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 #ifdef 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: Operation and DB incompatible, or DB flags changed",
1129 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1130 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1131 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1135 mdb_strerror(int err)
1139 return ("Successful return: 0");
1141 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1142 i = err - MDB_KEYEXIST;
1143 return mdb_errstr[i];
1146 return strerror(err);
1150 /** Display a key in hexadecimal and return the address of the result.
1151 * @param[in] key the key to display
1152 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1153 * @return The key in hexadecimal form.
1156 mdb_dkey(MDB_val *key, char *buf)
1159 unsigned char *c = key->mv_data;
1165 if (key->mv_size > MDB_MAXKEYSIZE)
1166 return "MDB_MAXKEYSIZE";
1167 /* may want to make this a dynamic check: if the key is mostly
1168 * printable characters, print it as-is instead of converting to hex.
1172 for (i=0; i<key->mv_size; i++)
1173 ptr += sprintf(ptr, "%02x", *c++);
1175 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1180 /** Display all the keys in the page. */
1182 mdb_page_list(MDB_page *mp)
1185 unsigned int i, nkeys, nsize;
1189 nkeys = NUMKEYS(mp);
1190 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1191 for (i=0; i<nkeys; i++) {
1192 node = NODEPTR(mp, i);
1193 key.mv_size = node->mn_ksize;
1194 key.mv_data = node->mn_data;
1195 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1196 if (IS_BRANCH(mp)) {
1197 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1200 if (F_ISSET(node->mn_flags, F_BIGDATA))
1201 nsize += sizeof(pgno_t);
1203 nsize += NODEDSZ(node);
1204 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1210 mdb_cursor_chk(MDB_cursor *mc)
1216 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1217 for (i=0; i<mc->mc_top; i++) {
1219 node = NODEPTR(mp, mc->mc_ki[i]);
1220 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1223 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1229 /** Count all the pages in each DB and in the freelist
1230 * and make sure it matches the actual number of pages
1233 static void mdb_audit(MDB_txn *txn)
1237 MDB_ID freecount, count;
1242 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1243 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1244 freecount += *(MDB_ID *)data.mv_data;
1247 for (i = 0; i<txn->mt_numdbs; i++) {
1249 mdb_cursor_init(&mc, txn, i, &mx);
1250 if (txn->mt_dbs[i].md_root == P_INVALID)
1252 count += txn->mt_dbs[i].md_branch_pages +
1253 txn->mt_dbs[i].md_leaf_pages +
1254 txn->mt_dbs[i].md_overflow_pages;
1255 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1256 mdb_page_search(&mc, NULL, 0);
1260 mp = mc.mc_pg[mc.mc_top];
1261 for (j=0; j<NUMKEYS(mp); j++) {
1262 MDB_node *leaf = NODEPTR(mp, j);
1263 if (leaf->mn_flags & F_SUBDATA) {
1265 memcpy(&db, NODEDATA(leaf), sizeof(db));
1266 count += db.md_branch_pages + db.md_leaf_pages +
1267 db.md_overflow_pages;
1271 while (mdb_cursor_sibling(&mc, 1) == 0);
1274 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1275 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1276 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1282 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1284 return txn->mt_dbxs[dbi].md_cmp(a, b);
1288 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1290 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1293 /** Allocate memory for a page.
1294 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1297 mdb_page_malloc(MDB_txn *txn, unsigned num)
1299 MDB_env *env = txn->mt_env;
1300 MDB_page *ret = env->me_dpages;
1301 size_t sz = env->me_psize;
1304 VGMEMP_ALLOC(env, ret, sz);
1305 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1306 env->me_dpages = ret->mp_next;
1312 if ((ret = malloc(sz)) != NULL) {
1313 VGMEMP_ALLOC(env, ret, sz);
1318 /** Free a single page.
1319 * Saves single pages to a list, for future reuse.
1320 * (This is not used for multi-page overflow pages.)
1323 mdb_page_free(MDB_env *env, MDB_page *mp)
1325 mp->mp_next = env->me_dpages;
1326 VGMEMP_FREE(env, mp);
1327 env->me_dpages = mp;
1330 /* Free a dirty page */
1332 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1334 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1335 mdb_page_free(env, dp);
1337 /* large pages just get freed directly */
1338 VGMEMP_FREE(env, dp);
1343 /** Return all dirty pages to dpage list */
1345 mdb_dlist_free(MDB_txn *txn)
1347 MDB_env *env = txn->mt_env;
1348 MDB_ID2L dl = txn->mt_u.dirty_list;
1349 unsigned i, n = dl[0].mid;
1351 for (i = 1; i <= n; i++) {
1352 mdb_dpage_free(env, dl[i].mptr);
1357 /* Set or clear P_KEEP in non-overflow, non-sub pages in this txn's cursors.
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;
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 for (m3 = mc; m3->mc_flags & C_INITIALIZED; m3 = &mx->mx_cursor) {
1375 for (j=0; j<m3->mc_snum; j++)
1376 if ((m3->mc_pg[j]->mp_flags & (P_SUBP|P_DIRTY|P_KEEP))
1378 m3->mc_pg[j]->mp_flags ^= P_KEEP;
1379 mx = m3->mc_xcursor;
1389 static int mdb_page_flush(MDB_txn *txn);
1391 /** Spill pages from the dirty list back to disk.
1392 * This is intended to prevent running into #MDB_TXN_FULL situations,
1393 * but note that they may still occur in a few cases:
1394 * 1) pages in #MDB_DUPSORT sub-DBs are never spilled, so if there
1395 * are too many of these dirtied in one txn, the txn may still get
1397 * 2) child txns may run out of space if their parents dirtied a
1398 * lot of pages and never spilled them. TODO: we probably should do
1399 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1400 * the parent's dirty_room is below a given threshold.
1401 * 3) our estimate of the txn size could be too small. At the
1402 * moment this seems unlikely.
1404 * Otherwise, if not using nested txns, it is expected that apps will
1405 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1406 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1407 * If the txn never references them again, they can be left alone.
1408 * If the txn only reads them, they can be used without any fuss.
1409 * If the txn writes them again, they can be dirtied immediately without
1410 * going thru all of the work of #mdb_page_touch(). Such references are
1411 * handled by #mdb_page_unspill().
1413 * Also note, we never spill DB root pages, nor pages of active cursors,
1414 * because we'll need these back again soon anyway. And in nested txns,
1415 * we can't spill a page in a child txn if it was already spilled in a
1416 * parent txn. That would alter the parent txns' data even though
1417 * the child hasn't committed yet, and we'd have no way to undo it if
1418 * the child aborted.
1420 * @param[in] m0 cursor A cursor handle identifying the transaction and
1421 * database for which we are checking space.
1422 * @param[in] key For a put operation, the key being stored.
1423 * @param[in] data For a put operation, the data being stored.
1424 * @return 0 on success, non-zero on failure.
1427 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1429 MDB_txn *txn = m0->mc_txn;
1431 MDB_ID2L dl = txn->mt_u.dirty_list;
1435 if (m0->mc_flags & C_SUB)
1438 /* Estimate how much space this op will take */
1439 i = m0->mc_db->md_depth;
1440 /* Named DBs also dirty the main DB */
1441 if (m0->mc_dbi > MAIN_DBI)
1442 i += txn->mt_dbs[MAIN_DBI].md_depth;
1443 /* For puts, roughly factor in the key+data size */
1445 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1446 i += i; /* double it for good measure */
1448 if (txn->mt_dirty_room > i)
1451 if (!txn->mt_spill_pgs) {
1452 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1453 if (!txn->mt_spill_pgs)
1457 /* Mark all the dirty root pages we want to preserve */
1458 for (i=0; i<txn->mt_numdbs; i++) {
1459 if (txn->mt_dbflags[i] & DB_DIRTY) {
1460 pgno_t pgno = txn->mt_dbs[i].md_root;
1461 if (pgno == P_INVALID)
1463 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1465 if ((dp->mp_flags & P_DIRTY) && level <= 1)
1466 dp->mp_flags |= P_KEEP;
1470 /* Preserve pages used by cursors */
1471 mdb_cursorpages_mark(m0, P_DIRTY);
1473 /* Save the page IDs of all the pages we're flushing */
1474 for (i=1; i<=dl[0].mid; i++) {
1476 if (dp->mp_flags & P_KEEP)
1478 /* Can't spill twice, make sure it's not already in a parent's
1481 if (txn->mt_parent) {
1483 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1484 if (tx2->mt_spill_pgs) {
1485 j = mdb_midl_search(tx2->mt_spill_pgs, dl[i].mid);
1486 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == dl[i].mid) {
1487 dp->mp_flags |= P_KEEP;
1495 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, dl[i].mid)))
1498 mdb_midl_sort(txn->mt_spill_pgs);
1500 rc = mdb_page_flush(txn);
1502 mdb_cursorpages_mark(m0, P_DIRTY|P_KEEP);
1506 if (txn->mt_parent) {
1508 pgno_t pgno = dl[i].mid;
1509 txn->mt_dirty_room = txn->mt_parent->mt_dirty_room - dl[0].mid;
1510 /* dirty pages that are dirty in an ancestor don't
1511 * count against this txn's dirty_room.
1513 for (i=1; i<=dl[0].mid; i++) {
1514 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1515 j = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1516 if (j <= tx2->mt_u.dirty_list[0].mid &&
1517 tx2->mt_u.dirty_list[j].mid == pgno) {
1518 txn->mt_dirty_room++;
1524 txn->mt_dirty_room = MDB_IDL_UM_MAX - dl[0].mid;
1526 txn->mt_flags |= MDB_TXN_SPILLS;
1528 txn->mt_flags |= MDB_TXN_ERROR;
1533 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1535 mdb_find_oldest(MDB_txn *txn)
1538 txnid_t mr, oldest = txn->mt_txnid - 1;
1539 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1540 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1550 /** Add a page to the txn's dirty list */
1552 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1555 int (*insert)(MDB_ID2L, MDB_ID2 *);
1557 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1558 insert = mdb_mid2l_append;
1560 insert = mdb_mid2l_insert;
1562 mid.mid = mp->mp_pgno;
1564 insert(txn->mt_u.dirty_list, &mid);
1565 txn->mt_dirty_room--;
1568 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1569 * me_pghead and mt_next_pgno.
1571 * If there are free pages available from older transactions, they
1572 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1573 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1574 * and move me_pglast to say which records were consumed. Only this
1575 * function can create me_pghead and move me_pglast/mt_next_pgno.
1576 * @param[in] mc cursor A cursor handle identifying the transaction and
1577 * database for which we are allocating.
1578 * @param[in] num the number of pages to allocate.
1579 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1580 * will always be satisfied by a single contiguous chunk of memory.
1581 * @return 0 on success, non-zero on failure.
1584 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1586 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1587 /* Get at most <Max_retries> more freeDB records once me_pghead
1588 * has enough pages. If not enough, use new pages from the map.
1589 * If <Paranoid> and mc is updating the freeDB, only get new
1590 * records if me_pghead is empty. Then the freelist cannot play
1591 * catch-up with itself by growing while trying to save it.
1593 enum { Paranoid = 1, Max_retries = 500 };
1595 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1597 int rc, n2 = num-1, retry = Max_retries;
1598 MDB_txn *txn = mc->mc_txn;
1599 MDB_env *env = txn->mt_env;
1600 pgno_t pgno, *mop = env->me_pghead;
1601 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1603 txnid_t oldest = 0, last;
1609 /* If our dirty list is already full, we can't do anything */
1610 if (txn->mt_dirty_room == 0)
1611 return MDB_TXN_FULL;
1613 for (op = MDB_FIRST;; op = MDB_NEXT) {
1616 pgno_t *idl, old_id, new_id;
1618 /* Seek a big enough contiguous page range. Prefer
1619 * pages at the tail, just truncating the list.
1621 if (mop_len >= (unsigned)num) {
1625 if (mop[i-n2] == pgno+n2)
1627 } while (--i >= (unsigned)num);
1628 if (Max_retries < INT_MAX && --retry < 0)
1632 if (op == MDB_FIRST) { /* 1st iteration */
1633 /* Prepare to fetch more and coalesce */
1634 oldest = mdb_find_oldest(txn);
1635 last = env->me_pglast;
1636 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1639 key.mv_data = &last; /* will look up last+1 */
1640 key.mv_size = sizeof(last);
1642 if (Paranoid && mc->mc_dbi == FREE_DBI)
1645 if (Paranoid && retry < 0 && mop_len)
1649 /* Do not fetch more if the record will be too recent */
1652 rc = mdb_cursor_get(&m2, &key, NULL, op);
1654 if (rc == MDB_NOTFOUND)
1658 last = *(txnid_t*)key.mv_data;
1661 np = m2.mc_pg[m2.mc_top];
1662 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1663 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1666 idl = (MDB_ID *) data.mv_data;
1669 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1672 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1674 mop = env->me_pghead;
1676 env->me_pglast = last;
1678 DPRINTF("IDL read txn %"Z"u root %"Z"u num %u",
1679 last, txn->mt_dbs[FREE_DBI].md_root, i);
1681 DPRINTF("IDL %"Z"u", idl[k]);
1683 /* Merge in descending sorted order */
1686 mop[0] = (pgno_t)-1;
1690 for (; old_id < new_id; old_id = mop[--j])
1697 /* Use new pages from the map when nothing suitable in the freeDB */
1699 pgno = txn->mt_next_pgno;
1700 if (pgno + num >= env->me_maxpg) {
1701 DPUTS("DB size maxed out");
1702 return MDB_MAP_FULL;
1706 if (env->me_flags & MDB_WRITEMAP) {
1707 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1709 if (!(np = mdb_page_malloc(txn, num)))
1713 mop[0] = mop_len -= num;
1714 /* Move any stragglers down */
1715 for (j = i-num; j < mop_len; )
1716 mop[++j] = mop[++i];
1718 txn->mt_next_pgno = pgno + num;
1721 mdb_page_dirty(txn, np);
1727 /** Copy the used portions of a non-overflow page.
1728 * @param[in] dst page to copy into
1729 * @param[in] src page to copy from
1730 * @param[in] psize size of a page
1733 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1735 enum { Align = sizeof(pgno_t) };
1736 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1738 /* If page isn't full, just copy the used portion. Adjust
1739 * alignment so memcpy may copy words instead of bytes.
1741 if ((unused &= -Align) && !IS_LEAF2(src)) {
1743 memcpy(dst, src, (lower + (Align-1)) & -Align);
1744 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1747 memcpy(dst, src, psize - unused);
1751 /** Pull a page off the txn's spill list, if present.
1752 * If a page being referenced was spilled to disk in this txn, bring
1753 * it back and make it dirty/writable again.
1754 * @param[in] tx0 the transaction handle.
1755 * @param[in] mp the page being referenced.
1756 * @param[out] ret the writable page, if any. ret is unchanged if
1757 * mp wasn't spilled.
