc->mc_top]++;
4923 DPRINTF("==> cursor points to page %"Z"u with %u keys, key index %u",
4924 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4927 key->mv_size = mc->mc_db->md_pad;
4928 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4932 assert(IS_LEAF(mp));
4933 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4935 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4936 mdb_xcursor_init1(mc, leaf);
4939 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4942 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4943 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4944 if (rc != MDB_SUCCESS)
4949 MDB_GET_KEY(leaf, key);
4953 /** Move the cursor to the previous data item. */
4955 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4961 assert(mc->mc_flags & C_INITIALIZED);
4963 mp = mc->mc_pg[mc->mc_top];
4965 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4966 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4967 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4968 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4969 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4970 if (op != MDB_PREV || rc != MDB_NOTFOUND)
4973 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4974 if (op == MDB_PREV_DUP)
4975 return MDB_NOTFOUND;
4980 DPRINTF("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc);
4982 if (mc->mc_ki[mc->mc_top] == 0) {
4983 DPUTS("=====> move to prev sibling page");
4984 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
4987 mp = mc->mc_pg[mc->mc_top];
4988 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4989 DPRINTF("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4991 mc->mc_ki[mc->mc_top]--;
4993 mc->mc_flags &= ~C_EOF;
4995 DPRINTF("==> cursor points to page %"Z"u with %u keys, key index %u",
4996 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4999 key->mv_size = mc->mc_db->md_pad;
5000 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5004 assert(IS_LEAF(mp));
5005 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5007 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5008 mdb_xcursor_init1(mc, leaf);
5011 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5014 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5015 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5016 if (rc != MDB_SUCCESS)
5021 MDB_GET_KEY(leaf, key);
5025 /** Set the cursor on a specific data item. */
5027 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5028 MDB_cursor_op op, int *exactp)
5032 MDB_node *leaf = NULL;
5037 assert(key->mv_size > 0);
5040 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5042 /* See if we're already on the right page */
5043 if (mc->mc_flags & C_INITIALIZED) {
5046 mp = mc->mc_pg[mc->mc_top];
5048 mc->mc_ki[mc->mc_top] = 0;
5049 return MDB_NOTFOUND;
5051 if (mp->mp_flags & P_LEAF2) {
5052 nodekey.mv_size = mc->mc_db->md_pad;
5053 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5055 leaf = NODEPTR(mp, 0);
5056 MDB_GET_KEY(leaf, &nodekey);
5058 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5060 /* Probably happens rarely, but first node on the page
5061 * was the one we wanted.
5063 mc->mc_ki[mc->mc_top] = 0;
5070 unsigned int nkeys = NUMKEYS(mp);
5072 if (mp->mp_flags & P_LEAF2) {
5073 nodekey.mv_data = LEAF2KEY(mp,
5074 nkeys-1, nodekey.mv_size);
5076 leaf = NODEPTR(mp, nkeys-1);
5077 MDB_GET_KEY(leaf, &nodekey);
5079 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5081 /* last node was the one we wanted */
5082 mc->mc_ki[mc->mc_top] = nkeys-1;
5088 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5089 /* This is definitely the right page, skip search_page */
5090 if (mp->mp_flags & P_LEAF2) {
5091 nodekey.mv_data = LEAF2KEY(mp,
5092 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5094 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5095 MDB_GET_KEY(leaf, &nodekey);
5097 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5099 /* current node was the one we wanted */
5109 /* If any parents have right-sibs, search.
5110 * Otherwise, there's nothing further.
5112 for (i=0; i<mc->mc_top; i++)
5114 NUMKEYS(mc->mc_pg[i])-1)
5116 if (i == mc->mc_top) {
5117 /* There are no other pages */
5118 mc->mc_ki[mc->mc_top] = nkeys;
5119 return MDB_NOTFOUND;
5123 /* There are no other pages */
5124 mc->mc_ki[mc->mc_top] = 0;
5125 return MDB_NOTFOUND;
5129 rc = mdb_page_search(mc, key, 0);
5130 if (rc != MDB_SUCCESS)
5133 mp = mc->mc_pg[mc->mc_top];
5134 assert(IS_LEAF(mp));
5137 leaf = mdb_node_search(mc, key, exactp);
5138 if (exactp != NULL && !*exactp) {
5139 /* MDB_SET specified and not an exact match. */
5140 return MDB_NOTFOUND;
5144 DPUTS("===> inexact leaf not found, goto sibling");
5145 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5146 return rc; /* no entries matched */
5147 mp = mc->mc_pg[mc->mc_top];
5148 assert(IS_LEAF(mp));
5149 leaf = NODEPTR(mp, 0);
5153 mc->mc_flags |= C_INITIALIZED;
5154 mc->mc_flags &= ~C_EOF;
5157 key->mv_size = mc->mc_db->md_pad;
5158 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5162 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5163 mdb_xcursor_init1(mc, leaf);
5166 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5167 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5168 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5171 if (op == MDB_GET_BOTH) {
5177 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5178 if (rc != MDB_SUCCESS)
5181 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5183 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5185 rc = mc->mc_dbx->md_dcmp(data, &d2);
5187 if (op == MDB_GET_BOTH || rc > 0)
5188 return MDB_NOTFOUND;
5193 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5194 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5199 /* The key already matches in all other cases */
5200 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5201 MDB_GET_KEY(leaf, key);
5202 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
5207 /** Move the cursor to the first item in the database. */
5209 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5215 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5217 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5218 rc = mdb_page_search(mc, NULL, 0);
5219 if (rc != MDB_SUCCESS)
5222 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5224 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5225 mc->mc_flags |= C_INITIALIZED;
5226 mc->mc_flags &= ~C_EOF;
5228 mc->mc_ki[mc->mc_top] = 0;
5230 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5231 key->mv_size = mc->mc_db->md_pad;
5232 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5237 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5238 mdb_xcursor_init1(mc, leaf);
5239 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5243 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5247 MDB_GET_KEY(leaf, key);
5251 /** Move the cursor to the last item in the database. */
5253 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5259 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5261 if (!(mc->mc_flags & C_EOF)) {
5263 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5266 lkey.mv_size = MDB_MAXKEYSIZE+1;
5267 lkey.mv_data = NULL;
5268 rc = mdb_page_search(mc, &lkey, 0);
5269 if (rc != MDB_SUCCESS)
5272 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5275 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5276 mc->mc_flags |= C_INITIALIZED|C_EOF;
5277 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5279 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5280 key->mv_size = mc->mc_db->md_pad;
5281 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5286 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5287 mdb_xcursor_init1(mc, leaf);
5288 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5292 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5297 MDB_GET_KEY(leaf, key);
5302 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5311 case MDB_GET_CURRENT:
5312 if (!(mc->mc_flags & C_INITIALIZED)) {
5315 MDB_page *mp = mc->mc_pg[mc->mc_top];
5317 mc->mc_ki[mc->mc_top] = 0;
5323 key->mv_size = mc->mc_db->md_pad;
5324 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5326 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5327 MDB_GET_KEY(leaf, key);
5329 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5330 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5332 rc = mdb_node_read(mc->mc_txn, leaf, data);
5339 case MDB_GET_BOTH_RANGE:
5340 if (data == NULL || mc->mc_xcursor == NULL) {
5348 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
5350 } else if (op == MDB_SET_RANGE)
5351 rc = mdb_cursor_set(mc, key, data, op, NULL);
5353 rc = mdb_cursor_set(mc, key, data, op, &exact);
5355 case MDB_GET_MULTIPLE:
5357 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
5358 !(mc->mc_flags & C_INITIALIZED)) {
5363 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5364 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5367 case MDB_NEXT_MULTIPLE:
5369 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5373 if (!(mc->mc_flags & C_INITIALIZED))
5374 rc = mdb_cursor_first(mc, key, data);
5376 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5377 if (rc == MDB_SUCCESS) {
5378 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5381 mx = &mc->mc_xcursor->mx_cursor;
5382 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5384 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5385 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5393 case MDB_NEXT_NODUP:
5394 if (!(mc->mc_flags & C_INITIALIZED))
5395 rc = mdb_cursor_first(mc, key, data);
5397 rc = mdb_cursor_next(mc, key, data, op);
5401 case MDB_PREV_NODUP:
5402 if (!(mc->mc_flags & C_INITIALIZED)) {
5403 rc = mdb_cursor_last(mc, key, data);
5406 mc->mc_flags |= C_INITIALIZED;
5407 mc->mc_ki[mc->mc_top]++;
5409 rc = mdb_cursor_prev(mc, key, data, op);
5412 rc = mdb_cursor_first(mc, key, data);
5416 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
5417 !(mc->mc_flags & C_INITIALIZED) ||
5418 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5422 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5425 rc = mdb_cursor_last(mc, key, data);
5429 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
5430 !(mc->mc_flags & C_INITIALIZED) ||
5431 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5435 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5438 DPRINTF("unhandled/unimplemented cursor operation %u", op);
5446 /** Touch all the pages in the cursor stack.