1760 mdb_page_unspill(MDB_txn *tx0, MDB_page *mp, MDB_page **ret)
1762 MDB_env *env = tx0->mt_env;
1765 pgno_t pgno = mp->mp_pgno;
1767 for (txn = tx0; txn; txn=txn->mt_parent) {
1768 if (!txn->mt_spill_pgs)
1770 x = mdb_midl_search(txn->mt_spill_pgs, pgno);
1771 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pgno) {
1774 if (IS_OVERFLOW(mp))
1778 if (env->me_flags & MDB_WRITEMAP) {
1781 np = mdb_page_malloc(txn, num);
1785 memcpy(np, mp, num * env->me_psize);
1787 mdb_page_copy(np, mp, env->me_psize);
1790 /* If in current txn, this page is no longer spilled */
1791 for (; x < txn->mt_spill_pgs[0]; x++)
1792 txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
1793 txn->mt_spill_pgs[0]--;
1794 } /* otherwise, if belonging to a parent txn, the
1795 * page remains spilled until child commits
1798 if (txn->mt_parent) {
1800 /* If this page is also in a parent's dirty list, then
1801 * it's already accounted in dirty_room, and we need to
1802 * cancel out the decrement that mdb_page_dirty does.
1804 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1805 x = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1806 if (x <= tx2->mt_u.dirty_list[0].mid &&
1807 tx2->mt_u.dirty_list[x].mid == pgno) {
1808 txn->mt_dirty_room++;
1813 mdb_page_dirty(tx0, np);
1814 np->mp_flags |= P_DIRTY;
1822 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1823 * @param[in] mc cursor pointing to the page to be touched
1824 * @return 0 on success, non-zero on failure.
1827 mdb_page_touch(MDB_cursor *mc)
1829 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1830 MDB_txn *txn = mc->mc_txn;
1831 MDB_cursor *m2, *m3;
1836 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1837 if (txn->mt_flags & MDB_TXN_SPILLS) {
1839 rc = mdb_page_unspill(txn, mp, &np);
1845 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1846 (rc = mdb_page_alloc(mc, 1, &np)))
1849 DPRINTF("touched db %u page %"Z"u -> %"Z"u", mc->mc_dbi,mp->mp_pgno,pgno);
1850 assert(mp->mp_pgno != pgno);
1851 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1852 /* Update the parent page, if any, to point to the new page */
1854 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1855 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1856 SETPGNO(node, pgno);
1858 mc->mc_db->md_root = pgno;
1860 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1861 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1863 /* If txn has a parent, make sure the page is in our
1867 unsigned x = mdb_mid2l_search(dl, pgno);
1868 if (x <= dl[0].mid && dl[x].mid == pgno) {
1869 if (mp != dl[x].mptr) { /* bad cursor? */
1870 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1871 return MDB_CORRUPTED;
1876 assert(dl[0].mid < MDB_IDL_UM_MAX);
1878 np = mdb_page_malloc(txn, 1);
1883 mdb_mid2l_insert(dl, &mid);
1888 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1890 np->mp_flags |= P_DIRTY;
1893 /* Adjust cursors pointing to mp */
1894 mc->mc_pg[mc->mc_top] = np;
1896 if (mc->mc_flags & C_SUB) {
1898 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1899 m3 = &m2->mc_xcursor->mx_cursor;
1900 if (m3->mc_snum < mc->mc_snum) continue;
1901 if (m3->mc_pg[mc->mc_top] == mp)
1902 m3->mc_pg[mc->mc_top] = np;
1905 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1906 if (m2->mc_snum < mc->mc_snum) continue;
1907 if (m2->mc_pg[mc->mc_top] == mp) {
1908 m2->mc_pg[mc->mc_top] = np;
1909 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1910 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1912 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1913 if (!(leaf->mn_flags & F_SUBDATA))
1914 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1923 mdb_env_sync(MDB_env *env, int force)
1926 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1927 if (env->me_flags & MDB_WRITEMAP) {
1928 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1929 ? MS_ASYNC : MS_SYNC;
1930 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1933 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1937 if (MDB_FDATASYNC(env->me_fd))
1944 /** Back up parent txn's cursors, then grab the originals for tracking */
1946 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1948 MDB_cursor *mc, *bk;
1953 for (i = src->mt_numdbs; --i >= 0; ) {
1954 if ((mc = src->mt_cursors[i]) != NULL) {
1955 size = sizeof(MDB_cursor);
1957 size += sizeof(MDB_xcursor);
1958 for (; mc; mc = bk->mc_next) {
1964 mc->mc_db = &dst->mt_dbs[i];
1965 /* Kill pointers into src - and dst to reduce abuse: The
1966 * user may not use mc until dst ends. Otherwise we'd...
1968 mc->mc_txn = NULL; /* ...set this to dst */
1969 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
1970 if ((mx = mc->mc_xcursor) != NULL) {
1971 *(MDB_xcursor *)(bk+1) = *mx;
1972 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
1974 mc->mc_next = dst->mt_cursors[i];
1975 dst->mt_cursors[i] = mc;
1982 /** Close this write txn's cursors, give parent txn's cursors back to parent.
1983 * @param[in] txn the transaction handle.
1984 * @param[in] merge true to keep changes to parent cursors, false to revert.
1985 * @return 0 on success, non-zero on failure.
1988 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1990 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
1994 for (i = txn->mt_numdbs; --i >= 0; ) {
1995 for (mc = cursors[i]; mc; mc = next) {
1997 if ((bk = mc->mc_backup) != NULL) {
1999 /* Commit changes to parent txn */
2000 mc->mc_next = bk->mc_next;
2001 mc->mc_backup = bk->mc_backup;
2002 mc->mc_txn = bk->mc_txn;
2003 mc->mc_db = bk->mc_db;
2004 mc->mc_dbflag = bk->mc_dbflag;
2005 if ((mx = mc->mc_xcursor) != NULL)
2006 mx->mx_cursor.mc_txn = bk->mc_txn;
2008 /* Abort nested txn */
2010 if ((mx = mc->mc_xcursor) != NULL)
2011 *mx = *(MDB_xcursor *)(bk+1);
2015 /* Only malloced cursors are permanently tracked. */
2022 #ifdef MDB_DEBUG_SKIP
2023 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2026 mdb_txn_reset0(MDB_txn *txn, const char *act);
2028 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2034 Pidset = F_SETLK, Pidcheck = F_GETLK
2038 /** Set or check a pid lock. Set returns 0 on success.
2039 * Check returns 0 if the process is certainly dead, nonzero if it may
2040 * be alive (the lock exists or an error happened so we do not know).
2042 * On Windows Pidset is a no-op, we merely check for the existence
2043 * of the process with the given pid. On POSIX we use a single byte
2044 * lock on the lockfile, set at an offset equal to the pid.
2047 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2049 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2052 if (op == Pidcheck) {
2053 h = OpenProcess(env->me_pidquery, FALSE, pid);
2054 /* No documented "no such process" code, but other program use this: */
2056 return ErrCode() != ERROR_INVALID_PARAMETER;
2057 /* A process exists until all handles to it close. Has it exited? */
2058 ret = WaitForSingleObject(h, 0) != 0;
2065 struct flock lock_info;
2066 memset(&lock_info, 0, sizeof(lock_info));
2067 lock_info.l_type = F_WRLCK;
2068 lock_info.l_whence = SEEK_SET;
2069 lock_info.l_start = pid;
2070 lock_info.l_len = 1;
2071 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2072 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2074 } else if ((rc = ErrCode()) == EINTR) {
2082 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2083 * @param[in] txn the transaction handle to initialize
2084 * @return 0 on success, non-zero on failure.
2087 mdb_txn_renew0(MDB_txn *txn)
2089 MDB_env *env = txn->mt_env;
2092 int rc, new_notls = 0;
2095 txn->mt_numdbs = env->me_numdbs;
2096 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2098 if (txn->mt_flags & MDB_TXN_RDONLY) {
2099 if (!env->me_txns) {
2100 i = mdb_env_pick_meta(env);
2101 txn->mt_txnid = env->me_metas[i]->mm_txnid;
2102 txn->mt_u.reader = NULL;
2104 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2105 pthread_getspecific(env->me_txkey);
2107 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2108 return MDB_BAD_RSLOT;
2110 pid_t pid = env->me_pid;
2111 pthread_t tid = pthread_self();
2113 if (!(env->me_flags & MDB_LIVE_READER)) {
2114 rc = mdb_reader_pid(env, Pidset, pid);
2116 UNLOCK_MUTEX_R(env);
2119 env->me_flags |= MDB_LIVE_READER;
2123 for (i=0; i<env->me_txns->mti_numreaders; i++)
2124 if (env->me_txns->mti_readers[i].mr_pid == 0)
2126 if (i == env->me_maxreaders) {
2127 UNLOCK_MUTEX_R(env);
2128 return MDB_READERS_FULL;
2130 env->me_txns->mti_readers[i].mr_pid = pid;
2131 env->me_txns->mti_readers[i].mr_tid = tid;
2132 if (i >= env->me_txns->mti_numreaders)
2133 env->me_txns->mti_numreaders = i+1;
2134 /* Save numreaders for un-mutexed mdb_env_close() */
2135 env->me_numreaders = env->me_txns->mti_numreaders;
2136 UNLOCK_MUTEX_R(env);
2137 r = &env->me_txns->mti_readers[i];
2138 new_notls = (env->me_flags & MDB_NOTLS);
2139 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2144 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2145 txn->mt_u.reader = r;
2147 txn->mt_toggle = txn->mt_txnid & 1;
2151 txn->mt_txnid = env->me_txns->mti_txnid;
2152 txn->mt_toggle = txn->mt_txnid & 1;
2155 if (txn->mt_txnid == mdb_debug_start)
2158 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2159 txn->mt_u.dirty_list = env->me_dirty_list;
2160 txn->mt_u.dirty_list[0].mid = 0;
2161 txn->mt_free_pgs = env->me_free_pgs;
2162 txn->mt_free_pgs[0] = 0;
2163 txn->mt_spill_pgs = NULL;
2167 /* Copy the DB info and flags */
2168 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
2170 /* Moved to here to avoid a data race in read TXNs */
2171 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
2173 for (i=2; i<txn->mt_numdbs; i++) {
2174 x = env->me_dbflags[i];
2175 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2176 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2178 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2180 if (env->me_maxpg < txn->mt_next_pgno) {
2181 mdb_txn_reset0(txn, "renew0-mapfail");
2183 txn->mt_u.reader->mr_pid = 0;
2184 txn->mt_u.reader = NULL;
2186 return MDB_MAP_RESIZED;
2193 mdb_txn_renew(MDB_txn *txn)
2197 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2200 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2201 DPUTS("environment had fatal error, must shutdown!");
2205 rc = mdb_txn_renew0(txn);
2206 if (rc == MDB_SUCCESS) {
2207 DPRINTF("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2208 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2209 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2215 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2219 int rc, size, tsize = sizeof(MDB_txn);
2221 if (env->me_flags & MDB_FATAL_ERROR) {
2222 DPUTS("environment had fatal error, must shutdown!");
2225 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2228 /* Nested transactions: Max 1 child, write txns only, no writemap */
2229 if (parent->mt_child ||
2230 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
2231 (env->me_flags & MDB_WRITEMAP))
2235 tsize = sizeof(MDB_ntxn);
2237 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2238 if (!(flags & MDB_RDONLY))
2239 size += env->me_maxdbs * sizeof(MDB_cursor *);
2241 if ((txn = calloc(1, size)) == NULL) {
2242 DPRINTF("calloc: %s", strerror(ErrCode()));
2245 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2246 if (flags & MDB_RDONLY) {
2247 txn->mt_flags |= MDB_TXN_RDONLY;
2248 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2250 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2251 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2257 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2258 if (!txn->mt_u.dirty_list ||
2259 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2261 free(txn->mt_u.dirty_list);
2265 txn->mt_txnid = parent->mt_txnid;
2266 txn->mt_toggle = parent->mt_toggle;
2267 txn->mt_dirty_room = parent->mt_dirty_room;
2268 txn->mt_u.dirty_list[0].mid = 0;
2269 txn->mt_spill_pgs = NULL;
2270 txn->mt_next_pgno = parent->mt_next_pgno;
2271 parent->mt_child = txn;
2272 txn->mt_parent = parent;
2273 txn->mt_numdbs = parent->mt_numdbs;
2274 txn->mt_flags = parent->mt_flags;
2275 txn->mt_dbxs = parent->mt_dbxs;
2276 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2277 /* Copy parent's mt_dbflags, but clear DB_NEW */
2278 for (i=0; i<txn->mt_numdbs; i++)
2279 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2281 ntxn = (MDB_ntxn *)txn;
2282 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2283 if (env->me_pghead) {
2284 size = MDB_IDL_SIZEOF(env->me_pghead);
2285 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2287 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2292 rc = mdb_cursor_shadow(parent, txn);
2294 mdb_txn_reset0(txn, "beginchild-fail");
2296 rc = mdb_txn_renew0(txn);
2302 DPRINTF("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2303 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2304 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
2311 mdb_txn_env(MDB_txn *txn)
2313 if(!txn) return NULL;
2317 /** Export or close DBI handles opened in this txn. */
2319 mdb_dbis_update(MDB_txn *txn, int keep)
2322 MDB_dbi n = txn->mt_numdbs;
2323 MDB_env *env = txn->mt_env;
2324 unsigned char *tdbflags = txn->mt_dbflags;
2326 for (i = n; --i >= 2;) {
2327 if (tdbflags[i] & DB_NEW) {
2329 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2331 char *ptr = env->me_dbxs[i].md_name.mv_data;
2332 env->me_dbxs[i].md_name.mv_data = NULL;
2333 env->me_dbxs[i].md_name.mv_size = 0;
2334 env->me_dbflags[i] = 0;
2339 if (keep && env->me_numdbs < n)
2343 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2344 * May be called twice for readonly txns: First reset it, then abort.