5447 * Makes sure all the pages are writable, before attempting a write operation.
5448 * @param[in] mc The cursor to operate on.
5451 mdb_cursor_touch(MDB_cursor *mc)
5455 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5458 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5459 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5462 *mc->mc_dbflag |= DB_DIRTY;
5464 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5465 rc = mdb_page_touch(mc);
5469 mc->mc_top = mc->mc_snum-1;
5473 /** Do not spill pages to disk if txn is getting full, may fail instead */
5474 #define MDB_NOSPILL 0x8000
5477 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5480 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5481 MDB_node *leaf = NULL;
5482 MDB_val xdata, *rdata, dkey;
5485 int do_sub = 0, insert = 0;
5486 unsigned int mcount = 0, dcount = 0, nospill;
5490 char dbuf[MDB_MAXKEYSIZE+1];
5491 unsigned int nflags;
5494 /* Check this first so counter will always be zero on any
5497 if (flags & MDB_MULTIPLE) {
5498 dcount = data[1].mv_size;
5499 data[1].mv_size = 0;
5500 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5504 nospill = flags & MDB_NOSPILL;
5505 flags &= ~MDB_NOSPILL;
5507 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5510 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5513 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5516 #if SIZE_MAX > MAXDATASIZE
5517 if (data->mv_size > MAXDATASIZE)
5521 DPRINTF("==> put db %u key [%s], size %"Z"u, data size %"Z"u",
5522 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
5526 if (flags == MDB_CURRENT) {
5527 if (!(mc->mc_flags & C_INITIALIZED))
5530 } else if (mc->mc_db->md_root == P_INVALID) {
5531 /* new database, cursor has nothing to point to */
5533 mc->mc_flags &= ~C_INITIALIZED;
5538 if (flags & MDB_APPEND) {
5540 rc = mdb_cursor_last(mc, &k2, &d2);
5542 rc = mc->mc_dbx->md_cmp(key, &k2);
5545 mc->mc_ki[mc->mc_top]++;
5547 /* new key is <= last key */
5552 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5554 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5555 DPRINTF("duplicate key [%s]", DKEY(key));
5557 return MDB_KEYEXIST;
5559 if (rc && rc != MDB_NOTFOUND)
5563 /* Cursor is positioned, check for room in the dirty list */
5565 if (flags & MDB_MULTIPLE) {
5567 xdata.mv_size = data->mv_size * dcount;
5571 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5575 if (rc == MDB_NO_ROOT) {
5577 /* new database, write a root leaf page */
5578 DPUTS("allocating new root leaf page");
5579 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5582 mdb_cursor_push(mc, np);
5583 mc->mc_db->md_root = np->mp_pgno;
5584 mc->mc_db->md_depth++;
5585 *mc->mc_dbflag |= DB_DIRTY;
5586 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5588 np->mp_flags |= P_LEAF2;
5589 mc->mc_flags |= C_INITIALIZED;
5591 /* make sure all cursor pages are writable */
5592 rc2 = mdb_cursor_touch(mc);
5597 /* The key already exists */
5598 if (rc == MDB_SUCCESS) {
5599 /* there's only a key anyway, so this is a no-op */
5600 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5601 unsigned int ksize = mc->mc_db->md_pad;
5602 if (key->mv_size != ksize)
5604 if (flags == MDB_CURRENT) {
5605 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5606 memcpy(ptr, key->mv_data, ksize);
5611 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5614 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5615 /* Was a single item before, must convert now */
5617 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5618 /* Just overwrite the current item */
5619 if (flags == MDB_CURRENT)
5622 dkey.mv_size = NODEDSZ(leaf);
5623 dkey.mv_data = NODEDATA(leaf);
5624 #if UINT_MAX < SIZE_MAX
5625 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5626 #ifdef MISALIGNED_OK
5627 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5629 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5632 /* if data matches, ignore it */
5633 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5634 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5636 /* create a fake page for the dup items */
5637 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5638 dkey.mv_data = dbuf;
5639 fp = (MDB_page *)&pbuf;
5640 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5641 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5642 fp->mp_lower = PAGEHDRSZ;
5643 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5644 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5645 fp->mp_flags |= P_LEAF2;
5646 fp->mp_pad = data->mv_size;
5647 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5649 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5650 (dkey.mv_size & 1) + (data->mv_size & 1);
5652 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5655 xdata.mv_size = fp->mp_upper;
5660 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5661 /* See if we need to convert from fake page to subDB */
5663 unsigned int offset;
5667 fp = NODEDATA(leaf);
5668 if (flags == MDB_CURRENT) {
5670 fp->mp_flags |= P_DIRTY;
5671 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5672 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5676 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5677 offset = fp->mp_pad;
5678 if (SIZELEFT(fp) >= offset)
5680 offset *= 4; /* space for 4 more */
5682 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5684 offset += offset & 1;
5685 fp_flags = fp->mp_flags;
5686 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5687 offset >= mc->mc_txn->mt_env->me_nodemax) {
5688 /* yes, convert it */
5690 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5691 dummy.md_pad = fp->mp_pad;
5692 dummy.md_flags = MDB_DUPFIXED;
5693 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5694 dummy.md_flags |= MDB_INTEGERKEY;
5697 dummy.md_branch_pages = 0;
5698 dummy.md_leaf_pages = 1;
5699 dummy.md_overflow_pages = 0;
5700 dummy.md_entries = NUMKEYS(fp);
5702 xdata.mv_size = sizeof(MDB_db);
5703 xdata.mv_data = &dummy;
5704 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5706 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5707 flags |= F_DUPDATA|F_SUBDATA;
5708 dummy.md_root = mp->mp_pgno;
5709 fp_flags &= ~P_SUBP;
5711 /* no, just grow it */
5713 xdata.mv_size = NODEDSZ(leaf) + offset;
5714 xdata.mv_data = &pbuf;
5715 mp = (MDB_page *)&pbuf;
5716 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5719 mp->mp_flags = fp_flags | P_DIRTY;
5720 mp->mp_pad = fp->mp_pad;
5721 mp->mp_lower = fp->mp_lower;
5722 mp->mp_upper = fp->mp_upper + offset;
5724 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5726 nsize = NODEDSZ(leaf) - fp->mp_upper;
5727 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5728 for (i=0; i<NUMKEYS(fp); i++)
5729 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5731 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5735 /* data is on sub-DB, just store it */
5736 flags |= F_DUPDATA|F_SUBDATA;
5740 /* overflow page overwrites need special handling */
5741 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5744 unsigned psize = mc->mc_txn->mt_env->me_psize;
5745 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5747 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5748 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5750 ovpages = omp->mp_pages;
5752 /* Is the ov page large enough? */
5753 if (ovpages >= dpages) {
5754 if (!(omp->mp_flags & P_DIRTY) &&
5755 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5757 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5760 level = 0; /* dirty in this txn or clean */
5763 if (omp->mp_flags & P_DIRTY) {
5764 /* yes, overwrite it. Note in this case we don't
5765 * bother to try shrinking the page if the new data
5766 * is smaller than the overflow threshold.
5769 /* It is writable only in a parent txn */
5770 size_t sz = (size_t) psize * ovpages, off;
5771 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5777 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5778 if (!(flags & MDB_RESERVE)) {
5779 /* Copy end of page, adjusting alignment so
5780 * compiler may copy words instead of bytes.
5782 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5783 memcpy((size_t *)((char *)np + off),
5784 (size_t *)((char *)omp + off), sz - off);
5787 memcpy(np, omp, sz); /* Copy beginning of page */
5790 SETDSZ(leaf, data->mv_size);
5791 if (F_ISSET(flags, MDB_RESERVE))
5792 data->mv_data = METADATA(omp);
5794 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5798 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5800 } else if (NODEDSZ(leaf) == data->mv_size) {
5801 /* same size, just replace it. Note that we could
5802 * also reuse this node if the new data is smaller,
5803 * but instead we opt to shrink the node in that case.