2345 * @param[in] txn the transaction handle to reset
2346 * @param[in] act why the transaction is being reset
2349 mdb_txn_reset0(MDB_txn *txn, const char *act)
2351 MDB_env *env = txn->mt_env;
2353 /* Close any DBI handles opened in this txn */
2354 mdb_dbis_update(txn, 0);
2356 DPRINTF("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2357 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2358 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2360 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2361 if (txn->mt_u.reader) {
2362 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2363 if (!(env->me_flags & MDB_NOTLS))
2364 txn->mt_u.reader = NULL; /* txn does not own reader */
2366 txn->mt_numdbs = 0; /* close nothing if called again */
2367 txn->mt_dbxs = NULL; /* mark txn as reset */
2369 mdb_cursors_close(txn, 0);
2371 if (!(env->me_flags & MDB_WRITEMAP)) {
2372 mdb_dlist_free(txn);
2374 mdb_midl_free(env->me_pghead);
2376 if (txn->mt_parent) {
2377 txn->mt_parent->mt_child = NULL;
2378 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2379 mdb_midl_free(txn->mt_free_pgs);
2380 mdb_midl_free(txn->mt_spill_pgs);
2381 free(txn->mt_u.dirty_list);
2385 if (mdb_midl_shrink(&txn->mt_free_pgs))
2386 env->me_free_pgs = txn->mt_free_pgs;
2387 env->me_pghead = NULL;
2391 /* The writer mutex was locked in mdb_txn_begin. */
2392 UNLOCK_MUTEX_W(env);
2397 mdb_txn_reset(MDB_txn *txn)
2402 /* This call is only valid for read-only txns */
2403 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2406 mdb_txn_reset0(txn, "reset");
2410 mdb_txn_abort(MDB_txn *txn)
2416 mdb_txn_abort(txn->mt_child);
2418 mdb_txn_reset0(txn, "abort");
2419 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2420 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2421 txn->mt_u.reader->mr_pid = 0;
2426 /** Save the freelist as of this transaction to the freeDB.
2427 * This changes the freelist. Keep trying until it stabilizes.
2430 mdb_freelist_save(MDB_txn *txn)
2432 /* env->me_pghead[] can grow and shrink during this call.
2433 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2434 * Page numbers cannot disappear from txn->mt_free_pgs[].
2437 MDB_env *env = txn->mt_env;
2438 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2439 txnid_t pglast = 0, head_id = 0;
2440 pgno_t freecnt = 0, *free_pgs, *mop;
2441 ssize_t head_room = 0, total_room = 0, mop_len;
2443 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2445 if (env->me_pghead) {
2446 /* Make sure first page of freeDB is touched and on freelist */
2447 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2448 if (rc && rc != MDB_NOTFOUND)
2453 /* Come back here after each Put() in case freelist changed */
2456 /* If using records from freeDB which we have not yet
2457 * deleted, delete them and any we reserved for me_pghead.
2459 while (pglast < env->me_pglast) {
2460 rc = mdb_cursor_first(&mc, &key, NULL);
2463 pglast = head_id = *(txnid_t *)key.mv_data;
2464 total_room = head_room = 0;
2465 assert(pglast <= env->me_pglast);
2466 rc = mdb_cursor_del(&mc, 0);
2471 /* Save the IDL of pages freed by this txn, to a single record */
2472 if (freecnt < txn->mt_free_pgs[0]) {
2474 /* Make sure last page of freeDB is touched and on freelist */
2475 key.mv_size = MDB_MAXKEYSIZE+1;
2477 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2478 if (rc && rc != MDB_NOTFOUND)
2481 free_pgs = txn->mt_free_pgs;
2482 /* Write to last page of freeDB */
2483 key.mv_size = sizeof(txn->mt_txnid);
2484 key.mv_data = &txn->mt_txnid;
2486 freecnt = free_pgs[0];
2487 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2488 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2491 /* Retry if mt_free_pgs[] grew during the Put() */
2492 free_pgs = txn->mt_free_pgs;
2493 } while (freecnt < free_pgs[0]);
2494 mdb_midl_sort(free_pgs);
2495 memcpy(data.mv_data, free_pgs, data.mv_size);
2498 unsigned int i = free_pgs[0];
2499 DPRINTF("IDL write txn %"Z"u root %"Z"u num %u",
2500 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2502 DPRINTF("IDL %"Z"u", free_pgs[i]);
2508 mop = env->me_pghead;
2509 mop_len = mop ? mop[0] : 0;
2511 /* Reserve records for me_pghead[]. Split it if multi-page,
2512 * to avoid searching freeDB for a page range. Use keys in
2513 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2515 if (total_room >= mop_len) {
2516 if (total_room == mop_len || --more < 0)
2518 } else if (head_room >= maxfree_1pg && head_id > 1) {
2519 /* Keep current record (overflow page), add a new one */
2523 /* (Re)write {key = head_id, IDL length = head_room} */
2524 total_room -= head_room;
2525 head_room = mop_len - total_room;
2526 if (head_room > maxfree_1pg && head_id > 1) {
2527 /* Overflow multi-page for part of me_pghead */
2528 head_room /= head_id; /* amortize page sizes */
2529 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2530 } else if (head_room < 0) {
2531 /* Rare case, not bothering to delete this record */
2534 key.mv_size = sizeof(head_id);
2535 key.mv_data = &head_id;
2536 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2537 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2540 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2541 total_room += head_room;
2544 /* Fill in the reserved, touched me_pghead records */
2550 rc = mdb_cursor_first(&mc, &key, &data);
2551 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2552 unsigned flags = MDB_CURRENT;
2553 txnid_t id = *(txnid_t *)key.mv_data;
2554 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2557 assert(len >= 0 && id <= env->me_pglast);
2559 if (len > mop_len) {
2561 data.mv_size = (len + 1) * sizeof(MDB_ID);
2564 data.mv_data = mop -= len;
2567 rc = mdb_cursor_put(&mc, &key, &data, flags);
2569 if (rc || !(mop_len -= len))
2576 /** Flush dirty pages to the map, after clearing their dirty flag.
2579 mdb_page_flush(MDB_txn *txn)
2581 MDB_env *env = txn->mt_env;
2582 MDB_ID2L dl = txn->mt_u.dirty_list;
2583 unsigned psize = env->me_psize, j;
2584 int i, pagecount = dl[0].mid, rc;
2585 size_t size = 0, pos = 0;
2587 MDB_page *dp = NULL;
2591 struct iovec iov[MDB_COMMIT_PAGES];
2592 ssize_t wpos = 0, wsize = 0, wres;
2593 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2598 if (env->me_flags & MDB_WRITEMAP) {
2599 /* Clear dirty flags */
2600 for (i=1; i<=pagecount; i++) {
2602 /* Don't flush this page yet */
2603 if (dp->mp_flags & P_KEEP) {
2604 dp->mp_flags ^= P_KEEP;
2608 dp->mp_flags &= ~P_DIRTY;
2614 /* Write the pages */
2616 if (i <= pagecount) {
2618 /* Don't flush this page yet */
2619 if (dp->mp_flags & P_KEEP) {
2620 dp->mp_flags ^= P_KEEP;
2625 /* clear dirty flag */
2626 dp->mp_flags &= ~P_DIRTY;
2629 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2634 /* Windows actually supports scatter/gather I/O, but only on
2635 * unbuffered file handles. Since we're relying on the OS page
2636 * cache for all our data, that's self-defeating. So we just
2637 * write pages one at a time. We use the ov structure to set
2638 * the write offset, to at least save the overhead of a Seek
2641 DPRINTF("committing page %"Z"u", pgno);
2642 memset(&ov, 0, sizeof(ov));
2643 ov.Offset = pos & 0xffffffff;
2644 ov.OffsetHigh = pos >> 16 >> 16;
2645 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2647 DPRINTF("WriteFile: %d", rc);
2651 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2652 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2654 /* Write previous page(s) */
2655 #ifdef MDB_USE_PWRITEV
2656 wres = pwritev(env->me_fd, iov, n, wpos);
2659 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2661 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2663 DPRINTF("lseek: %s", strerror(rc));
2666 wres = writev(env->me_fd, iov, n);
2669 if (wres != wsize) {
2672 DPRINTF("Write error: %s", strerror(rc));
2674 rc = EIO; /* TODO: Use which error code? */
2675 DPUTS("short write, filesystem full?");
2686 DPRINTF("committing page %"Z"u", pgno);
2687 next_pos = pos + size;
2688 iov[n].iov_len = size;
2689 iov[n].iov_base = (char *)dp;
2696 for (i=1; i<=pagecount; i++) {
2698 /* This is a page we skipped above */
2701 dl[j].mid = dp->mp_pgno;
2704 mdb_dpage_free(env, dp);
2712 mdb_txn_commit(MDB_txn *txn)
2718 assert(txn != NULL);
2719 assert(txn->mt_env != NULL);
2721 if (txn->mt_child) {
2722 rc = mdb_txn_commit(txn->mt_child);
2723 txn->mt_child = NULL;
2730 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2731 mdb_dbis_update(txn, 1);
2732 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2737 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2738 DPUTS("error flag is set, can't commit");
2740 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2745 if (txn->mt_parent) {
2746 MDB_txn *parent = txn->mt_parent;
2750 /* Append our free list to parent's */
2751 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2754 mdb_midl_free(txn->mt_free_pgs);
2756 parent->mt_next_pgno = txn->mt_next_pgno;
2757 parent->mt_flags = txn->mt_flags;
2759 /* Merge our cursors into parent's and close them */
2760 mdb_cursors_close(txn, 1);
2762 /* Update parent's DB table. */
2763 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2764 parent->mt_numdbs = txn->mt_numdbs;
2765 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2766 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2767 for (i=2; i<txn->mt_numdbs; i++) {
2768 /* preserve parent's DB_NEW status */
2769 x = parent->mt_dbflags[i] & DB_NEW;
2770 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2773 dst = parent->mt_u.dirty_list;
2774 src = txn->mt_u.dirty_list;
2775 /* Remove anything in our dirty list from parent's spill list */
2776 if (parent->mt_spill_pgs) {
2777 x = parent->mt_spill_pgs[0];
2779 /* zero out our dirty pages in parent spill list */
2780 for (i=1; i<=src[0].mid; i++) {
2781 if (src[i].mid < parent->mt_spill_pgs[x])
2783 if (src[i].mid > parent->mt_spill_pgs[x]) {
2789 parent->mt_spill_pgs[x] = 0;
2792 /* OK, we had a few hits, squash zeros from the spill list */
2793 if (len < parent->mt_spill_pgs[0]) {
2795 for (y=1; y<=parent->mt_spill_pgs[0]; y++) {
2796 if (parent->mt_spill_pgs[y]) {
2798 parent->mt_spill_pgs[x] = parent->mt_spill_pgs[y];
2803 parent->mt_spill_pgs[0] = len;
2806 /* Find len = length of merging our dirty list with parent's */
2808 dst[0].mid = 0; /* simplify loops */
2809 if (parent->mt_parent) {
2810 len = x + src[0].mid;
2811 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2812 for (i = x; y && i; y--) {
2813 pgno_t yp = src[y].mid;
2814 while (yp < dst[i].mid)
2816 if (yp == dst[i].mid) {
2821 } else { /* Simplify the above for single-ancestor case */
2822 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2824 /* Merge our dirty list with parent's */
2826 for (i = len; y; dst[i--] = src[y--]) {
2827 pgno_t yp = src[y].mid;
2828 while (yp < dst[x].mid)
2829 dst[i--] = dst[x--];
2830 if (yp == dst[x].mid)
2831 free(dst[x--].mptr);
2835 free(txn->mt_u.dirty_list);
2836 parent->mt_dirty_room = txn->mt_dirty_room;
2837 if (txn->mt_spill_pgs) {
2838 if (parent->mt_spill_pgs) {
2839 mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2840 mdb_midl_free(txn->mt_spill_pgs);
2841 mdb_midl_sort(parent->mt_spill_pgs);
2843 parent->mt_spill_pgs = txn->mt_spill_pgs;
2847 parent->mt_child = NULL;
2848 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2853 if (txn != env->me_txn) {
2854 DPUTS("attempt to commit unknown transaction");
2859 mdb_cursors_close(txn, 0);
2861 if (!txn->mt_u.dirty_list[0].mid &&
2862 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2865 DPRINTF("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2866 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2868 /* Update DB root pointers */
2869 if (txn->mt_numdbs > 2) {
2873 data.mv_size = sizeof(MDB_db);
2875 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2876 for (i = 2; i < txn->mt_numdbs; i++) {
2877 if (txn->mt_dbflags[i] & DB_DIRTY) {
2878 data.mv_data = &txn->mt_dbs[i];
2879 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2886 rc = mdb_freelist_save(txn);
2890 mdb_midl_free(env->me_pghead);
2891 env->me_pghead = NULL;
2892 if (mdb_midl_shrink(&txn->mt_free_pgs))
2893 env->me_free_pgs = txn->mt_free_pgs;
2899 if ((rc = mdb_page_flush(txn)) ||
2900 (rc = mdb_env_sync(env, 0)) ||
2901 (rc = mdb_env_write_meta(txn)))
2907 mdb_dbis_update(txn, 1);
2909 UNLOCK_MUTEX_W(env);
2919 /** Read the environment parameters of a DB environment before
2920 * mapping it into memory.
2921 * @param[in] env the environment handle
2922 * @param[out] meta address of where to store the meta information
2923 * @return 0 on success, non-zero on failure.
2926 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2933 /* We don't know the page size yet, so use a minimum value.
2934 * Read both meta pages so we can use the latest one.
2937 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2941 memset(&ov, 0, sizeof(ov));
2943 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2944 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2947 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2949 if (rc != MDB_PAGESIZE) {
2950 if (rc == 0 && off == 0)
2952 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2953 DPRINTF("read: %s", mdb_strerror(rc));
2957 p = (MDB_page *)&pbuf;
2959 if (!F_ISSET(p->mp_flags, P_META)) {
2960 DPRINTF("page %"Z"u not a meta page", p->mp_pgno);
2965 if (m->mm_magic != MDB_MAGIC) {
2966 DPUTS("meta has invalid magic");
2970 if (m->mm_version != MDB_DATA_VERSION) {
2971 DPRINTF("database is version %u, expected version %u",
2972 m->mm_version, MDB_DATA_VERSION);
2973 return MDB_VERSION_MISMATCH;
2976 if (off == 0 || m->mm_txnid > meta->mm_txnid)
2982 /** Write the environment parameters of a freshly created DB environment.
2983 * @param[in] env the environment handle
2984 * @param[out] meta address of where to store the meta information
2985 * @return 0 on success, non-zero on failure.
2988 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2996 memset(&ov, 0, sizeof(ov));
2997 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
2999 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3002 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3003 len = pwrite(fd, ptr, size, pos); \
3004 rc = (len >= 0); } while(0)
3007 DPUTS("writing new meta page");
3009 GET_PAGESIZE(psize);
3011 meta->mm_magic = MDB_MAGIC;
3012 meta->mm_version = MDB_DATA_VERSION;
3013 meta->mm_mapsize = env->me_mapsize;
3014 meta->mm_psize = psize;
3015 meta->mm_last_pg = 1;
3016 meta->mm_flags = env->me_flags & 0xffff;
3017 meta->mm_flags |= MDB_INTEGERKEY;
3018 meta->mm_dbs[0].md_root = P_INVALID;
3019 meta->mm_dbs[1].md_root = P_INVALID;
3021 p = calloc(2, psize);
3023 p->mp_flags = P_META;
3024 *(MDB_meta *)METADATA(p) = *meta;
3026 q = (MDB_page *)((char *)p + psize);
3028 q->mp_flags = P_META;
3029 *(MDB_meta *)METADATA(q) = *meta;
3031 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3034 else if ((unsigned) len == psize * 2)
3042 /** Update the environment info to commit a transaction.