5805 if (F_ISSET(flags, MDB_RESERVE))
5806 data->mv_data = NODEDATA(leaf);
5807 else if (data->mv_size)
5808 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5810 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5813 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5814 mc->mc_db->md_entries--;
5816 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5823 nflags = flags & NODE_ADD_FLAGS;
5824 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5825 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5826 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5827 nflags &= ~MDB_APPEND;
5829 nflags |= MDB_SPLIT_REPLACE;
5830 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5832 /* There is room already in this leaf page. */
5833 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5834 if (rc == 0 && !do_sub && insert) {
5835 /* Adjust other cursors pointing to mp */
5836 MDB_cursor *m2, *m3;
5837 MDB_dbi dbi = mc->mc_dbi;
5838 unsigned i = mc->mc_top;
5839 MDB_page *mp = mc->mc_pg[i];
5841 if (mc->mc_flags & C_SUB)
5844 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5845 if (mc->mc_flags & C_SUB)
5846 m3 = &m2->mc_xcursor->mx_cursor;
5849 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5850 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5857 if (rc != MDB_SUCCESS)
5858 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5860 /* Now store the actual data in the child DB. Note that we're
5861 * storing the user data in the keys field, so there are strict
5862 * size limits on dupdata. The actual data fields of the child
5863 * DB are all zero size.
5870 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5871 if (flags & MDB_CURRENT) {
5872 xflags = MDB_CURRENT|MDB_NOSPILL;
5874 mdb_xcursor_init1(mc, leaf);
5875 xflags = (flags & MDB_NODUPDATA) ?
5876 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
5878 /* converted, write the original data first */
5880 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5884 /* Adjust other cursors pointing to mp */
5886 unsigned i = mc->mc_top;
5887 MDB_page *mp = mc->mc_pg[i];
5889 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5890 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5891 if (!(m2->mc_flags & C_INITIALIZED)) continue;
5892 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5893 mdb_xcursor_init1(m2, leaf);
5897 /* we've done our job */
5900 if (flags & MDB_APPENDDUP)
5901 xflags |= MDB_APPEND;
5902 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5903 if (flags & F_SUBDATA) {
5904 void *db = NODEDATA(leaf);
5905 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5908 /* sub-writes might have failed so check rc again.
5909 * Don't increment count if we just replaced an existing item.
5911 if (!rc && !(flags & MDB_CURRENT))
5912 mc->mc_db->md_entries++;
5913 if (flags & MDB_MULTIPLE) {
5916 if (mcount < dcount) {
5917 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5918 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5922 /* let caller know how many succeeded, if any */
5923 data[1].mv_size = mcount;
5927 /* If we succeeded and the key didn't exist before, make sure
5928 * the cursor is marked valid.
5931 mc->mc_flags |= C_INITIALIZED;
5936 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5941 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5944 if (!(mc->mc_flags & C_INITIALIZED))
5947 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
5949 flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
5951 rc = mdb_cursor_touch(mc);
5955 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5957 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5958 if (flags != MDB_NODUPDATA) {
5959 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5960 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5962 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
5963 /* If sub-DB still has entries, we're done */
5964 if (mc->mc_xcursor->mx_db.md_entries) {
5965 if (leaf->mn_flags & F_SUBDATA) {
5966 /* update subDB info */
5967 void *db = NODEDATA(leaf);
5968 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5971 /* shrink fake page */
5972 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5973 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5974 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5975 /* fix other sub-DB cursors pointed at this fake page */
5976 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5977 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5978 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5979 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5980 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5983 mc->mc_db->md_entries--;
5986 /* otherwise fall thru and delete the sub-DB */
5989 if (leaf->mn_flags & F_SUBDATA) {
5990 /* add all the child DB's pages to the free list */
5991 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5992 if (rc == MDB_SUCCESS) {
5993 mc->mc_db->md_entries -=
5994 mc->mc_xcursor->mx_db.md_entries;
5999 return mdb_cursor_del0(mc, leaf);
6002 /** Allocate and initialize new pages for a database.
6003 * @param[in] mc a cursor on the database being added to.
6004 * @param[in] flags flags defining what type of page is being allocated.
6005 * @param[in] num the number of pages to allocate. This is usually 1,
6006 * unless allocating overflow pages for a large record.
6007 * @param[out] mp Address of a page, or NULL on failure.
6008 * @return 0 on success, non-zero on failure.
6011 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6016 if ((rc = mdb_page_alloc(mc, num, &np)))
6018 DPRINTF("allocated new mpage %"Z"u, page size %u",
6019 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
6020 np->mp_flags = flags | P_DIRTY;
6021 np->mp_lower = PAGEHDRSZ;
6022 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6025 mc->mc_db->md_branch_pages++;
6026 else if (IS_LEAF(np))
6027 mc->mc_db->md_leaf_pages++;
6028 else if (IS_OVERFLOW(np)) {
6029 mc->mc_db->md_overflow_pages += num;
6037 /** Calculate the size of a leaf node.
6038 * The size depends on the environment's page size; if a data item
6039 * is too large it will be put onto an overflow page and the node
6040 * size will only include the key and not the data. Sizes are always
6041 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6042 * of the #MDB_node headers.
6043 * @param[in] env The environment handle.
6044 * @param[in] key The key for the node.
6045 * @param[in] data The data for the node.
6046 * @return The number of bytes needed to store the node.
6049 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6053 sz = LEAFSIZE(key, data);
6054 if (sz >= env->me_nodemax) {
6055 /* put on overflow page */
6056 sz -= data->mv_size - sizeof(pgno_t);
6060 return sz + sizeof(indx_t);
6063 /** Calculate the size of a branch node.
6064 * The size should depend on the environment's page size but since
6065 * we currently don't support spilling large keys onto overflow
6066 * pages, it's simply the size of the #MDB_node header plus the
6067 * size of the key. Sizes are always rounded up to an even number
6068 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6069 * @param[in] env The environment handle.
6070 * @param[in] key The key for the node.
6071 * @return The number of bytes needed to store the node.
6074 mdb_branch_size(MDB_env *env, MDB_val *key)
6079 if (sz >= env->me_nodemax) {
6080 /* put on overflow page */
6081 /* not implemented */
6082 /* sz -= key->size - sizeof(pgno_t); */
6085 return sz + sizeof(indx_t);
6088 /** Add a node to the page pointed to by the cursor.
6089 * @param[in] mc The cursor for this operation.
6090 * @param[in] indx The index on the page where the new node should be added.
6091 * @param[in] key The key for the new node.
6092 * @param[in] data The data for the new node, if any.
6093 * @param[in] pgno The page number, if adding a branch node.
6094 * @param[in] flags Flags for the node.
6095 * @return 0 on success, non-zero on failure. Possible errors are:
6097 * <li>ENOMEM - failed to allocate overflow pages for the node.
6098 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6099 * should never happen since all callers already calculate the
6100 * page's free space before calling this function.
6104 mdb_node_add(MDB_cursor *mc, indx_t indx,
6105 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6108 size_t node_size = NODESIZE;
6111 MDB_page *mp = mc->mc_pg[mc->mc_top];
6112 MDB_page *ofp = NULL; /* overflow page */
6115 assert(mp->mp_upper >= mp->mp_lower);
6117 DPRINTF("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6118 IS_LEAF(mp) ? "leaf" : "branch",
6119 IS_SUBP(mp) ? "sub-" : "",
6120 mp->mp_pgno, indx, data ? data->mv_size : 0,
6121 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
6124 /* Move higher keys up one slot. */
6125 int ksize = mc->mc_db->md_pad, dif;
6126 char *ptr = LEAF2KEY(mp, indx, ksize);
6127 dif = NUMKEYS(mp) - indx;
6129 memmove(ptr+ksize, ptr, dif*ksize);
6130 /* insert new key */
6131 memcpy(ptr, key->mv_data, ksize);
6133 /* Just using these for counting */
6134 mp->mp_lower += sizeof(indx_t);
6135 mp->mp_upper -= ksize - sizeof(indx_t);
6140 node_size += key->mv_size;
6144 if (F_ISSET(flags, F_BIGDATA)) {
6145 /* Data already on overflow page. */
6146 node_size += sizeof(pgno_t);
6147 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6148 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6150 /* Put data on overflow page. */
6151 DPRINTF("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6152 data->mv_size, node_size+data->mv_size);
6153 node_size += sizeof(pgno_t);
6154 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6156 DPRINTF("allocated overflow page %"Z"u", ofp->mp_pgno);
6159 node_size += data->mv_size;
6162 node_size += node_size & 1;
6164 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6165 DPRINTF("not enough room in page %"Z"u, got %u ptrs",
6166 mp->mp_pgno, NUMKEYS(mp));
6167 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6168 mp->mp_upper - mp->mp_lower);
6169 DPRINTF("node size = %"Z"u", node_size);
6170 return MDB_PAGE_FULL;
6173 /* Move higher pointers up one slot. */
6174 for (i = NUMKEYS(mp); i > indx; i--)
6175 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6177 /* Adjust free space offsets. */
6178 ofs = mp->mp_upper - node_size;
6179 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6180 mp->mp_ptrs[indx] = ofs;
6182 mp->mp_lower += sizeof(indx_t);
6184 /* Write the node data. */
6185 node = NODEPTR(mp, indx);
6186 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6187 node->mn_flags = flags;
6189 SETDSZ(node,data->mv_size);
6194 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6199 if (F_ISSET(flags, F_BIGDATA))
6200 memcpy(node->mn_data + key->mv_size, data->mv_data,
6202 else if (F_ISSET(flags, MDB_RESERVE))
6203 data->mv_data = node->mn_data + key->mv_size;
6205 memcpy(node->mn_data + key->mv_size, data->mv_data,
6208 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6210 if (F_ISSET(flags, MDB_RESERVE))
6211 data->mv_data = METADATA(ofp);
6213 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6220 /** Delete the specified node from a page.
6221 * @param[in] mp The page to operate on.
6222 * @param[in] indx The index of the node to delete.
6223 * @param[in] ksize The size of a node. Only used if the page is
6224 * part of a #MDB_DUPFIXED database.
6227 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6230 indx_t i, j, numkeys, ptr;
6237 COPY_PGNO(pgno, mp->mp_pgno);
6238 DPRINTF("delete node %u on %s page %"Z"u", indx,
6239 IS_LEAF(mp) ? "leaf" : "branch", pgno);
6242 assert(indx < NUMKEYS(mp));
6245 int x = NUMKEYS(mp) - 1 - indx;
6246 base = LEAF2KEY(mp, indx, ksize);
6248 memmove(base, base + ksize, x * ksize);
6249 mp->mp_lower -= sizeof(indx_t);
6250 mp->mp_upper += ksize - sizeof(indx_t);
6254 node = NODEPTR(mp, indx);
6255 sz = NODESIZE + node->mn_ksize;
6257 if (F_ISSET(node->mn_flags, F_BIGDATA))
6258 sz += sizeof(pgno_t);
6260 sz += NODEDSZ(node);
6264 ptr = mp->mp_ptrs[indx];
6265 numkeys = NUMKEYS(mp);
6266 for (i = j = 0; i < numkeys; i++) {
6268 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6269 if (mp->mp_ptrs[i] < ptr)
6270 mp->mp_ptrs[j] += sz;
6275 base = (char *)mp + mp->mp_upper;
6276 memmove(base + sz, base, ptr - mp->mp_upper);
6278 mp->mp_lower -= sizeof(indx_t);
6282 /** Compact the main page after deleting a node on a subpage.
6283 * @param[in] mp The main page to operate on.
6284 * @param[in] indx The index of the subpage on the main page.
6287 mdb_node_shrink(MDB_page *mp, indx_t indx)
6294 indx_t i, numkeys, ptr;
6296 node = NODEPTR(mp, indx);
6297 sp = (MDB_page *)NODEDATA(node);
6298 osize = NODEDSZ(node);
6300 delta = sp->mp_upper - sp->mp_lower;
6301 SETDSZ(node, osize - delta);
6302 xp = (MDB_page *)((char *)sp + delta);
6304 /* shift subpage upward */
6306 nsize = NUMKEYS(sp) * sp->mp_pad;
6307 memmove(METADATA(xp), METADATA(sp), nsize);
6310 nsize = osize - sp->mp_upper;
6311 numkeys = NUMKEYS(sp);
6312 for (i=numkeys-1; i>=0; i--)
6313 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6315 xp->mp_upper = sp->mp_lower;
6316 xp->mp_lower = sp->mp_lower;
6317 xp->mp_flags = sp->mp_flags;
6318 xp->mp_pad = sp->mp_pad;
6319 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6321 /* shift lower nodes upward */
6322 ptr = mp->mp_ptrs[indx];
6323 numkeys = NUMKEYS(mp);
6324 for (i = 0; i < numkeys; i++) {
6325 if (mp->mp_ptrs[i] <= ptr)
6326 mp->mp_ptrs[i] += delta;
6329 base = (char *)mp + mp->mp_upper;
6330 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6331 mp->mp_upper += delta;
6334 /** Initial setup of a sorted-dups cursor.
6335 * Sorted duplicates are implemented as a sub-database for the given key.
6336 * The duplicate data items are actually keys of the sub-database.
6337 * Operations on the duplicate data items are performed using a sub-cursor
6338 * initialized when the sub-database is first accessed. This function does
6339 * the preliminary setup of the sub-cursor, filling in the fields that
6340 * depend only on the parent DB.
6341 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6344 mdb_xcursor_init0(MDB_cursor *mc)
6346 MDB_xcursor *mx = mc->mc_xcursor;
6348 mx->mx_cursor.mc_xcursor = NULL;
6349 mx->mx_cursor.mc_txn = mc->mc_txn;
6350 mx->mx_cursor.mc_db = &mx->mx_db;
6351 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6352 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6353 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6354 mx->mx_cursor.mc_snum = 0;
6355 mx->mx_cursor.mc_top = 0;
6356 mx->mx_cursor.mc_flags = C_SUB;
6357 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6358 mx->mx_dbx.md_dcmp = NULL;
6359 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6362 /** Final setup of a sorted-dups cursor.
6363 * Sets up the fields that depend on the data from the main cursor.
6364 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6365 * @param[in] node The data containing the #MDB_db record for the
6366 * sorted-dup database.
6369 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6371 MDB_xcursor *mx = mc->mc_xcursor;
6373 if (node->mn_flags & F_SUBDATA) {
6374 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6375 mx->mx_cursor.mc_pg[0] = 0;
6376 mx->mx_cursor.mc_snum = 0;
6377 mx->mx_cursor.mc_flags = C_SUB;
6379 MDB_page *fp = NODEDATA(node);
6380 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6381 mx->mx_db.md_flags = 0;
6382 mx->mx_db.md_depth = 1;
6383 mx->mx_db.md_branch_pages = 0;
6384 mx->mx_db.md_leaf_pages = 1;
6385 mx->mx_db.md_overflow_pages = 0;
6386 mx->mx_db.md_entries = NUMKEYS(fp);
6387 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6388 mx->mx_cursor.mc_snum = 1;
6389 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6390 mx->mx_cursor.mc_top = 0;
6391 mx->mx_cursor.mc_pg[0] = fp;
6392 mx->mx_cursor.mc_ki[0] = 0;
6393 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6394 mx->mx_db.md_flags = MDB_DUPFIXED;
6395 mx->mx_db.md_pad = fp->mp_pad;
6396 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6397 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6400 DPRINTF("Sub-db %u for db %u root page %"Z"u", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6402 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6404 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6405 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6406 #if UINT_MAX < SIZE_MAX
6407 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6408 #ifdef MISALIGNED_OK
6409 mx->mx_dbx.md_cmp = mdb_cmp_long;
6411 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6416 /** Initialize a cursor for a given transaction and database. */
6418 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6421 mc->mc_backup = NULL;
6424 mc->mc_db = &txn->mt_dbs[dbi];
6425 mc->mc_dbx = &txn->mt_dbxs[dbi];
6426 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6431 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6433 mc->mc_xcursor = mx;
6434 mdb_xcursor_init0(mc);
6436 mc->mc_xcursor = NULL;
6438 if (*mc->mc_dbflag & DB_STALE) {
6439 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6444 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6447 size_t size = sizeof(MDB_cursor);
6449 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6452 /* Allow read access to the freelist */
6453 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6456 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6457 size += sizeof(MDB_xcursor);
6459 if ((mc = malloc(size)) != NULL) {
6460 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6461 if (txn->mt_cursors) {
6462 mc->mc_next = txn->mt_cursors[dbi];
6463 txn->mt_cursors[dbi] = mc;
6464 mc->mc_flags |= C_UNTRACK;
6476 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6478 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6481 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6484 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6488 /* Return the count of duplicate data items for the current key */
6490 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6494 if (mc == NULL || countp == NULL)
6497 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
6500 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6501 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6504 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6507 *countp = mc->mc_xcursor->mx_db.md_entries;
6513 mdb_cursor_close(MDB_cursor *mc)
6515 if (mc && !mc->mc_backup) {
6516 /* remove from txn, if tracked */
6517 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6518 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6519 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6521 *prev = mc->mc_next;
6528 mdb_cursor_txn(MDB_cursor *mc)
6530 if (!mc) return NULL;
6535 mdb_cursor_dbi(MDB_cursor *mc)
6541 /** Replace the key for a node with a new key.
6542 * @param[in] mc Cursor pointing to the node to operate on.
6543 * @param[in] key The new key to use.
6544 * @return 0 on success, non-zero on failure.