3043 * @param[in] txn the transaction that's being committed
3044 * @return 0 on success, non-zero on failure.
3047 mdb_env_write_meta(MDB_txn *txn)
3050 MDB_meta meta, metab, *mp;
3052 int rc, len, toggle;
3061 assert(txn != NULL);
3062 assert(txn->mt_env != NULL);
3064 toggle = !txn->mt_toggle;
3065 DPRINTF("writing meta page %d for root page %"Z"u",
3066 toggle, txn->mt_dbs[MAIN_DBI].md_root);
3069 mp = env->me_metas[toggle];
3071 if (env->me_flags & MDB_WRITEMAP) {
3072 /* Persist any increases of mapsize config */
3073 if (env->me_mapsize > mp->mm_mapsize)
3074 mp->mm_mapsize = env->me_mapsize;
3075 mp->mm_dbs[0] = txn->mt_dbs[0];
3076 mp->mm_dbs[1] = txn->mt_dbs[1];
3077 mp->mm_last_pg = txn->mt_next_pgno - 1;
3078 mp->mm_txnid = txn->mt_txnid;
3079 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3080 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3083 ptr += env->me_psize;
3084 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3091 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3092 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3094 ptr = (char *)&meta;
3095 if (env->me_mapsize > mp->mm_mapsize) {
3096 /* Persist any increases of mapsize config */
3097 meta.mm_mapsize = env->me_mapsize;
3098 off = offsetof(MDB_meta, mm_mapsize);
3100 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3102 len = sizeof(MDB_meta) - off;
3105 meta.mm_dbs[0] = txn->mt_dbs[0];
3106 meta.mm_dbs[1] = txn->mt_dbs[1];
3107 meta.mm_last_pg = txn->mt_next_pgno - 1;
3108 meta.mm_txnid = txn->mt_txnid;
3111 off += env->me_psize;
3114 /* Write to the SYNC fd */
3115 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3116 env->me_fd : env->me_mfd;
3119 memset(&ov, 0, sizeof(ov));
3121 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3125 rc = pwrite(mfd, ptr, len, off);
3128 rc = rc < 0 ? ErrCode() : EIO;
3129 DPUTS("write failed, disk error?");
3130 /* On a failure, the pagecache still contains the new data.
3131 * Write some old data back, to prevent it from being used.
3132 * Use the non-SYNC fd; we know it will fail anyway.
3134 meta.mm_last_pg = metab.mm_last_pg;
3135 meta.mm_txnid = metab.mm_txnid;
3137 memset(&ov, 0, sizeof(ov));
3139 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3141 r2 = pwrite(env->me_fd, ptr, len, off);
3144 env->me_flags |= MDB_FATAL_ERROR;
3148 /* Memory ordering issues are irrelevant; since the entire writer
3149 * is wrapped by wmutex, all of these changes will become visible
3150 * after the wmutex is unlocked. Since the DB is multi-version,
3151 * readers will get consistent data regardless of how fresh or
3152 * how stale their view of these values is.
3154 env->me_txns->mti_txnid = txn->mt_txnid;
3159 /** Check both meta pages to see which one is newer.
3160 * @param[in] env the environment handle
3161 * @return meta toggle (0 or 1).
3164 mdb_env_pick_meta(const MDB_env *env)
3166 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3170 mdb_env_create(MDB_env **env)
3174 e = calloc(1, sizeof(MDB_env));
3178 e->me_maxreaders = DEFAULT_READERS;
3179 e->me_maxdbs = e->me_numdbs = 2;
3180 e->me_fd = INVALID_HANDLE_VALUE;
3181 e->me_lfd = INVALID_HANDLE_VALUE;
3182 e->me_mfd = INVALID_HANDLE_VALUE;
3183 #ifdef MDB_USE_POSIX_SEM
3184 e->me_rmutex = SEM_FAILED;
3185 e->me_wmutex = SEM_FAILED;
3187 e->me_pid = getpid();
3188 VGMEMP_CREATE(e,0,0);
3194 mdb_env_set_mapsize(MDB_env *env, size_t size)
3198 env->me_mapsize = size;
3200 env->me_maxpg = env->me_mapsize / env->me_psize;
3205 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3209 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3214 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3216 if (env->me_map || readers < 1)
3218 env->me_maxreaders = readers;
3223 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3225 if (!env || !readers)
3227 *readers = env->me_maxreaders;
3231 /** Further setup required for opening an MDB environment
3234 mdb_env_open2(MDB_env *env)
3236 unsigned int flags = env->me_flags;
3244 memset(&meta, 0, sizeof(meta));
3246 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3249 DPUTS("new mdbenv");
3253 /* Was a mapsize configured? */
3254 if (!env->me_mapsize) {
3255 /* If this is a new environment, take the default,
3256 * else use the size recorded in the existing env.
3258 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3259 } else if (env->me_mapsize < meta.mm_mapsize) {
3260 /* If the configured size is smaller, make sure it's
3261 * still big enough. Silently round up to minimum if not.
3263 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3264 if (env->me_mapsize < minsize)
3265 env->me_mapsize = minsize;
3272 LONG sizelo, sizehi;
3273 sizelo = env->me_mapsize & 0xffffffff;
3274 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3276 /* See if we should use QueryLimited */
3278 if ((rc & 0xff) > 5)
3279 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3281 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3283 /* Windows won't create mappings for zero length files.
3284 * Just allocate the maxsize right now.
3287 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3288 || !SetEndOfFile(env->me_fd)
3289 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3292 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3293 PAGE_READWRITE : PAGE_READONLY,
3294 sizehi, sizelo, NULL);
3297 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3298 FILE_MAP_WRITE : FILE_MAP_READ,
3299 0, 0, env->me_mapsize, meta.mm_address);
3300 rc = env->me_map ? 0 : ErrCode();
3308 if (flags & MDB_WRITEMAP) {
3310 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3313 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
3315 if (env->me_map == MAP_FAILED) {
3319 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3321 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3323 #ifdef POSIX_MADV_RANDOM
3324 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3325 #endif /* POSIX_MADV_RANDOM */
3326 #endif /* MADV_RANDOM */
3330 if (flags & MDB_FIXEDMAP)
3331 meta.mm_address = env->me_map;
3332 i = mdb_env_init_meta(env, &meta);
3333 if (i != MDB_SUCCESS) {
3336 } else if (meta.mm_address && env->me_map != meta.mm_address) {
3337 /* Can happen because the address argument to mmap() is just a
3338 * hint. mmap() can pick another, e.g. if the range is in use.
3339 * The MAP_FIXED flag would prevent that, but then mmap could
3340 * instead unmap existing pages to make room for the new map.
3342 return EBUSY; /* TODO: Make a new MDB_* error code? */
3344 env->me_psize = meta.mm_psize;
3345 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3346 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3348 env->me_maxpg = env->me_mapsize / env->me_psize;
3350 p = (MDB_page *)env->me_map;
3351 env->me_metas[0] = METADATA(p);
3352 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
3356 int toggle = mdb_env_pick_meta(env);
3357 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3359 DPRINTF("opened database version %u, pagesize %u",
3360 env->me_metas[0]->mm_version, env->me_psize);
3361 DPRINTF("using meta page %d", toggle);
3362 DPRINTF("depth: %u", db->md_depth);
3363 DPRINTF("entries: %"Z"u", db->md_entries);
3364 DPRINTF("branch pages: %"Z"u", db->md_branch_pages);
3365 DPRINTF("leaf pages: %"Z"u", db->md_leaf_pages);
3366 DPRINTF("overflow pages: %"Z"u", db->md_overflow_pages);
3367 DPRINTF("root: %"Z"u", db->md_root);
3375 /** Release a reader thread's slot in the reader lock table.
3376 * This function is called automatically when a thread exits.
3377 * @param[in] ptr This points to the slot in the reader lock table.
3380 mdb_env_reader_dest(void *ptr)
3382 MDB_reader *reader = ptr;
3388 /** Junk for arranging thread-specific callbacks on Windows. This is
3389 * necessarily platform and compiler-specific. Windows supports up
3390 * to 1088 keys. Let's assume nobody opens more than 64 environments
3391 * in a single process, for now. They can override this if needed.
3393 #ifndef MAX_TLS_KEYS
3394 #define MAX_TLS_KEYS 64
3396 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3397 static int mdb_tls_nkeys;
3399 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3403 case DLL_PROCESS_ATTACH: break;
3404 case DLL_THREAD_ATTACH: break;
3405 case DLL_THREAD_DETACH:
3406 for (i=0; i<mdb_tls_nkeys; i++) {
3407 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3408 mdb_env_reader_dest(r);
3411 case DLL_PROCESS_DETACH: break;
3416 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3418 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3422 /* Force some symbol references.
3423 * _tls_used forces the linker to create the TLS directory if not already done
3424 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3426 #pragma comment(linker, "/INCLUDE:_tls_used")
3427 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3428 #pragma const_seg(".CRT$XLB")
3429 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3430 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3433 #pragma comment(linker, "/INCLUDE:__tls_used")
3434 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3435 #pragma data_seg(".CRT$XLB")
3436 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3438 #endif /* WIN 32/64 */
3439 #endif /* !__GNUC__ */
3442 /** Downgrade the exclusive lock on the region back to shared */
3444 mdb_env_share_locks(MDB_env *env, int *excl)
3446 int rc = 0, toggle = mdb_env_pick_meta(env);
3448 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3453 /* First acquire a shared lock. The Unlock will
3454 * then release the existing exclusive lock.
3456 memset(&ov, 0, sizeof(ov));
3457 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3460 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3466 struct flock lock_info;
3467 /* The shared lock replaces the existing lock */
3468 memset((void *)&lock_info, 0, sizeof(lock_info));
3469 lock_info.l_type = F_RDLCK;
3470 lock_info.l_whence = SEEK_SET;
3471 lock_info.l_start = 0;
3472 lock_info.l_len = 1;
3473 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3474 (rc = ErrCode()) == EINTR) ;
3475 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3482 /** Try to get exlusive lock, otherwise shared.
3483 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3486 mdb_env_excl_lock(MDB_env *env, int *excl)
3490 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3494 memset(&ov, 0, sizeof(ov));
3495 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3502 struct flock lock_info;
3503 memset((void *)&lock_info, 0, sizeof(lock_info));
3504 lock_info.l_type = F_WRLCK;
3505 lock_info.l_whence = SEEK_SET;
3506 lock_info.l_start = 0;
3507 lock_info.l_len = 1;
3508 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3509 (rc = ErrCode()) == EINTR) ;
3513 # ifdef MDB_USE_POSIX_SEM
3514 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3517 lock_info.l_type = F_RDLCK;
3518 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3519 (rc = ErrCode()) == EINTR) ;
3527 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3529 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3531 * @(#) $Revision: 5.1 $
3532 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3533 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3535 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3539 * Please do not copyright this code. This code is in the public domain.
3541 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3542 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3543 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3544 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3545 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3546 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3547 * PERFORMANCE OF THIS SOFTWARE.
3550 * chongo <Landon Curt Noll> /\oo/\
3551 * http://www.isthe.com/chongo/
3553 * Share and Enjoy! :-)
3556 typedef unsigned long long mdb_hash_t;
3557 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3559 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3560 * @param[in] val value to hash
3561 * @param[in] hval initial value for hash
3562 * @return 64 bit hash
3564 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3565 * hval arg on the first call.
3568 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3570 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3571 unsigned char *end = s + val->mv_size;
3573 * FNV-1a hash each octet of the string
3576 /* xor the bottom with the current octet */
3577 hval ^= (mdb_hash_t)*s++;
3579 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3580 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3581 (hval << 7) + (hval << 8) + (hval << 40);
3583 /* return our new hash value */
3587 /** Hash the string and output the encoded hash.
3588 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3589 * very short name limits. We don't care about the encoding being reversible,
3590 * we just want to preserve as many bits of the input as possible in a
3591 * small printable string.
3592 * @param[in] str string to hash
3593 * @param[out] encbuf an array of 11 chars to hold the hash
3595 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3598 mdb_pack85(unsigned long l, char *out)
3602 for (i=0; i<5; i++) {
3603 *out++ = mdb_a85[l % 85];
3609 mdb_hash_enc(MDB_val *val, char *encbuf)
3611 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3612 unsigned long *l = (unsigned long *)&h;
3614 mdb_pack85(l[0], encbuf);
3615 mdb_pack85(l[1], encbuf+5);
3620 /** Open and/or initialize the lock region for the environment.
3621 * @param[in] env The MDB environment.
3622 * @param[in] lpath The pathname of the file used for the lock region.
3623 * @param[in] mode The Unix permissions for the file, if we create it.
3624 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3625 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3626 * @return 0 on success, non-zero on failure.
3629 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3632 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3634 # define MDB_ERRCODE_ROFS EROFS
3635 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3636 # define MDB_CLOEXEC O_CLOEXEC
3639 # define MDB_CLOEXEC 0
3646 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3647 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3648 FILE_ATTRIBUTE_NORMAL, NULL);
3650 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3652 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3654 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3659 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3660 /* Lose record locks when exec*() */
3661 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3662 fcntl(env->me_lfd, F_SETFD, fdflags);
3665 if (!(env->me_flags & MDB_NOTLS)) {
3666 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3669 env->me_flags |= MDB_ENV_TXKEY;
3671 /* Windows TLS callbacks need help finding their TLS info. */
3672 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3676 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3680 /* Try to get exclusive lock. If we succeed, then
3681 * nobody is using the lock region and we should initialize it.
3683 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3686 size = GetFileSize(env->me_lfd, NULL);
3688 size = lseek(env->me_lfd, 0, SEEK_END);
3689 if (size == -1) goto fail_errno;
3691 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3692 if (size < rsize && *excl > 0) {
3694 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3695 || !SetEndOfFile(env->me_lfd))
3698 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3702 size = rsize - sizeof(MDB_txninfo);
3703 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3708 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3710 if (!mh) goto fail_errno;
3711 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3713 if (!env->me_txns) goto fail_errno;
3715 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3717 if (m == MAP_FAILED) goto fail_errno;
3723 BY_HANDLE_FILE_INFORMATION stbuf;
3732 if (!mdb_sec_inited) {
3733 InitializeSecurityDescriptor(&mdb_null_sd,
3734 SECURITY_DESCRIPTOR_REVISION);
3735 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3736 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3737 mdb_all_sa.bInheritHandle = FALSE;
3738 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3741 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3742 idbuf.volume = stbuf.dwVolumeSerialNumber;
3743 idbuf.nhigh = stbuf.nFileIndexHigh;
3744 idbuf.nlow = stbuf.nFileIndexLow;
3745 val.mv_data = &idbuf;
3746 val.mv_size = sizeof(idbuf);
3747 mdb_hash_enc(&val, encbuf);
3748 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3749 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3750 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3751 if (!env->me_rmutex) goto fail_errno;
3752 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3753 if (!env->me_wmutex) goto fail_errno;
3754 #elif defined(MDB_USE_POSIX_SEM)
3763 #if defined(__NetBSD__)
3764 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3766 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3767 idbuf.dev = stbuf.st_dev;
3768 idbuf.ino = stbuf.st_ino;
3769 val.mv_data = &idbuf;
3770 val.mv_size = sizeof(idbuf);
3771 mdb_hash_enc(&val, encbuf);
3772 #ifdef MDB_SHORT_SEMNAMES
3773 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3775 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3776 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3777 /* Clean up after a previous run, if needed: Try to
3778 * remove both semaphores before doing anything else.
3780 sem_unlink(env->me_txns->mti_rmname);
3781 sem_unlink(env->me_txns->mti_wmname);
3782 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3783 O_CREAT|O_EXCL, mode, 1);
3784 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3785 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3786 O_CREAT|O_EXCL, mode, 1);
3787 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3788 #else /* MDB_USE_POSIX_SEM */
3789 pthread_mutexattr_t mattr;
3791 if ((rc = pthread_mutexattr_init(&mattr))
3792 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3793 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3794 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3796 pthread_mutexattr_destroy(&mattr);
3797 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3799 env->me_txns->mti_magic = MDB_MAGIC;
3800 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3801 env->me_txns->mti_txnid = 0;
3802 env->me_txns->mti_numreaders = 0;
3805 if (env->me_txns->mti_magic != MDB_MAGIC) {
3806 DPUTS("lock region has invalid magic");
3810 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3811 DPRINTF("lock region has format+version 0x%x, expected 0x%x",
3812 env->me_txns->mti_format, MDB_LOCK_FORMAT);
3813 rc = MDB_VERSION_MISMATCH;
3817 if (rc && rc != EACCES && rc != EAGAIN) {
3821 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3822 if (!env->me_rmutex) goto fail_errno;
3823 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3824 if (!env->me_wmutex) goto fail_errno;
3825 #elif defined(MDB_USE_POSIX_SEM)
3826 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3827 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3828 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3829 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3840 /** The name of the lock file in the DB environment */
3841 #define LOCKNAME "/lock.mdb"
3842 /** The name of the data file in the DB environment */
3843 #define DATANAME "/data.mdb"
3844 /** The suffix of the lock file when no subdir is used */
3845 #define LOCKSUFF "-lock"
3846 /** Only a subset of the @ref mdb_env flags can be changed
3847 * at runtime. Changing other flags requires closing the
3848 * environment and re-opening it with the new flags.
3850 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3851 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3854 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3856 int oflags, rc, len, excl = -1;
3857 char *lpath, *dpath;
3859 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3863 if (flags & MDB_NOSUBDIR) {
3864 rc = len + sizeof(LOCKSUFF) + len + 1;
3866 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3871 if (flags & MDB_NOSUBDIR) {
3872 dpath = lpath + len + sizeof(LOCKSUFF);
3873 sprintf(lpath, "%s" LOCKSUFF, path);
3874 strcpy(dpath, path);
3876 dpath = lpath + len + sizeof(LOCKNAME);
3877 sprintf(lpath, "%s" LOCKNAME, path);
3878 sprintf(dpath, "%s" DATANAME, path);
3882 flags |= env->me_flags;
3883 if (flags & MDB_RDONLY) {
3884 /* silently ignore WRITEMAP when we're only getting read access */
3885 flags &= ~MDB_WRITEMAP;
3887 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3888 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3891 env->me_flags = flags |= MDB_ENV_ACTIVE;
3895 env->me_path = strdup(path);
3896 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3897 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3898 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3903 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3908 if (F_ISSET(flags, MDB_RDONLY)) {
3909 oflags = GENERIC_READ;
3910 len = OPEN_EXISTING;
3912 oflags = GENERIC_READ|GENERIC_WRITE;
3915 mode = FILE_ATTRIBUTE_NORMAL;
3916 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3917 NULL, len, mode, NULL);
3919 if (F_ISSET(flags, MDB_RDONLY))
3922 oflags = O_RDWR | O_CREAT;
3924 env->me_fd = open(dpath, oflags, mode);
3926 if (env->me_fd == INVALID_HANDLE_VALUE) {
3931 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3932 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3933 env->me_mfd = env->me_fd;
3935 /* Synchronous fd for meta writes. Needed even with
3936 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3939 env->me_mfd = CreateFile(dpath, oflags,
3940 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3941 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3943 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3945 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3950 DPRINTF("opened dbenv %p", (void *) env);
3952 rc = mdb_env_share_locks(env, &excl);
3958 mdb_env_close0(env, excl);
3964 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3966 mdb_env_close0(MDB_env *env, int excl)
3970 if (!(env->me_flags & MDB_ENV_ACTIVE))
3973 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3974 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3975 free(env->me_dbxs[i].md_name.mv_data);
3977 free(env->me_dbflags);
3980 free(env->me_dirty_list);
3981 mdb_midl_free(env->me_free_pgs);
3983 if (env->me_flags & MDB_ENV_TXKEY) {
3984 pthread_key_delete(env->me_txkey);
3986 /* Delete our key from the global list */
3987 for (i=0; i<mdb_tls_nkeys; i++)
3988 if (mdb_tls_keys[i] == env->me_txkey) {
3989 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3997 munmap(env->me_map, env->me_mapsize);
3999 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4000 (void) close(env->me_mfd);
4001 if (env->me_fd != INVALID_HANDLE_VALUE)
4002 (void) close(env->me_fd);
4004 pid_t pid = env->me_pid;
4005 /* Clearing readers is done in this function because
4006 * me_txkey with its destructor must be disabled first.
4008 for (i = env->me_numreaders; --i >= 0; )
4009 if (env->me_txns->mti_readers[i].mr_pid == pid)
4010 env->me_txns->mti_readers[i].mr_pid = 0;
4012 if (env->me_rmutex) {
4013 CloseHandle(env->me_rmutex);
4014 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4016 /* Windows automatically destroys the mutexes when
4017 * the last handle closes.
4019 #elif defined(MDB_USE_POSIX_SEM)
4020 if (env->me_rmutex != SEM_FAILED) {
4021 sem_close(env->me_rmutex);
4022 if (env->me_wmutex != SEM_FAILED)
4023 sem_close(env->me_wmutex);
4024 /* If we have the filelock: If we are the
4025 * only remaining user, clean up semaphores.
4028 mdb_env_excl_lock(env, &excl);
4030 sem_unlink(env->me_txns->mti_rmname);
4031 sem_unlink(env->me_txns->mti_wmname);
4035 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4037 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4040 /* Unlock the lockfile. Windows would have unlocked it
4041 * after closing anyway, but not necessarily at once.
4043 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4046 (void) close(env->me_lfd);
4049 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4053 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4055 MDB_txn *txn = NULL;
4061 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4065 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4068 /* Do the lock/unlock of the reader mutex before starting the
4069 * write txn. Otherwise other read txns could block writers.
4071 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4076 /* We must start the actual read txn after blocking writers */
4077 mdb_txn_reset0(txn, "reset-stage1");
4079 /* Temporarily block writers until we snapshot the meta pages */
4082 rc = mdb_txn_renew0(txn);
4084 UNLOCK_MUTEX_W(env);
4089 wsize = env->me_psize * 2;
4093 DO_WRITE(rc, fd, ptr, w2, len);
4097 } else if (len > 0) {
4103 /* Non-blocking or async handles are not supported */
4109 UNLOCK_MUTEX_W(env);
4114 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4116 if (wsize > MAX_WRITE)
4120 DO_WRITE(rc, fd, ptr, w2, len);
4124 } else if (len > 0) {
4141 mdb_env_copy(MDB_env *env, const char *path)
4145 HANDLE newfd = INVALID_HANDLE_VALUE;
4147 if (env->me_flags & MDB_NOSUBDIR) {
4148 lpath = (char *)path;
4151 len += sizeof(DATANAME);
4152 lpath = malloc(len);
4155 sprintf(lpath, "%s" DATANAME, path);
4158 /* The destination path must exist, but the destination file must not.
4159 * We don't want the OS to cache the writes, since the source data is
4160 * already in the OS cache.
4163 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4164 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4166 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
4172 if (newfd == INVALID_HANDLE_VALUE) {
4177 #ifdef F_NOCACHE /* __APPLE__ */
4178 rc = fcntl(newfd, F_NOCACHE, 1);
4185 rc = mdb_env_copyfd(env, newfd);
4188 if (!(env->me_flags & MDB_NOSUBDIR))
4190 if (newfd != INVALID_HANDLE_VALUE)
4191 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4198 mdb_env_close(MDB_env *env)
4205 VGMEMP_DESTROY(env);
4206 while ((dp = env->me_dpages) != NULL) {
4207 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4208 env->me_dpages = dp->mp_next;
4212 mdb_env_close0(env, 0);
4216 /** Compare two items pointing at aligned size_t's */
4218 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4220 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4221 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4224 /** Compare two items pointing at aligned int's */
4226 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4228 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4229 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4232 /** Compare two items pointing at ints of unknown alignment.
4233 * Nodes and keys are guaranteed to be 2-byte aligned.
4236 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4238 #if BYTE_ORDER == LITTLE_ENDIAN
4239 unsigned short *u, *c;
4242 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4243 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4246 } while(!x && u > (unsigned short *)a->mv_data);
4249 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4253 /** Compare two items lexically */
4255 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4262 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4268 diff = memcmp(a->mv_data, b->mv_data, len);
4269 return diff ? diff : len_diff<0 ? -1 : len_diff;
4272 /** Compare two items in reverse byte order */
4274 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4276 const unsigned char *p1, *p2, *p1_lim;
4280 p1_lim = (const unsigned char *)a->mv_data;
4281 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4282 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4284 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4290 while (p1 > p1_lim) {
4291 diff = *--p1 - *--p2;
4295 return len_diff<0 ? -1 : len_diff;
4298 /** Search for key within a page, using binary search.
4299 * Returns the smallest entry larger or equal to the key.
4300 * If exactp is non-null, stores whether the found entry was an exact match
4301 * in *exactp (1 or 0).
4302 * Updates the cursor index with the index of the found entry.
4303 * If no entry larger or equal to the key is found, returns NULL.
4306 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4308 unsigned int i = 0, nkeys;
4311 MDB_page *mp = mc->mc_pg[mc->mc_top];
4312 MDB_node *node = NULL;
4317 nkeys = NUMKEYS(mp);
4322 COPY_PGNO(pgno, mp->mp_pgno);
4323 DPRINTF("searching %u keys in %s %spage %"Z"u",
4324 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4331 low = IS_LEAF(mp) ? 0 : 1;
4333 cmp = mc->mc_dbx->md_cmp;
4335 /* Branch pages have no data, so if using integer keys,
4336 * alignment is guaranteed. Use faster mdb_cmp_int.
4338 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4339 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4346 nodekey.mv_size = mc->mc_db->md_pad;
4347 node = NODEPTR(mp, 0); /* fake */
4348 while (low <= high) {
4349 i = (low + high) >> 1;
4350 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4351 rc = cmp(key, &nodekey);
4352 DPRINTF("found leaf index %u [%s], rc = %i",
4353 i, DKEY(&nodekey), rc);
4362 while (low <= high) {
4363 i = (low + high) >> 1;
4365 node = NODEPTR(mp, i);
4366 nodekey.mv_size = NODEKSZ(node);
4367 nodekey.mv_data = NODEKEY(node);
4369 rc = cmp(key, &nodekey);
4372 DPRINTF("found leaf index %u [%s], rc = %i",
4373 i, DKEY(&nodekey), rc);
4375 DPRINTF("found branch index %u [%s -> %"Z"u], rc = %i",
4376 i, DKEY(&nodekey), NODEPGNO(node), rc);
4387 if (rc > 0) { /* Found entry is less than the key. */
4388 i++; /* Skip to get the smallest entry larger than key. */
4390 node = NODEPTR(mp, i);
4393 *exactp = (rc == 0);
4394 /* store the key index */
4395 mc->mc_ki[mc->mc_top] = i;
4397 /* There is no entry larger or equal to the key. */
4400 /* nodeptr is fake for LEAF2 */
4406 mdb_cursor_adjust(MDB_cursor *mc, func)
4410 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4411 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4418 /** Pop a page off the top of the cursor's stack. */
4420 mdb_cursor_pop(MDB_cursor *mc)
4423 #ifndef MDB_DEBUG_SKIP
4424 MDB_page *top = mc->mc_pg[mc->mc_top];
4430 DPRINTF("popped page %"Z"u off db %u cursor %p", top->mp_pgno,
4431 mc->mc_dbi, (void *) mc);
4435 /** Push a page onto the top of the cursor's stack. */
4437 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4439 DPRINTF("pushing page %"Z"u on db %u cursor %p", mp->mp_pgno,
4440 mc->mc_dbi, (void *) mc);
4442 if (mc->mc_snum >= CURSOR_STACK) {
4443 assert(mc->mc_snum < CURSOR_STACK);
4444 return MDB_CURSOR_FULL;
4447 mc->mc_top = mc->mc_snum++;
4448 mc->mc_pg[mc->mc_top] = mp;
4449 mc->mc_ki[mc->mc_top] = 0;
4454 /** Find the address of the page corresponding to a given page number.
4455 * @param[in] txn the transaction for this access.
4456 * @param[in] pgno the page number for the page to retrieve.
4457 * @param[out] ret address of a pointer where the page's address will be stored.
4458 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4459 * @return 0 on success, non-zero on failure.
4462 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4464 MDB_env *env = txn->mt_env;
4468 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4472 MDB_ID2L dl = tx2->mt_u.dirty_list;
4474 /* Spilled pages were dirtied in this txn and flushed
4475 * because the dirty list got full. Bring this page
4476 * back in from the map (but don't unspill it here,
4477 * leave that unless page_touch happens again).
4479 if (tx2->mt_spill_pgs) {
4480 x = mdb_midl_search(tx2->mt_spill_pgs, pgno);
4481 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pgno) {
4482 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4487 unsigned x = mdb_mid2l_search(dl, pgno);
4488 if (x <= dl[0].mid && dl[x].mid == pgno) {
4494 } while ((tx2 = tx2->mt_parent) != NULL);
4497 if (pgno < txn->mt_next_pgno) {
4499 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4501 DPRINTF("page %"Z"u not found", pgno);
4503 return MDB_PAGE_NOTFOUND;
4513 /** Search for the page a given key should be in.