6547 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6554 indx_t ptr, i, numkeys, indx;
6557 indx = mc->mc_ki[mc->mc_top];
6558 mp = mc->mc_pg[mc->mc_top];
6559 node = NODEPTR(mp, indx);
6560 ptr = mp->mp_ptrs[indx];
6564 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6565 k2.mv_data = NODEKEY(node);
6566 k2.mv_size = node->mn_ksize;
6567 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6569 mdb_dkey(&k2, kbuf2),
6575 delta0 = delta = key->mv_size - node->mn_ksize;
6577 /* Must be 2-byte aligned. If new key is
6578 * shorter by 1, the shift will be skipped.
6580 delta += (delta & 1);
6582 if (delta > 0 && SIZELEFT(mp) < delta) {
6584 /* not enough space left, do a delete and split */
6585 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6586 pgno = NODEPGNO(node);
6587 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6588 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6591 numkeys = NUMKEYS(mp);
6592 for (i = 0; i < numkeys; i++) {
6593 if (mp->mp_ptrs[i] <= ptr)
6594 mp->mp_ptrs[i] -= delta;
6597 base = (char *)mp + mp->mp_upper;
6598 len = ptr - mp->mp_upper + NODESIZE;
6599 memmove(base - delta, base, len);
6600 mp->mp_upper -= delta;
6602 node = NODEPTR(mp, indx);
6605 /* But even if no shift was needed, update ksize */
6607 node->mn_ksize = key->mv_size;
6610 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6616 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6618 /** Move a node from csrc to cdst.
6621 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6628 unsigned short flags;
6632 /* Mark src and dst as dirty. */
6633 if ((rc = mdb_page_touch(csrc)) ||
6634 (rc = mdb_page_touch(cdst)))
6637 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6638 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6639 key.mv_size = csrc->mc_db->md_pad;
6640 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6642 data.mv_data = NULL;
6646 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6647 assert(!((long)srcnode&1));
6648 srcpg = NODEPGNO(srcnode);
6649 flags = srcnode->mn_flags;
6650 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6651 unsigned int snum = csrc->mc_snum;
6653 /* must find the lowest key below src */
6654 mdb_page_search_lowest(csrc);
6655 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6656 key.mv_size = csrc->mc_db->md_pad;
6657 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6659 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6660 key.mv_size = NODEKSZ(s2);
6661 key.mv_data = NODEKEY(s2);
6663 csrc->mc_snum = snum--;
6664 csrc->mc_top = snum;
6666 key.mv_size = NODEKSZ(srcnode);
6667 key.mv_data = NODEKEY(srcnode);
6669 data.mv_size = NODEDSZ(srcnode);
6670 data.mv_data = NODEDATA(srcnode);
6672 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6673 unsigned int snum = cdst->mc_snum;
6676 /* must find the lowest key below dst */
6677 mdb_page_search_lowest(cdst);
6678 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6679 bkey.mv_size = cdst->mc_db->md_pad;
6680 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6682 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6683 bkey.mv_size = NODEKSZ(s2);
6684 bkey.mv_data = NODEKEY(s2);
6686 cdst->mc_snum = snum--;
6687 cdst->mc_top = snum;
6688 mdb_cursor_copy(cdst, &mn);
6690 rc = mdb_update_key(&mn, &bkey);
6695 DPRINTF("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6696 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6697 csrc->mc_ki[csrc->mc_top],
6699 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6700 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6702 /* Add the node to the destination page.
6704 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6705 if (rc != MDB_SUCCESS)
6708 /* Delete the node from the source page.
6710 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6713 /* Adjust other cursors pointing to mp */
6714 MDB_cursor *m2, *m3;
6715 MDB_dbi dbi = csrc->mc_dbi;
6716 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6718 if (csrc->mc_flags & C_SUB)
6721 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6722 if (csrc->mc_flags & C_SUB)
6723 m3 = &m2->mc_xcursor->mx_cursor;
6726 if (m3 == csrc) continue;
6727 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6728 csrc->mc_ki[csrc->mc_top]) {
6729 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6730 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6735 /* Update the parent separators.
6737 if (csrc->mc_ki[csrc->mc_top] == 0) {
6738 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6739 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6740 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6742 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6743 key.mv_size = NODEKSZ(srcnode);
6744 key.mv_data = NODEKEY(srcnode);
6746 DPRINTF("update separator for source page %"Z"u to [%s]",
6747 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6748 mdb_cursor_copy(csrc, &mn);
6751 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6754 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6756 indx_t ix = csrc->mc_ki[csrc->mc_top];
6757 nullkey.mv_size = 0;
6758 csrc->mc_ki[csrc->mc_top] = 0;
6759 rc = mdb_update_key(csrc, &nullkey);
6760 csrc->mc_ki[csrc->mc_top] = ix;
6761 assert(rc == MDB_SUCCESS);
6765 if (cdst->mc_ki[cdst->mc_top] == 0) {
6766 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6767 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6768 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6770 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6771 key.mv_size = NODEKSZ(srcnode);
6772 key.mv_data = NODEKEY(srcnode);
6774 DPRINTF("update separator for destination page %"Z"u to [%s]",
6775 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6776 mdb_cursor_copy(cdst, &mn);
6779 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6782 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6784 indx_t ix = cdst->mc_ki[cdst->mc_top];
6785 nullkey.mv_size = 0;
6786 cdst->mc_ki[cdst->mc_top] = 0;
6787 rc = mdb_update_key(cdst, &nullkey);
6788 cdst->mc_ki[cdst->mc_top] = ix;
6789 assert(rc == MDB_SUCCESS);
6796 /** Merge one page into another.
6797 * The nodes from the page pointed to by \b csrc will
6798 * be copied to the page pointed to by \b cdst and then
6799 * the \b csrc page will be freed.
6800 * @param[in] csrc Cursor pointing to the source page.
6801 * @param[in] cdst Cursor pointing to the destination page.
6804 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6812 DPRINTF("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6813 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6815 assert(csrc->mc_snum > 1); /* can't merge root page */
6816 assert(cdst->mc_snum > 1);
6818 /* Mark dst as dirty. */
6819 if ((rc = mdb_page_touch(cdst)))
6822 /* Move all nodes from src to dst.
6824 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6825 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6826 key.mv_size = csrc->mc_db->md_pad;
6827 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6828 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6829 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6830 if (rc != MDB_SUCCESS)
6832 key.mv_data = (char *)key.mv_data + key.mv_size;
6835 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6836 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6837 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6838 unsigned int snum = csrc->mc_snum;
6840 /* must find the lowest key below src */
6841 mdb_page_search_lowest(csrc);
6842 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6843 key.mv_size = csrc->mc_db->md_pad;
6844 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6846 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6847 key.mv_size = NODEKSZ(s2);
6848 key.mv_data = NODEKEY(s2);
6850 csrc->mc_snum = snum--;
6851 csrc->mc_top = snum;
6853 key.mv_size = srcnode->mn_ksize;
6854 key.mv_data = NODEKEY(srcnode);
6857 data.mv_size = NODEDSZ(srcnode);
6858 data.mv_data = NODEDATA(srcnode);
6859 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6860 if (rc != MDB_SUCCESS)
6865 DPRINTF("dst page %"Z"u now has %u keys (%.1f%% filled)",
6866 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]), (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10);
6868 /* Unlink the src page from parent and add to free list.
6870 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6871 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6874 rc = mdb_update_key(csrc, &key);
6880 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6881 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6884 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6885 csrc->mc_db->md_leaf_pages--;
6887 csrc->mc_db->md_branch_pages--;
6889 /* Adjust other cursors pointing to mp */
6890 MDB_cursor *m2, *m3;
6891 MDB_dbi dbi = csrc->mc_dbi;
6892 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6894 if (csrc->mc_flags & C_SUB)
6897 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6898 if (csrc->mc_flags & C_SUB)
6899 m3 = &m2->mc_xcursor->mx_cursor;
6902 if (m3 == csrc) continue;
6903 if (m3->mc_snum < csrc->mc_snum) continue;
6904 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6905 m3->mc_pg[csrc->mc_top] = mp;
6906 m3->mc_ki[csrc->mc_top] += nkeys;
6910 mdb_cursor_pop(csrc);
6912 return mdb_rebalance(csrc);
6915 /** Copy the contents of a cursor.
6916 * @param[in] csrc The cursor to copy from.
6917 * @param[out] cdst The cursor to copy to.
6920 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6924 cdst->mc_txn = csrc->mc_txn;
6925 cdst->mc_dbi = csrc->mc_dbi;
6926 cdst->mc_db = csrc->mc_db;
6927 cdst->mc_dbx = csrc->mc_dbx;
6928 cdst->mc_snum = csrc->mc_snum;
6929 cdst->mc_top = csrc->mc_top;
6930 cdst->mc_flags = csrc->mc_flags;
6932 for (i=0; i<csrc->mc_snum; i++) {
6933 cdst->mc_pg[i] = csrc->mc_pg[i];
6934 cdst->mc_ki[i] = csrc->mc_ki[i];
6938 /** Rebalance the tree after a delete operation.