4514 * Pushes parent pages on the cursor stack. This function continues a
4515 * search on a cursor that has already been initialized. (Usually by
4516 * #mdb_page_search() but also by #mdb_node_move().)
4517 * @param[in,out] mc the cursor for this operation.
4518 * @param[in] key the key to search for. If NULL, search for the lowest
4519 * page. (This is used by #mdb_cursor_first().)
4520 * @param[in] modify If true, visited pages are updated with new page numbers.
4521 * @return 0 on success, non-zero on failure.
4524 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4526 MDB_page *mp = mc->mc_pg[mc->mc_top];
4531 while (IS_BRANCH(mp)) {
4535 DPRINTF("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp));
4536 assert(NUMKEYS(mp) > 1);
4537 DPRINTF("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0)));
4539 if (key == NULL) /* Initialize cursor to first page. */
4541 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4542 /* cursor to last page */
4546 node = mdb_node_search(mc, key, &exact);
4548 i = NUMKEYS(mp) - 1;
4550 i = mc->mc_ki[mc->mc_top];
4559 DPRINTF("following index %u for key [%s]",
4561 assert(i < NUMKEYS(mp));
4562 node = NODEPTR(mp, i);
4564 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4567 mc->mc_ki[mc->mc_top] = i;
4568 if ((rc = mdb_cursor_push(mc, mp)))
4572 if ((rc = mdb_page_touch(mc)) != 0)
4574 mp = mc->mc_pg[mc->mc_top];
4579 DPRINTF("internal error, index points to a %02X page!?",
4581 return MDB_CORRUPTED;
4584 DPRINTF("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4585 key ? DKEY(key) : NULL);
4586 mc->mc_flags |= C_INITIALIZED;
4587 mc->mc_flags &= ~C_EOF;
4592 /** Search for the lowest key under the current branch page.
4593 * This just bypasses a NUMKEYS check in the current page
4594 * before calling mdb_page_search_root(), because the callers
4595 * are all in situations where the current page is known to
4599 mdb_page_search_lowest(MDB_cursor *mc)
4601 MDB_page *mp = mc->mc_pg[mc->mc_top];
4602 MDB_node *node = NODEPTR(mp, 0);
4605 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4608 mc->mc_ki[mc->mc_top] = 0;
4609 if ((rc = mdb_cursor_push(mc, mp)))
4611 return mdb_page_search_root(mc, NULL, 0);
4614 /** Search for the page a given key should be in.
4615 * Pushes parent pages on the cursor stack. This function just sets up
4616 * the search; it finds the root page for \b mc's database and sets this
4617 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4618 * called to complete the search.
4619 * @param[in,out] mc the cursor for this operation.
4620 * @param[in] key the key to search for. If NULL, search for the lowest
4621 * page. (This is used by #mdb_cursor_first().)
4622 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4623 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4624 * @return 0 on success, non-zero on failure.
4627 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4632 /* Make sure the txn is still viable, then find the root from
4633 * the txn's db table.
4635 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4636 DPUTS("transaction has failed, must abort");
4639 /* Make sure we're using an up-to-date root */
4640 if (mc->mc_dbi > MAIN_DBI) {
4641 if ((*mc->mc_dbflag & DB_STALE) ||
4642 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4644 unsigned char dbflag = 0;
4645 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4646 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4649 if (*mc->mc_dbflag & DB_STALE) {
4653 MDB_node *leaf = mdb_node_search(&mc2,
4654 &mc->mc_dbx->md_name, &exact);
4656 return MDB_NOTFOUND;
4657 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4660 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4662 /* The txn may not know this DBI, or another process may
4663 * have dropped and recreated the DB with other flags.
4665 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4666 return MDB_INCOMPATIBLE;
4667 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4669 if (flags & MDB_PS_MODIFY)
4671 *mc->mc_dbflag &= ~DB_STALE;
4672 *mc->mc_dbflag |= dbflag;
4675 root = mc->mc_db->md_root;
4677 if (root == P_INVALID) { /* Tree is empty. */
4678 DPUTS("tree is empty");
4679 return MDB_NOTFOUND;
4684 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4685 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4691 DPRINTF("db %u root page %"Z"u has flags 0x%X",
4692 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4694 if (flags & MDB_PS_MODIFY) {
4695 if ((rc = mdb_page_touch(mc)))
4699 if (flags & MDB_PS_ROOTONLY)
4702 return mdb_page_search_root(mc, key, flags);
4706 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4708 MDB_txn *txn = mc->mc_txn;
4709 pgno_t pg = mp->mp_pgno;
4710 unsigned x = 0, ovpages = mp->mp_pages;
4711 MDB_env *env = txn->mt_env;
4712 MDB_IDL sl = txn->mt_spill_pgs;
4715 DPRINTF("free ov page %"Z"u (%d)", pg, ovpages);
4716 /* If the page is dirty or on the spill list we just acquired it,
4717 * so we should give it back to our current free list, if any.
4718 * Otherwise put it onto the list of pages we freed in this txn.
4720 * Won't create me_pghead: me_pglast must be inited along with it.
4721 * Unsupported in nested txns: They would need to hide the page
4722 * range in ancestor txns' dirty and spilled lists.
4724 if (env->me_pghead &&
4726 ((mp->mp_flags & P_DIRTY) ||
4727 (sl && (x = mdb_midl_search(sl, pg)) <= sl[0] && sl[x] == pg)))
4731 MDB_ID2 *dl, ix, iy;
4732 rc = mdb_midl_need(&env->me_pghead, ovpages);
4735 if (!(mp->mp_flags & P_DIRTY)) {
4736 /* This page is no longer spilled */
4737 for (; x < sl[0]; x++)
4742 /* Remove from dirty list */
4743 dl = txn->mt_u.dirty_list;
4745 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4753 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4754 txn->mt_flags |= MDB_TXN_ERROR;
4755 return MDB_CORRUPTED;
4758 if (!(env->me_flags & MDB_WRITEMAP))
4759 mdb_dpage_free(env, mp);
4761 /* Insert in me_pghead */
4762 mop = env->me_pghead;
4763 j = mop[0] + ovpages;
4764 for (i = mop[0]; i && mop[i] < pg; i--)
4770 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4774 mc->mc_db->md_overflow_pages -= ovpages;
4778 /** Return the data associated with a given node.
4779 * @param[in] txn The transaction for this operation.
4780 * @param[in] leaf The node being read.
4781 * @param[out] data Updated to point to the node's data.
4782 * @return 0 on success, non-zero on failure.
4785 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4787 MDB_page *omp; /* overflow page */
4791 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4792 data->mv_size = NODEDSZ(leaf);
4793 data->mv_data = NODEDATA(leaf);
4797 /* Read overflow data.
4799 data->mv_size = NODEDSZ(leaf);
4800 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4801 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4802 DPRINTF("read overflow page %"Z"u failed", pgno);
4805 data->mv_data = METADATA(omp);
4811 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4812 MDB_val *key, MDB_val *data)
4821 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4823 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4826 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4827 return MDB_BAD_VALSIZE;
4830 mdb_cursor_init(&mc, txn, dbi, &mx);
4831 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4834 /** Find a sibling for a page.
4835 * Replaces the page at the top of the cursor's stack with the
4836 * specified sibling, if one exists.
4837 * @param[in] mc The cursor for this operation.
4838 * @param[in] move_right Non-zero if the right sibling is requested,
4839 * otherwise the left sibling.
4840 * @return 0 on success, non-zero on failure.
4843 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4849 if (mc->mc_snum < 2) {
4850 return MDB_NOTFOUND; /* root has no siblings */
4854 DPRINTF("parent page is page %"Z"u, index %u",
4855 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4857 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4858 : (mc->mc_ki[mc->mc_top] == 0)) {
4859 DPRINTF("no more keys left, moving to %s sibling",
4860 move_right ? "right" : "left");
4861 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4862 /* undo cursor_pop before returning */
4869 mc->mc_ki[mc->mc_top]++;
4871 mc->mc_ki[mc->mc_top]--;
4872 DPRINTF("just moving to %s index key %u",
4873 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4875 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4877 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4878 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4881 mdb_cursor_push(mc, mp);
4883 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4888 /** Move the cursor to the next data item. */
4890 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4896 if (mc->mc_flags & C_EOF) {
4897 return MDB_NOTFOUND;
4900 assert(mc->mc_flags & C_INITIALIZED);
4902 mp = mc->mc_pg[mc->mc_top];
4904 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4905 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4906 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4907 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4908 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4909 if (op != MDB_NEXT || rc != MDB_NOTFOUND)
4913 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4914 if (op == MDB_NEXT_DUP)
4915 return MDB_NOTFOUND;
4919 DPRINTF("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc);
4921 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4922 DPUTS("=====> move to next sibling page");
4923 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4924 mc->mc_flags |= C_EOF;
4927 mp = mc->mc_pg[mc->mc_top];
4928 DPRINTF("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4930 mc->mc_ki[mc->mc_top]++;
4932 DPRINTF("==> cursor points to page %"Z"u with %u keys, key index %u",
4933 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4936 key->mv_size = mc->mc_db->md_pad;
4937 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4941 assert(IS_LEAF(mp));
4942 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4944 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4945 mdb_xcursor_init1(mc, leaf);
4948 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4951 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4952 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4953 if (rc != MDB_SUCCESS)
4958 MDB_GET_KEY(leaf, key);
4962 /** Move the cursor to the previous data item. */
4964 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4970 assert(mc->mc_flags & C_INITIALIZED);
4972 mp = mc->mc_pg[mc->mc_top];
4974 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4975 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4976 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4977 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4978 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4979 if (op != MDB_PREV || rc != MDB_NOTFOUND)
4982 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4983 if (op == MDB_PREV_DUP)
4984 return MDB_NOTFOUND;
4989 DPRINTF("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc);
4991 if (mc->mc_ki[mc->mc_top] == 0) {
4992 DPUTS("=====> move to prev sibling page");
4993 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
4996 mp = mc->mc_pg[mc->mc_top];
4997 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4998 DPRINTF("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
5000 mc->mc_ki[mc->mc_top]--;
5002 mc->mc_flags &= ~C_EOF;
5004 DPRINTF("==> cursor points to page %"Z"u with %u keys, key index %u",
5005 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
5008 key->mv_size = mc->mc_db->md_pad;
5009 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5013 assert(IS_LEAF(mp));
5014 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5016 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5017 mdb_xcursor_init1(mc, leaf);
5020 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5023 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5024 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5025 if (rc != MDB_SUCCESS)
5030 MDB_GET_KEY(leaf, key);
5034 /** Set the cursor on a specific data item. */
5036 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5037 MDB_cursor_op op, int *exactp)
5041 MDB_node *leaf = NULL;
5046 assert(key->mv_size > 0);
5049 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5051 /* See if we're already on the right page */
5052 if (mc->mc_flags & C_INITIALIZED) {
5055 mp = mc->mc_pg[mc->mc_top];
5057 mc->mc_ki[mc->mc_top] = 0;
5058 return MDB_NOTFOUND;
5060 if (mp->mp_flags & P_LEAF2) {
5061 nodekey.mv_size = mc->mc_db->md_pad;
5062 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5064 leaf = NODEPTR(mp, 0);
5065 MDB_GET_KEY(leaf, &nodekey);
5067 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5069 /* Probably happens rarely, but first node on the page
5070 * was the one we wanted.
5072 mc->mc_ki[mc->mc_top] = 0;
5079 unsigned int nkeys = NUMKEYS(mp);
5081 if (mp->mp_flags & P_LEAF2) {
5082 nodekey.mv_data = LEAF2KEY(mp,
5083 nkeys-1, nodekey.mv_size);
5085 leaf = NODEPTR(mp, nkeys-1);
5086 MDB_GET_KEY(leaf, &nodekey);
5088 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5090 /* last node was the one we wanted */
5091 mc->mc_ki[mc->mc_top] = nkeys-1;
5097 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5098 /* This is definitely the right page, skip search_page */
5099 if (mp->mp_flags & P_LEAF2) {
5100 nodekey.mv_data = LEAF2KEY(mp,
5101 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5103 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5104 MDB_GET_KEY(leaf, &nodekey);
5106 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5108 /* current node was the one we wanted */
5118 /* If any parents have right-sibs, search.
5119 * Otherwise, there's nothing further.