6939 * @param[in] mc Cursor pointing to the page where rebalancing
6941 * @return 0 on success, non-zero on failure.
6944 mdb_rebalance(MDB_cursor *mc)
6948 unsigned int ptop, minkeys;
6951 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6955 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6956 DPRINTF("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
6957 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6958 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6962 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6963 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6966 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6967 DPRINTF("no need to rebalance page %"Z"u, above fill threshold",
6973 if (mc->mc_snum < 2) {
6974 MDB_page *mp = mc->mc_pg[0];
6976 DPUTS("Can't rebalance a subpage, ignoring");
6979 if (NUMKEYS(mp) == 0) {
6980 DPUTS("tree is completely empty");
6981 mc->mc_db->md_root = P_INVALID;
6982 mc->mc_db->md_depth = 0;
6983 mc->mc_db->md_leaf_pages = 0;
6984 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6987 /* Adjust cursors pointing to mp */
6991 MDB_cursor *m2, *m3;
6992 MDB_dbi dbi = mc->mc_dbi;
6994 if (mc->mc_flags & C_SUB)
6997 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6998 if (mc->mc_flags & C_SUB)
6999 m3 = &m2->mc_xcursor->mx_cursor;
7002 if (m3->mc_snum < mc->mc_snum) continue;
7003 if (m3->mc_pg[0] == mp) {
7009 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7010 DPUTS("collapsing root page!");
7011 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7014 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7015 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7018 mc->mc_db->md_depth--;
7019 mc->mc_db->md_branch_pages--;
7020 mc->mc_ki[0] = mc->mc_ki[1];
7022 /* Adjust other cursors pointing to mp */
7023 MDB_cursor *m2, *m3;
7024 MDB_dbi dbi = mc->mc_dbi;
7026 if (mc->mc_flags & C_SUB)
7029 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7030 if (mc->mc_flags & C_SUB)
7031 m3 = &m2->mc_xcursor->mx_cursor;
7034 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7035 if (m3->mc_pg[0] == mp) {
7036 m3->mc_pg[0] = mc->mc_pg[0];
7039 m3->mc_ki[0] = m3->mc_ki[1];
7044 DPUTS("root page doesn't need rebalancing");
7048 /* The parent (branch page) must have at least 2 pointers,
7049 * otherwise the tree is invalid.
7051 ptop = mc->mc_top-1;
7052 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7054 /* Leaf page fill factor is below the threshold.
7055 * Try to move keys from left or right neighbor, or
7056 * merge with a neighbor page.
7061 mdb_cursor_copy(mc, &mn);
7062 mn.mc_xcursor = NULL;
7064 if (mc->mc_ki[ptop] == 0) {
7065 /* We're the leftmost leaf in our parent.
7067 DPUTS("reading right neighbor");
7069 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7070 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7073 mn.mc_ki[mn.mc_top] = 0;
7074 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7076 /* There is at least one neighbor to the left.
7078 DPUTS("reading left neighbor");
7080 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7081 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7084 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7085 mc->mc_ki[mc->mc_top] = 0;
7088 DPRINTF("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7089 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10);
7091 /* If the neighbor page is above threshold and has enough keys,
7092 * move one key from it. Otherwise we should try to merge them.
7093 * (A branch page must never have less than 2 keys.)
7095 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7096 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7097 return mdb_node_move(&mn, mc);
7099 if (mc->mc_ki[ptop] == 0)
7100 rc = mdb_page_merge(&mn, mc);
7102 rc = mdb_page_merge(mc, &mn);
7103 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7108 /** Complete a delete operation started by #mdb_cursor_del(). */
7110 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7116 mp = mc->mc_pg[mc->mc_top];
7117 ki = mc->mc_ki[mc->mc_top];
7119 /* add overflow pages to free list */
7120 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7124 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7125 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7126 (rc = mdb_ovpage_free(mc, omp)))
7129 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7130 mc->mc_db->md_entries--;
7131 rc = mdb_rebalance(mc);
7132 if (rc != MDB_SUCCESS)
7133 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7134 /* if mc points past last node in page, invalidate */
7135 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7136 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7139 /* Adjust other cursors pointing to mp */
7142 MDB_dbi dbi = mc->mc_dbi;
7144 mp = mc->mc_pg[mc->mc_top];
7145 nkeys = NUMKEYS(mp);
7146 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7149 if (!(m2->mc_flags & C_INITIALIZED))
7151 if (m2->mc_pg[mc->mc_top] == mp) {
7152 if (m2->mc_ki[mc->mc_top] > ki)
7153 m2->mc_ki[mc->mc_top]--;
7154 if (m2->mc_ki[mc->mc_top] >= nkeys)
7155 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
7164 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7165 MDB_val *key, MDB_val *data)
7170 MDB_val rdata, *xdata;
7174 assert(key != NULL);
7176 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
7178 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7181 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7185 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7189 mdb_cursor_init(&mc, txn, dbi, &mx);
7200 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7202 /* let mdb_page_split know about this cursor if needed:
7203 * delete will trigger a rebalance; if it needs to move
7204 * a node from one page to another, it will have to
7205 * update the parent's separator key(s). If the new sepkey
7206 * is larger than the current one, the parent page may
7207 * run out of space, triggering a split. We need this
7208 * cursor to be consistent until the end of the rebalance.
7210 mc.mc_flags |= C_UNTRACK;
7211 mc.mc_next = txn->mt_cursors[dbi];
7212 txn->mt_cursors[dbi] = &mc;
7213 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7214 txn->mt_cursors[dbi] = mc.mc_next;
7219 /** Split a page and insert a new node.
7220 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7221 * The cursor will be updated to point to the actual page and index where
7222 * the node got inserted after the split.
7223 * @param[in] newkey The key for the newly inserted node.
7224 * @param[in] newdata The data for the newly inserted node.
7225 * @param[in] newpgno The page number, if the new node is a branch node.
7226 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7227 * @return 0 on success, non-zero on failure.
7230 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7231 unsigned int nflags)
7234 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7237 unsigned int i, j, split_indx, nkeys, pmax;
7239 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7241 MDB_page *mp, *rp, *pp;
7246 mp = mc->mc_pg[mc->mc_top];
7247 newindx = mc->mc_ki[mc->mc_top];
7249 DPRINTF("-----> splitting %s page %"Z"u and adding [%s] at index %i",
7250 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7251 DKEY(newkey), mc->mc_ki[mc->mc_top]);
7253 /* Create a right sibling. */
7254 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7256 DPRINTF("new right sibling: page %"Z"u", rp->mp_pgno);
7258 if (mc->mc_snum < 2) {
7259 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7261 /* shift current top to make room for new parent */
7262 mc->mc_pg[1] = mc->mc_pg[0];
7263 mc->mc_ki[1] = mc->mc_ki[0];
7266 mc->mc_db->md_root = pp->mp_pgno;
7267 DPRINTF("root split! new root = %"Z"u", pp->mp_pgno);
7268 mc->mc_db->md_depth++;
7271 /* Add left (implicit) pointer. */
7272 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7273 /* undo the pre-push */
7274 mc->mc_pg[0] = mc->mc_pg[1];
7275 mc->mc_ki[0] = mc->mc_ki[1];
7276 mc->mc_db->md_root = mp->mp_pgno;
7277 mc->mc_db->md_depth--;
7284 ptop = mc->mc_top-1;
7285 DPRINTF("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno);
7288 mc->mc_flags |= C_SPLITTING;
7289 mdb_cursor_copy(mc, &mn);
7290 mn.mc_pg[mn.mc_top] = rp;
7291 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7293 if (nflags & MDB_APPEND) {
7294 mn.mc_ki[mn.mc_top] = 0;
7296 split_indx = newindx;
7301 nkeys = NUMKEYS(mp);
7302 split_indx = nkeys / 2;
7303 if (newindx < split_indx)
7309 unsigned int lsize, rsize, ksize;
7310 /* Move half of the keys to the right sibling */
7312 x = mc->mc_ki[mc->mc_top] - split_indx;
7313 ksize = mc->mc_db->md_pad;
7314 split = LEAF2KEY(mp, split_indx, ksize);
7315 rsize = (nkeys - split_indx) * ksize;
7316 lsize = (nkeys - split_indx) * sizeof(indx_t);
7317 mp->mp_lower -= lsize;
7318 rp->mp_lower += lsize;
7319 mp->mp_upper += rsize - lsize;
7320 rp->mp_upper -= rsize - lsize;
7321 sepkey.mv_size = ksize;
7322 if (newindx == split_indx) {
7323 sepkey.mv_data = newkey->mv_data;
7325 sepkey.mv_data = split;
7328 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7329 memcpy(rp->mp_ptrs, split, rsize);
7330 sepkey.mv_data = rp->mp_ptrs;
7331 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7332 memcpy(ins, newkey->mv_data, ksize);
7333 mp->mp_lower += sizeof(indx_t);
7334 mp->mp_upper -= ksize - sizeof(indx_t);
7337 memcpy(rp->mp_ptrs, split, x * ksize);
7338 ins = LEAF2KEY(rp, x, ksize);
7339 memcpy(ins, newkey->mv_data, ksize);
7340 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7341 rp->mp_lower += sizeof(indx_t);
7342 rp->mp_upper -= ksize - sizeof(indx_t);
7343 mc->mc_ki[mc->mc_top] = x;
7344 mc->mc_pg[mc->mc_top] = rp;
7349 /* For leaf pages, check the split point based on what
7350 * fits where, since otherwise mdb_node_add can fail.