5121 for (i=0; i<mc->mc_top; i++)
5123 NUMKEYS(mc->mc_pg[i])-1)
5125 if (i == mc->mc_top) {
5126 /* There are no other pages */
5127 mc->mc_ki[mc->mc_top] = nkeys;
5128 return MDB_NOTFOUND;
5132 /* There are no other pages */
5133 mc->mc_ki[mc->mc_top] = 0;
5134 return MDB_NOTFOUND;
5138 rc = mdb_page_search(mc, key, 0);
5139 if (rc != MDB_SUCCESS)
5142 mp = mc->mc_pg[mc->mc_top];
5143 assert(IS_LEAF(mp));
5146 leaf = mdb_node_search(mc, key, exactp);
5147 if (exactp != NULL && !*exactp) {
5148 /* MDB_SET specified and not an exact match. */
5149 return MDB_NOTFOUND;
5153 DPUTS("===> inexact leaf not found, goto sibling");
5154 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5155 return rc; /* no entries matched */
5156 mp = mc->mc_pg[mc->mc_top];
5157 assert(IS_LEAF(mp));
5158 leaf = NODEPTR(mp, 0);
5162 mc->mc_flags |= C_INITIALIZED;
5163 mc->mc_flags &= ~C_EOF;
5166 key->mv_size = mc->mc_db->md_pad;
5167 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5171 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5172 mdb_xcursor_init1(mc, leaf);
5175 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5176 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5177 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5180 if (op == MDB_GET_BOTH) {
5186 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5187 if (rc != MDB_SUCCESS)
5190 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5192 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5194 rc = mc->mc_dbx->md_dcmp(data, &d2);
5196 if (op == MDB_GET_BOTH || rc > 0)
5197 return MDB_NOTFOUND;
5202 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5203 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5208 /* The key already matches in all other cases */
5209 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5210 MDB_GET_KEY(leaf, key);
5211 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
5216 /** Move the cursor to the first item in the database. */
5218 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5224 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5226 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5227 rc = mdb_page_search(mc, NULL, 0);
5228 if (rc != MDB_SUCCESS)
5231 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5233 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5234 mc->mc_flags |= C_INITIALIZED;
5235 mc->mc_flags &= ~C_EOF;
5237 mc->mc_ki[mc->mc_top] = 0;
5239 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5240 key->mv_size = mc->mc_db->md_pad;
5241 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5246 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5247 mdb_xcursor_init1(mc, leaf);
5248 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5252 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5256 MDB_GET_KEY(leaf, key);
5260 /** Move the cursor to the last item in the database. */
5262 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5268 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5270 if (!(mc->mc_flags & C_EOF)) {
5272 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5275 lkey.mv_size = MDB_MAXKEYSIZE+1;
5276 lkey.mv_data = NULL;
5277 rc = mdb_page_search(mc, &lkey, 0);
5278 if (rc != MDB_SUCCESS)
5281 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5284 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5285 mc->mc_flags |= C_INITIALIZED|C_EOF;
5286 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5288 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5289 key->mv_size = mc->mc_db->md_pad;
5290 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5295 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5296 mdb_xcursor_init1(mc, leaf);
5297 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5301 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5306 MDB_GET_KEY(leaf, key);
5311 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5316 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5321 case MDB_GET_CURRENT:
5322 if (!(mc->mc_flags & C_INITIALIZED)) {
5325 MDB_page *mp = mc->mc_pg[mc->mc_top];
5327 mc->mc_ki[mc->mc_top] = 0;
5333 key->mv_size = mc->mc_db->md_pad;
5334 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5336 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5337 MDB_GET_KEY(leaf, key);
5339 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5340 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5342 rc = mdb_node_read(mc->mc_txn, leaf, data);
5349 case MDB_GET_BOTH_RANGE:
5354 if (mc->mc_xcursor == NULL) {
5355 rc = MDB_INCOMPATIBLE;
5364 } else if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
5365 rc = MDB_BAD_VALSIZE;
5366 } else if (op == MDB_SET_RANGE)
5367 rc = mdb_cursor_set(mc, key, data, op, NULL);
5369 rc = mdb_cursor_set(mc, key, data, op, &exact);
5371 case MDB_GET_MULTIPLE:
5372 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5376 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5377 rc = MDB_INCOMPATIBLE;
5381 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5382 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5385 case MDB_NEXT_MULTIPLE:
5390 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5391 rc = MDB_INCOMPATIBLE;
5394 if (!(mc->mc_flags & C_INITIALIZED))
5395 rc = mdb_cursor_first(mc, key, data);
5397 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5398 if (rc == MDB_SUCCESS) {
5399 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5402 mx = &mc->mc_xcursor->mx_cursor;
5403 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5405 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5406 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5414 case MDB_NEXT_NODUP:
5415 if (!(mc->mc_flags & C_INITIALIZED))
5416 rc = mdb_cursor_first(mc, key, data);
5418 rc = mdb_cursor_next(mc, key, data, op);
5422 case MDB_PREV_NODUP:
5423 if (!(mc->mc_flags & C_INITIALIZED)) {
5424 rc = mdb_cursor_last(mc, key, data);
5427 mc->mc_flags |= C_INITIALIZED;
5428 mc->mc_ki[mc->mc_top]++;
5430 rc = mdb_cursor_prev(mc, key, data, op);
5433 rc = mdb_cursor_first(mc, key, data);
5436 mfunc = mdb_cursor_first;
5438 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5442 if (mc->mc_xcursor == NULL) {
5443 rc = MDB_INCOMPATIBLE;
5446 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5450 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5453 rc = mdb_cursor_last(mc, key, data);
5456 mfunc = mdb_cursor_last;
5459 DPRINTF("unhandled/unimplemented cursor operation %u", op);
5467 /** Touch all the pages in the cursor stack.
5468 * Makes sure all the pages are writable, before attempting a write operation.
5469 * @param[in] mc The cursor to operate on.
5472 mdb_cursor_touch(MDB_cursor *mc)
5476 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5479 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5480 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5483 *mc->mc_dbflag |= DB_DIRTY;
5485 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5486 rc = mdb_page_touch(mc);
5490 mc->mc_top = mc->mc_snum-1;
5494 /** Do not spill pages to disk if txn is getting full, may fail instead */
5495 #define MDB_NOSPILL 0x8000
5498 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5501 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5502 MDB_node *leaf = NULL;
5503 MDB_val xdata, *rdata, dkey;
5506 int do_sub = 0, insert = 0;
5507 unsigned int mcount = 0, dcount = 0, nospill;
5511 char dbuf[MDB_MAXKEYSIZE+1];
5512 unsigned int nflags;
5515 /* Check this first so counter will always be zero on any
5518 if (flags & MDB_MULTIPLE) {
5519 dcount = data[1].mv_size;
5520 data[1].mv_size = 0;
5521 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5522 return MDB_INCOMPATIBLE;
5525 nospill = flags & MDB_NOSPILL;
5526 flags &= ~MDB_NOSPILL;
5528 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5531 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5532 return MDB_BAD_VALSIZE;
5534 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5535 return MDB_BAD_VALSIZE;
5537 #if SIZE_MAX > MAXDATASIZE
5538 if (data->mv_size > MAXDATASIZE)
5539 return MDB_BAD_VALSIZE;
5542 DPRINTF("==> put db %u key [%s], size %"Z"u, data size %"Z"u",
5543 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
5547 if (flags == MDB_CURRENT) {
5548 if (!(mc->mc_flags & C_INITIALIZED))
5551 } else if (mc->mc_db->md_root == P_INVALID) {
5552 /* new database, cursor has nothing to point to */
5554 mc->mc_flags &= ~C_INITIALIZED;
5559 if (flags & MDB_APPEND) {
5561 rc = mdb_cursor_last(mc, &k2, &d2);
5563 rc = mc->mc_dbx->md_cmp(key, &k2);
5566 mc->mc_ki[mc->mc_top]++;
5568 /* new key is <= last key */
5573 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5575 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5576 DPRINTF("duplicate key [%s]", DKEY(key));
5578 return MDB_KEYEXIST;
5580 if (rc && rc != MDB_NOTFOUND)
5584 /* Cursor is positioned, check for room in the dirty list */
5586 if (flags & MDB_MULTIPLE) {
5588 xdata.mv_size = data->mv_size * dcount;
5592 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5596 if (rc == MDB_NO_ROOT) {
5598 /* new database, write a root leaf page */
5599 DPUTS("allocating new root leaf page");
5600 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5603 mdb_cursor_push(mc, np);
5604 mc->mc_db->md_root = np->mp_pgno;
5605 mc->mc_db->md_depth++;
5606 *mc->mc_dbflag |= DB_DIRTY;
5607 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5609 np->mp_flags |= P_LEAF2;
5610 mc->mc_flags |= C_INITIALIZED;
5612 /* make sure all cursor pages are writable */
5613 rc2 = mdb_cursor_touch(mc);
5618 /* The key already exists */
5619 if (rc == MDB_SUCCESS) {
5620 /* there's only a key anyway, so this is a no-op */
5621 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5622 unsigned int ksize = mc->mc_db->md_pad;
5623 if (key->mv_size != ksize)
5624 return MDB_BAD_VALSIZE;
5625 if (flags == MDB_CURRENT) {
5626 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5627 memcpy(ptr, key->mv_data, ksize);
5632 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5635 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5636 /* Was a single item before, must convert now */
5638 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5639 /* Just overwrite the current item */
5640 if (flags == MDB_CURRENT)
5643 dkey.mv_size = NODEDSZ(leaf);
5644 dkey.mv_data = NODEDATA(leaf);
5645 #if UINT_MAX < SIZE_MAX
5646 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5647 #ifdef MISALIGNED_OK
5648 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5650 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5653 /* if data matches, ignore it */
5654 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5655 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5657 /* create a fake page for the dup items */
5658 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5659 dkey.mv_data = dbuf;
5660 fp = (MDB_page *)&pbuf;
5661 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5662 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5663 fp->mp_lower = PAGEHDRSZ;
5664 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5665 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5666 fp->mp_flags |= P_LEAF2;
5667 fp->mp_pad = data->mv_size;
5668 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5670 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5671 (dkey.mv_size & 1) + (data->mv_size & 1);
5673 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5676 xdata.mv_size = fp->mp_upper;
5681 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5682 /* See if we need to convert from fake page to subDB */
5684 unsigned int offset;
5688 fp = NODEDATA(leaf);
5689 if (flags == MDB_CURRENT) {
5691 fp->mp_flags |= P_DIRTY;
5692 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5693 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5697 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5698 offset = fp->mp_pad;
5699 if (SIZELEFT(fp) >= offset)
5701 offset *= 4; /* space for 4 more */
5703 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5705 offset += offset & 1;
5706 fp_flags = fp->mp_flags;
5707 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5708 offset >= mc->mc_txn->mt_env->me_nodemax) {
5709 /* yes, convert it */
5711 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5712 dummy.md_pad = fp->mp_pad;
5713 dummy.md_flags = MDB_DUPFIXED;
5714 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5715 dummy.md_flags |= MDB_INTEGERKEY;
5718 dummy.md_branch_pages = 0;
5719 dummy.md_leaf_pages = 1;
5720 dummy.md_overflow_pages = 0;
5721 dummy.md_entries = NUMKEYS(fp);
5723 xdata.mv_size = sizeof(MDB_db);
5724 xdata.mv_data = &dummy;
5725 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5727 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5728 flags |= F_DUPDATA|F_SUBDATA;
5729 dummy.md_root = mp->mp_pgno;
5730 fp_flags &= ~P_SUBP;
5732 /* no, just grow it */
5734 xdata.mv_size = NODEDSZ(leaf) + offset;
5735 xdata.mv_data = &pbuf;
5736 mp = (MDB_page *)&pbuf;
5737 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5740 mp->mp_flags = fp_flags | P_DIRTY;
5741 mp->mp_pad = fp->mp_pad;
5742 mp->mp_lower = fp->mp_lower;
5743 mp->mp_upper = fp->mp_upper + offset;
5745 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5747 nsize = NODEDSZ(leaf) - fp->mp_upper;
5748 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5749 for (i=0; i<NUMKEYS(fp); i++)
5750 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5752 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5756 /* data is on sub-DB, just store it */
5757 flags |= F_DUPDATA|F_SUBDATA;
5761 /* overflow page overwrites need special handling */
5762 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5765 unsigned psize = mc->mc_txn->mt_env->me_psize;
5766 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5768 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5769 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5771 ovpages = omp->mp_pages;
5773 /* Is the ov page large enough? */
5774 if (ovpages >= dpages) {
5775 if (!(omp->mp_flags & P_DIRTY) &&
5776 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5778 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5781 level = 0; /* dirty in this txn or clean */
5784 if (omp->mp_flags & P_DIRTY) {
5785 /* yes, overwrite it. Note in this case we don't
5786 * bother to try shrinking the page if the new data
5787 * is smaller than the overflow threshold.
5790 /* It is writable only in a parent txn */
5791 size_t sz = (size_t) psize * ovpages, off;
5792 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5798 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5799 if (!(flags & MDB_RESERVE)) {
5800 /* Copy end of page, adjusting alignment so
5801 * compiler may copy words instead of bytes.
5803 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5804 memcpy((size_t *)((char *)np + off),
5805 (size_t *)((char *)omp + off), sz - off);
5808 memcpy(np, omp, sz); /* Copy beginning of page */
5811 SETDSZ(leaf, data->mv_size);
5812 if (F_ISSET(flags, MDB_RESERVE))
5813 data->mv_data = METADATA(omp);
5815 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5819 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5821 } else if (NODEDSZ(leaf) == data->mv_size) {
5822 /* same size, just replace it. Note that we could
5823 * also reuse this node if the new data is smaller,
5824 * but instead we opt to shrink the node in that case.
5826 if (F_ISSET(flags, MDB_RESERVE))
5827 data->mv_data = NODEDATA(leaf);
5828 else if (data->mv_size)
5829 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5831 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5834 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5835 mc->mc_db->md_entries--;
5837 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5844 nflags = flags & NODE_ADD_FLAGS;
5845 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5846 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5847 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5848 nflags &= ~MDB_APPEND;
5850 nflags |= MDB_SPLIT_REPLACE;
5851 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5853 /* There is room already in this leaf page. */
5854 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5855 if (rc == 0 && !do_sub && insert) {
5856 /* Adjust other cursors pointing to mp */
5857 MDB_cursor *m2, *m3;
5858 MDB_dbi dbi = mc->mc_dbi;
5859 unsigned i = mc->mc_top;
5860 MDB_page *mp = mc->mc_pg[i];
5862 if (mc->mc_flags & C_SUB)
5865 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5866 if (mc->mc_flags & C_SUB)
5867 m3 = &m2->mc_xcursor->mx_cursor;
5870 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5871 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5878 if (rc != MDB_SUCCESS)
5879 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5881 /* Now store the actual data in the child DB. Note that we're
5882 * storing the user data in the keys field, so there are strict
5883 * size limits on dupdata. The actual data fields of the child
5884 * DB are all zero size.
5891 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5892 if (flags & MDB_CURRENT) {
5893 xflags = MDB_CURRENT|MDB_NOSPILL;
5895 mdb_xcursor_init1(mc, leaf);
5896 xflags = (flags & MDB_NODUPDATA) ?
5897 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
5899 /* converted, write the original data first */
5901 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5905 /* Adjust other cursors pointing to mp */
5907 unsigned i = mc->mc_top;
5908 MDB_page *mp = mc->mc_pg[i];
5910 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5911 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5912 if (!(m2->mc_flags & C_INITIALIZED)) continue;
5913 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5914 mdb_xcursor_init1(m2, leaf);
5918 /* we've done our job */
5921 if (flags & MDB_APPENDDUP)
5922 xflags |= MDB_APPEND;
5923 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5924 if (flags & F_SUBDATA) {
5925 void *db = NODEDATA(leaf);
5926 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5929 /* sub-writes might have failed so check rc again.
5930 * Don't increment count if we just replaced an existing item.
5932 if (!rc && !(flags & MDB_CURRENT))
5933 mc->mc_db->md_entries++;
5934 if (flags & MDB_MULTIPLE) {
5937 if (mcount < dcount) {
5938 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5939 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5943 /* let caller know how many succeeded, if any */
5944 data[1].mv_size = mcount;
5948 /* If we succeeded and the key didn't exist before, make sure
5949 * the cursor is marked valid.
5952 mc->mc_flags |= C_INITIALIZED;
5957 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5962 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5965 if (!(mc->mc_flags & C_INITIALIZED))
5968 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
5970 flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
5972 rc = mdb_cursor_touch(mc);
5976 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5978 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5979 if (flags != MDB_NODUPDATA) {
5980 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5981 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5983 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
5984 /* If sub-DB still has entries, we're done */
5985 if (mc->mc_xcursor->mx_db.md_entries) {
5986 if (leaf->mn_flags & F_SUBDATA) {
5987 /* update subDB info */
5988 void *db = NODEDATA(leaf);
5989 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5992 /* shrink fake page */
5993 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5994 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5995 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5996 /* fix other sub-DB cursors pointed at this fake page */
5997 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5998 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5999 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
6000 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6001 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6004 mc->mc_db->md_entries--;
6007 /* otherwise fall thru and delete the sub-DB */
6010 if (leaf->mn_flags & F_SUBDATA) {
6011 /* add all the child DB's pages to the free list */
6012 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6013 if (rc == MDB_SUCCESS) {
6014 mc->mc_db->md_entries -=
6015 mc->mc_xcursor->mx_db.md_entries;
6020 return mdb_cursor_del0(mc, leaf);
6023 /** Allocate and initialize new pages for a database.