7352 * This check is only needed when the data items are
7353 * relatively large, such that being off by one will
7354 * make the difference between success or failure.
7356 * It's also relevant if a page happens to be laid out
7357 * such that one half of its nodes are all "small" and
7358 * the other half of its nodes are "large." If the new
7359 * item is also "large" and falls on the half with
7360 * "large" nodes, it also may not fit.
7363 unsigned int psize, nsize;
7364 /* Maximum free space in an empty page */
7365 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7366 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7367 if ((nkeys < 20) || (nsize > pmax/16)) {
7368 if (newindx <= split_indx) {
7371 for (i=0; i<split_indx; i++) {
7372 node = NODEPTR(mp, i);
7373 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7374 if (F_ISSET(node->mn_flags, F_BIGDATA))
7375 psize += sizeof(pgno_t);
7377 psize += NODEDSZ(node);
7381 split_indx = newindx;
7392 for (i=nkeys-1; i>=split_indx; i--) {
7393 node = NODEPTR(mp, i);
7394 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7395 if (F_ISSET(node->mn_flags, F_BIGDATA))
7396 psize += sizeof(pgno_t);
7398 psize += NODEDSZ(node);
7402 split_indx = newindx;
7413 /* First find the separating key between the split pages.
7414 * The case where newindx == split_indx is ambiguous; the
7415 * new item could go to the new page or stay on the original
7416 * page. If newpos == 1 it goes to the new page.
7418 if (newindx == split_indx && newpos) {
7419 sepkey.mv_size = newkey->mv_size;
7420 sepkey.mv_data = newkey->mv_data;
7422 node = NODEPTR(mp, split_indx);
7423 sepkey.mv_size = node->mn_ksize;
7424 sepkey.mv_data = NODEKEY(node);
7428 DPRINTF("separator is [%s]", DKEY(&sepkey));
7430 /* Copy separator key to the parent.
7432 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7436 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7439 if (mn.mc_snum == mc->mc_snum) {
7440 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7441 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7442 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7443 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7448 /* Right page might now have changed parent.
7449 * Check if left page also changed parent.
7451 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7452 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7453 for (i=0; i<ptop; i++) {
7454 mc->mc_pg[i] = mn.mc_pg[i];
7455 mc->mc_ki[i] = mn.mc_ki[i];
7457 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7458 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7462 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7465 mc->mc_flags ^= C_SPLITTING;
7466 if (rc != MDB_SUCCESS) {
7469 if (nflags & MDB_APPEND) {
7470 mc->mc_pg[mc->mc_top] = rp;
7471 mc->mc_ki[mc->mc_top] = 0;
7472 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7475 for (i=0; i<mc->mc_top; i++)
7476 mc->mc_ki[i] = mn.mc_ki[i];
7483 /* Move half of the keys to the right sibling. */
7485 /* grab a page to hold a temporary copy */
7486 copy = mdb_page_malloc(mc->mc_txn, 1);
7490 copy->mp_pgno = mp->mp_pgno;
7491 copy->mp_flags = mp->mp_flags;
7492 copy->mp_lower = PAGEHDRSZ;
7493 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7494 mc->mc_pg[mc->mc_top] = copy;
7495 for (i = j = 0; i <= nkeys; j++) {
7496 if (i == split_indx) {
7497 /* Insert in right sibling. */
7498 /* Reset insert index for right sibling. */
7499 if (i != newindx || (newpos ^ ins_new)) {
7501 mc->mc_pg[mc->mc_top] = rp;
7505 if (i == newindx && !ins_new) {
7506 /* Insert the original entry that caused the split. */
7507 rkey.mv_data = newkey->mv_data;
7508 rkey.mv_size = newkey->mv_size;
7517 /* Update index for the new key. */
7518 mc->mc_ki[mc->mc_top] = j;
7519 } else if (i == nkeys) {
7522 node = NODEPTR(mp, i);
7523 rkey.mv_data = NODEKEY(node);
7524 rkey.mv_size = node->mn_ksize;
7526 xdata.mv_data = NODEDATA(node);
7527 xdata.mv_size = NODEDSZ(node);
7530 pgno = NODEPGNO(node);
7531 flags = node->mn_flags;
7536 if (!IS_LEAF(mp) && j == 0) {
7537 /* First branch index doesn't need key data. */
7541 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7545 nkeys = NUMKEYS(copy);
7546 for (i=0; i<nkeys; i++)
7547 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7548 mp->mp_lower = copy->mp_lower;
7549 mp->mp_upper = copy->mp_upper;
7550 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7551 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7553 /* reset back to original page */
7554 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7555 mc->mc_pg[mc->mc_top] = mp;
7556 if (nflags & MDB_RESERVE) {
7557 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7558 if (!(node->mn_flags & F_BIGDATA))
7559 newdata->mv_data = NODEDATA(node);
7563 /* Make sure mc_ki is still valid.
7565 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7566 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7567 for (i=0; i<ptop; i++) {
7568 mc->mc_pg[i] = mn.mc_pg[i];
7569 mc->mc_ki[i] = mn.mc_ki[i];
7571 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7572 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7576 /* return tmp page to freelist */
7577 mdb_page_free(mc->mc_txn->mt_env, copy);
7580 /* Adjust other cursors pointing to mp */
7581 MDB_cursor *m2, *m3;
7582 MDB_dbi dbi = mc->mc_dbi;
7583 int fixup = NUMKEYS(mp);
7585 if (mc->mc_flags & C_SUB)
7588 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7589 if (mc->mc_flags & C_SUB)
7590 m3 = &m2->mc_xcursor->mx_cursor;
7595 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7597 if (m3->mc_flags & C_SPLITTING)
7602 for (k=m3->mc_top; k>=0; k--) {
7603 m3->mc_ki[k+1] = m3->mc_ki[k];
7604 m3->mc_pg[k+1] = m3->mc_pg[k];
7606 if (m3->mc_ki[0] >= split_indx) {
7611 m3->mc_pg[0] = mc->mc_pg[0];
7615 if (m3->mc_pg[mc->mc_top] == mp) {
7616 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7617 m3->mc_ki[mc->mc_top]++;
7618 if (m3->mc_ki[mc->mc_top] >= fixup) {
7619 m3->mc_pg[mc->mc_top] = rp;
7620 m3->mc_ki[mc->mc_top] -= fixup;
7621 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7623 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7624 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7633 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7634 MDB_val *key, MDB_val *data, unsigned int flags)
7639 assert(key != NULL);
7640 assert(data != NULL);
7642 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7645 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7649 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7653 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7656 mdb_cursor_init(&mc, txn, dbi, &mx);
7657 return mdb_cursor_put(&mc, key, data, flags);
7661 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7663 if ((flag & CHANGEABLE) != flag)
7666 env->me_flags |= flag;
7668 env->me_flags &= ~flag;
7673 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7678 *arg = env->me_flags;
7683 mdb_env_get_path(MDB_env *env, const char **arg)
7688 *arg = env->me_path;
7692 /** Common code for #mdb_stat() and #mdb_env_stat().
7693 * @param[in] env the environment to operate in.
7694 * @param[in] db the #MDB_db record containing the stats to return.
7695 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7696 * @return 0, this function always succeeds.
7699 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7701 arg->ms_psize = env->me_psize;
7702 arg->ms_depth = db->md_depth;
7703 arg->ms_branch_pages = db->md_branch_pages;
7704 arg->ms_leaf_pages = db->md_leaf_pages;
7705 arg->ms_overflow_pages = db->md_overflow_pages;
7706 arg->ms_entries = db->md_entries;
7711 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7715 if (env == NULL || arg == NULL)
7718 toggle = mdb_env_pick_meta(env);
7720 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7724 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7728 if (env == NULL || arg == NULL)
7731 toggle = mdb_env_pick_meta(env);
7732 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7733 arg->me_mapsize = env->me_mapsize;
7734 arg->me_maxreaders = env->me_maxreaders;
7736 /* me_numreaders may be zero if this process never used any readers. Use
7737 * the shared numreader count if it exists.
7739 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7741 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7742 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7746 /** Set the default comparison functions for a database.
7747 * Called immediately after a database is opened to set the defaults.
7748 * The user can then override them with #mdb_set_compare() or
7749 * #mdb_set_dupsort().
7750 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7751 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7754 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7756 uint16_t f = txn->mt_dbs[dbi].md_flags;
7758 txn->mt_dbxs[dbi].md_cmp =
7759 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7760 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7762 txn->mt_dbxs[dbi].md_dcmp =
7763 !(f & MDB_DUPSORT) ? 0 :
7764 ((f & MDB_INTEGERDUP)
7765 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7766 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7769 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7774 int rc, dbflag, exact;
7775 unsigned int unused = 0;
7778 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7779 mdb_default_cmp(txn, FREE_DBI);
7782 if ((flags & VALID_FLAGS) != flags)
7788 if (flags & PERSISTENT_FLAGS) {
7789 uint16_t f2 = flags & PERSISTENT_FLAGS;
7790 /* make sure flag changes get committed */
7791 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7792 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7793 txn->mt_flags |= MDB_TXN_DIRTY;
7796 mdb_default_cmp(txn, MAIN_DBI);
7800 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7801 mdb_default_cmp(txn, MAIN_DBI);
7804 /* Is the DB already open? */
7806 for (i=2; i<txn->mt_numdbs; i++) {
7807 if (!txn->mt_dbxs[i].md_name.mv_size) {
7808 /* Remember this free slot */
7809 if (!unused) unused = i;
7812 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7813 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7819 /* If no free slot and max hit, fail */
7820 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7821 return MDB_DBS_FULL;
7823 /* Cannot mix named databases with some mainDB flags */
7824 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7825 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7827 /* Find the DB info */
7828 dbflag = DB_NEW|DB_VALID;
7831 key.mv_data = (void *)name;
7832 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7833 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7834 if (rc == MDB_SUCCESS) {
7835 /* make sure this is actually a DB */
7836 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7837 if (!(node->mn_flags & F_SUBDATA))
7839 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7840 /* Create if requested */
7842 data.mv_size = sizeof(MDB_db);
7843 data.mv_data = &dummy;
7844 memset(&dummy, 0, sizeof(dummy));
7845 dummy.md_root = P_INVALID;
7846 dummy.md_flags = flags & PERSISTENT_FLAGS;
7847 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7851 /* OK, got info, add to table */
7852 if (rc == MDB_SUCCESS) {
7853 unsigned int slot = unused ? unused : txn->mt_numdbs;
7854 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7855 txn->mt_dbxs[slot].md_name.mv_size = len;
7856 txn->mt_dbxs[slot].md_rel = NULL;
7857 txn->mt_dbflags[slot] = dbflag;
7858 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7860 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7861 mdb_default_cmp(txn, slot);
7870 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7872 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7875 if (txn->mt_dbflags[dbi] & DB_STALE) {
7878 /* Stale, must read the DB's root. cursor_init does it for us. */
7879 mdb_cursor_init(&mc, txn, dbi, &mx);
7881 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7884 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7887 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7889 ptr = env->me_dbxs[dbi].md_name.mv_data;
7890 env->me_dbxs[dbi].md_name.mv_data = NULL;
7891 env->me_dbxs[dbi].md_name.mv_size = 0;
7892 env->me_dbflags[dbi] = 0;
7896 int mdb_dbi_flags(MDB_env *env, MDB_dbi dbi, unsigned int *flags)
7898 /* We could return the flags for the FREE_DBI too but what's the point? */
7899 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
7901 *flags = env->me_dbflags[dbi];
7905 /** Add all the DB's pages to the free list.
7906 * @param[in] mc Cursor on the DB to free.
7907 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7908 * @return 0 on success, non-zero on failure.
7911 mdb_drop0(MDB_cursor *mc, int subs)
7915 rc = mdb_page_search(mc, NULL, 0);
7916 if (rc == MDB_SUCCESS) {
7917 MDB_txn *txn = mc->mc_txn;
7922 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7923 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7926 mdb_cursor_copy(mc, &mx);
7927 while (mc->mc_snum > 0) {
7928 MDB_page *mp = mc->mc_pg[mc->mc_top];
7929 unsigned n = NUMKEYS(mp);
7931 for (i=0; i<n; i++) {
7932 ni = NODEPTR(mp, i);
7933 if (ni->mn_flags & F_BIGDATA) {
7936 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7937 rc = mdb_page_get(txn, pg, &omp, NULL);
7940 assert(IS_OVERFLOW(omp));
7941 rc = mdb_midl_append_range(&txn->mt_free_pgs,
7945 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7946 mdb_xcursor_init1(mc, ni);
7947 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7953 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
7955 for (i=0; i<n; i++) {
7957 ni = NODEPTR(mp, i);
7960 mdb_midl_xappend(txn->mt_free_pgs, pg);
7965 mc->mc_ki[mc->mc_top] = i;
7966 rc = mdb_cursor_sibling(mc, 1);
7968 /* no more siblings, go back to beginning
7969 * of previous level.
7973 for (i=1; i<mc->mc_snum; i++) {
7975 mc->mc_pg[i] = mx.mc_pg[i];
7980 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
7981 } else if (rc == MDB_NOTFOUND) {
7987 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7989 MDB_cursor *mc, *m2;
7992 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7995 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7998 rc = mdb_cursor_open(txn, dbi, &mc);
8002 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8003 /* Invalidate the dropped DB's cursors */
8004 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8005 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8009 /* Can't delete the main DB */
8010 if (del && dbi > MAIN_DBI) {
8011 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8013 txn->mt_dbflags[dbi] = DB_STALE;
8014 mdb_dbi_close(txn->mt_env, dbi);
8017 /* reset the DB record, mark it dirty */
8018 txn->mt_dbflags[dbi] |= DB_DIRTY;
8019 txn->mt_dbs[dbi].md_depth = 0;
8020 txn->mt_dbs[dbi].md_branch_pages = 0;
8021 txn->mt_dbs[dbi].md_leaf_pages = 0;
8022 txn->mt_dbs[dbi].md_overflow_pages = 0;
8023 txn->mt_dbs[dbi].md_entries = 0;
8024 txn->mt_dbs[dbi].md_root = P_INVALID;
8026 txn->mt_flags |= MDB_TXN_DIRTY;
8029 mdb_cursor_close(mc);
8033 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8035 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8038 txn->mt_dbxs[dbi].md_cmp = cmp;
8042 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8044 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8047 txn->mt_dbxs[dbi].md_dcmp = cmp;
8051 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8053 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8056 txn->mt_dbxs[dbi].md_rel = rel;
8060 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8062 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8065 txn->mt_dbxs[dbi].md_relctx = ctx;
8069 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8071 unsigned int i, rdrs;
8078 if (!env->me_txns) {
8079 return func("(no reader locks)\n", ctx);
8081 rdrs = env->me_txns->mti_numreaders;
8082 mr = env->me_txns->mti_readers;
8083 for (i=0; i<rdrs; i++) {
8088 if (mr[i].mr_txnid == (txnid_t)-1) {
8089 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8091 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8095 func(" pid thread txnid\n", ctx);
8097 rc = func(buf, ctx);
8103 func("(no active readers)\n", ctx);
8108 /* insert pid into list if not already present.
8109 * return -1 if already present.
8111 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8113 /* binary search of pid in list */
8115 unsigned cursor = 1;
8117 unsigned n = ids[0];
8120 unsigned pivot = n >> 1;
8121 cursor = base + pivot + 1;
8122 val = pid - ids[cursor];
8127 } else if ( val > 0 ) {
8132 /* found, so it's a duplicate */
8141 for (n = ids[0]; n > cursor; n--)
8147 int mdb_reader_check(MDB_env *env, int *dead)
8149 unsigned int i, j, rdrs;
8160 rdrs = env->me_txns->mti_numreaders;
8161 pids = malloc((rdrs+1) * sizeof(pid_t));
8165 mr = env->me_txns->mti_readers;
8167 for (i=0; i<rdrs; i++) {
8168 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8170 if (mdb_pid_insert(pids, pid) == 0) {
8171 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8173 /* Recheck, a new process may have reused pid */
8174 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8175 for (j=i; j<rdrs; j++)
8176 if (mr[j].mr_pid == pid) {
8181 UNLOCK_MUTEX_R(env);