6024 * @param[in] mc a cursor on the database being added to.
6025 * @param[in] flags flags defining what type of page is being allocated.
6026 * @param[in] num the number of pages to allocate. This is usually 1,
6027 * unless allocating overflow pages for a large record.
6028 * @param[out] mp Address of a page, or NULL on failure.
6029 * @return 0 on success, non-zero on failure.
6032 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6037 if ((rc = mdb_page_alloc(mc, num, &np)))
6039 DPRINTF("allocated new mpage %"Z"u, page size %u",
6040 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
6041 np->mp_flags = flags | P_DIRTY;
6042 np->mp_lower = PAGEHDRSZ;
6043 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6046 mc->mc_db->md_branch_pages++;
6047 else if (IS_LEAF(np))
6048 mc->mc_db->md_leaf_pages++;
6049 else if (IS_OVERFLOW(np)) {
6050 mc->mc_db->md_overflow_pages += num;
6058 /** Calculate the size of a leaf node.
6059 * The size depends on the environment's page size; if a data item
6060 * is too large it will be put onto an overflow page and the node
6061 * size will only include the key and not the data. Sizes are always
6062 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6063 * of the #MDB_node headers.
6064 * @param[in] env The environment handle.
6065 * @param[in] key The key for the node.
6066 * @param[in] data The data for the node.
6067 * @return The number of bytes needed to store the node.
6070 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6074 sz = LEAFSIZE(key, data);
6075 if (sz >= env->me_nodemax) {
6076 /* put on overflow page */
6077 sz -= data->mv_size - sizeof(pgno_t);
6081 return sz + sizeof(indx_t);
6084 /** Calculate the size of a branch node.
6085 * The size should depend on the environment's page size but since
6086 * we currently don't support spilling large keys onto overflow
6087 * pages, it's simply the size of the #MDB_node header plus the
6088 * size of the key. Sizes are always rounded up to an even number
6089 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6090 * @param[in] env The environment handle.
6091 * @param[in] key The key for the node.
6092 * @return The number of bytes needed to store the node.
6095 mdb_branch_size(MDB_env *env, MDB_val *key)
6100 if (sz >= env->me_nodemax) {
6101 /* put on overflow page */
6102 /* not implemented */
6103 /* sz -= key->size - sizeof(pgno_t); */
6106 return sz + sizeof(indx_t);
6109 /** Add a node to the page pointed to by the cursor.
6110 * @param[in] mc The cursor for this operation.
6111 * @param[in] indx The index on the page where the new node should be added.
6112 * @param[in] key The key for the new node.
6113 * @param[in] data The data for the new node, if any.
6114 * @param[in] pgno The page number, if adding a branch node.
6115 * @param[in] flags Flags for the node.
6116 * @return 0 on success, non-zero on failure. Possible errors are:
6118 * <li>ENOMEM - failed to allocate overflow pages for the node.
6119 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6120 * should never happen since all callers already calculate the
6121 * page's free space before calling this function.
6125 mdb_node_add(MDB_cursor *mc, indx_t indx,
6126 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6129 size_t node_size = NODESIZE;
6132 MDB_page *mp = mc->mc_pg[mc->mc_top];
6133 MDB_page *ofp = NULL; /* overflow page */
6136 assert(mp->mp_upper >= mp->mp_lower);
6138 DPRINTF("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6139 IS_LEAF(mp) ? "leaf" : "branch",
6140 IS_SUBP(mp) ? "sub-" : "",
6141 mp->mp_pgno, indx, data ? data->mv_size : 0,
6142 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
6145 /* Move higher keys up one slot. */
6146 int ksize = mc->mc_db->md_pad, dif;
6147 char *ptr = LEAF2KEY(mp, indx, ksize);
6148 dif = NUMKEYS(mp) - indx;
6150 memmove(ptr+ksize, ptr, dif*ksize);
6151 /* insert new key */
6152 memcpy(ptr, key->mv_data, ksize);
6154 /* Just using these for counting */
6155 mp->mp_lower += sizeof(indx_t);
6156 mp->mp_upper -= ksize - sizeof(indx_t);
6161 node_size += key->mv_size;
6165 if (F_ISSET(flags, F_BIGDATA)) {
6166 /* Data already on overflow page. */
6167 node_size += sizeof(pgno_t);
6168 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6169 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6171 /* Put data on overflow page. */
6172 DPRINTF("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6173 data->mv_size, node_size+data->mv_size);
6174 node_size += sizeof(pgno_t);
6175 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6177 DPRINTF("allocated overflow page %"Z"u", ofp->mp_pgno);
6180 node_size += data->mv_size;
6183 node_size += node_size & 1;
6185 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6186 DPRINTF("not enough room in page %"Z"u, got %u ptrs",
6187 mp->mp_pgno, NUMKEYS(mp));
6188 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6189 mp->mp_upper - mp->mp_lower);
6190 DPRINTF("node size = %"Z"u", node_size);
6191 return MDB_PAGE_FULL;
6194 /* Move higher pointers up one slot. */
6195 for (i = NUMKEYS(mp); i > indx; i--)
6196 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6198 /* Adjust free space offsets. */
6199 ofs = mp->mp_upper - node_size;
6200 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6201 mp->mp_ptrs[indx] = ofs;
6203 mp->mp_lower += sizeof(indx_t);
6205 /* Write the node data. */
6206 node = NODEPTR(mp, indx);
6207 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6208 node->mn_flags = flags;
6210 SETDSZ(node,data->mv_size);
6215 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6220 if (F_ISSET(flags, F_BIGDATA))
6221 memcpy(node->mn_data + key->mv_size, data->mv_data,
6223 else if (F_ISSET(flags, MDB_RESERVE))
6224 data->mv_data = node->mn_data + key->mv_size;
6226 memcpy(node->mn_data + key->mv_size, data->mv_data,
6229 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6231 if (F_ISSET(flags, MDB_RESERVE))
6232 data->mv_data = METADATA(ofp);
6234 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6241 /** Delete the specified node from a page.
6242 * @param[in] mp The page to operate on.
6243 * @param[in] indx The index of the node to delete.
6244 * @param[in] ksize The size of a node. Only used if the page is
6245 * part of a #MDB_DUPFIXED database.
6248 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6251 indx_t i, j, numkeys, ptr;
6258 COPY_PGNO(pgno, mp->mp_pgno);
6259 DPRINTF("delete node %u on %s page %"Z"u", indx,
6260 IS_LEAF(mp) ? "leaf" : "branch", pgno);
6263 assert(indx < NUMKEYS(mp));
6266 int x = NUMKEYS(mp) - 1 - indx;
6267 base = LEAF2KEY(mp, indx, ksize);
6269 memmove(base, base + ksize, x * ksize);
6270 mp->mp_lower -= sizeof(indx_t);
6271 mp->mp_upper += ksize - sizeof(indx_t);
6275 node = NODEPTR(mp, indx);
6276 sz = NODESIZE + node->mn_ksize;
6278 if (F_ISSET(node->mn_flags, F_BIGDATA))
6279 sz += sizeof(pgno_t);
6281 sz += NODEDSZ(node);
6285 ptr = mp->mp_ptrs[indx];
6286 numkeys = NUMKEYS(mp);
6287 for (i = j = 0; i < numkeys; i++) {
6289 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6290 if (mp->mp_ptrs[i] < ptr)
6291 mp->mp_ptrs[j] += sz;
6296 base = (char *)mp + mp->mp_upper;
6297 memmove(base + sz, base, ptr - mp->mp_upper);
6299 mp->mp_lower -= sizeof(indx_t);
6303 /** Compact the main page after deleting a node on a subpage.
6304 * @param[in] mp The main page to operate on.
6305 * @param[in] indx The index of the subpage on the main page.
6308 mdb_node_shrink(MDB_page *mp, indx_t indx)
6315 indx_t i, numkeys, ptr;
6317 node = NODEPTR(mp, indx);
6318 sp = (MDB_page *)NODEDATA(node);
6319 osize = NODEDSZ(node);
6321 delta = sp->mp_upper - sp->mp_lower;
6322 SETDSZ(node, osize - delta);
6323 xp = (MDB_page *)((char *)sp + delta);
6325 /* shift subpage upward */
6327 nsize = NUMKEYS(sp) * sp->mp_pad;
6328 memmove(METADATA(xp), METADATA(sp), nsize);
6331 nsize = osize - sp->mp_upper;
6332 numkeys = NUMKEYS(sp);
6333 for (i=numkeys-1; i>=0; i--)
6334 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6336 xp->mp_upper = sp->mp_lower;
6337 xp->mp_lower = sp->mp_lower;
6338 xp->mp_flags = sp->mp_flags;
6339 xp->mp_pad = sp->mp_pad;
6340 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6342 /* shift lower nodes upward */
6343 ptr = mp->mp_ptrs[indx];
6344 numkeys = NUMKEYS(mp);
6345 for (i = 0; i < numkeys; i++) {
6346 if (mp->mp_ptrs[i] <= ptr)
6347 mp->mp_ptrs[i] += delta;
6350 base = (char *)mp + mp->mp_upper;
6351 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6352 mp->mp_upper += delta;
6355 /** Initial setup of a sorted-dups cursor.
6356 * Sorted duplicates are implemented as a sub-database for the given key.
6357 * The duplicate data items are actually keys of the sub-database.
6358 * Operations on the duplicate data items are performed using a sub-cursor
6359 * initialized when the sub-database is first accessed. This function does
6360 * the preliminary setup of the sub-cursor, filling in the fields that
6361 * depend only on the parent DB.
6362 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6365 mdb_xcursor_init0(MDB_cursor *mc)
6367 MDB_xcursor *mx = mc->mc_xcursor;
6369 mx->mx_cursor.mc_xcursor = NULL;
6370 mx->mx_cursor.mc_txn = mc->mc_txn;
6371 mx->mx_cursor.mc_db = &mx->mx_db;
6372 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6373 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6374 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6375 mx->mx_cursor.mc_snum = 0;
6376 mx->mx_cursor.mc_top = 0;
6377 mx->mx_cursor.mc_flags = C_SUB;
6378 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6379 mx->mx_dbx.md_dcmp = NULL;
6380 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6383 /** Final setup of a sorted-dups cursor.
6384 * Sets up the fields that depend on the data from the main cursor.
6385 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6386 * @param[in] node The data containing the #MDB_db record for the
6387 * sorted-dup database.
6390 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6392 MDB_xcursor *mx = mc->mc_xcursor;
6394 if (node->mn_flags & F_SUBDATA) {
6395 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6396 mx->mx_cursor.mc_pg[0] = 0;
6397 mx->mx_cursor.mc_snum = 0;
6398 mx->mx_cursor.mc_flags = C_SUB;
6400 MDB_page *fp = NODEDATA(node);
6401 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6402 mx->mx_db.md_flags = 0;
6403 mx->mx_db.md_depth = 1;
6404 mx->mx_db.md_branch_pages = 0;
6405 mx->mx_db.md_leaf_pages = 1;
6406 mx->mx_db.md_overflow_pages = 0;
6407 mx->mx_db.md_entries = NUMKEYS(fp);
6408 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6409 mx->mx_cursor.mc_snum = 1;
6410 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6411 mx->mx_cursor.mc_top = 0;
6412 mx->mx_cursor.mc_pg[0] = fp;
6413 mx->mx_cursor.mc_ki[0] = 0;
6414 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6415 mx->mx_db.md_flags = MDB_DUPFIXED;
6416 mx->mx_db.md_pad = fp->mp_pad;
6417 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6418 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6421 DPRINTF("Sub-db %u for db %u root page %"Z"u", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6423 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6425 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6426 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6427 #if UINT_MAX < SIZE_MAX
6428 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6429 #ifdef MISALIGNED_OK
6430 mx->mx_dbx.md_cmp = mdb_cmp_long;
6432 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6437 /** Initialize a cursor for a given transaction and database. */
6439 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6442 mc->mc_backup = NULL;
6445 mc->mc_db = &txn->mt_dbs[dbi];
6446 mc->mc_dbx = &txn->mt_dbxs[dbi];
6447 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6452 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6454 mc->mc_xcursor = mx;
6455 mdb_xcursor_init0(mc);
6457 mc->mc_xcursor = NULL;
6459 if (*mc->mc_dbflag & DB_STALE) {
6460 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6465 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6468 size_t size = sizeof(MDB_cursor);
6470 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6473 /* Allow read access to the freelist */
6474 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6477 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6478 size += sizeof(MDB_xcursor);
6480 if ((mc = malloc(size)) != NULL) {
6481 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6482 if (txn->mt_cursors) {
6483 mc->mc_next = txn->mt_cursors[dbi];
6484 txn->mt_cursors[dbi] = mc;
6485 mc->mc_flags |= C_UNTRACK;
6497 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6499 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6502 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6505 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6509 /* Return the count of duplicate data items for the current key */
6511 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6515 if (mc == NULL || countp == NULL)
6518 if (mc->mc_xcursor == NULL)
6519 return MDB_INCOMPATIBLE;
6521 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6522 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6525 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6528 *countp = mc->mc_xcursor->mx_db.md_entries;
6534 mdb_cursor_close(MDB_cursor *mc)
6536 if (mc && !mc->mc_backup) {
6537 /* remove from txn, if tracked */
6538 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6539 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6540 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6542 *prev = mc->mc_next;
6549 mdb_cursor_txn(MDB_cursor *mc)
6551 if (!mc) return NULL;
6556 mdb_cursor_dbi(MDB_cursor *mc)
6562 /** Replace the key for a node with a new key.
6563 * @param[in] mc Cursor pointing to the node to operate on.
6564 * @param[in] key The new key to use.
6565 * @return 0 on success, non-zero on failure.
6568 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6575 indx_t ptr, i, numkeys, indx;
6578 indx = mc->mc_ki[mc->mc_top];
6579 mp = mc->mc_pg[mc->mc_top];
6580 node = NODEPTR(mp, indx);
6581 ptr = mp->mp_ptrs[indx];
6585 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6586 k2.mv_data = NODEKEY(node);
6587 k2.mv_size = node->mn_ksize;
6588 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6590 mdb_dkey(&k2, kbuf2),
6596 delta0 = delta = key->mv_size - node->mn_ksize;
6598 /* Must be 2-byte aligned. If new key is
6599 * shorter by 1, the shift will be skipped.
6601 delta += (delta & 1);
6603 if (delta > 0 && SIZELEFT(mp) < delta) {
6605 /* not enough space left, do a delete and split */
6606 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6607 pgno = NODEPGNO(node);
6608 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6609 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6612 numkeys = NUMKEYS(mp);
6613 for (i = 0; i < numkeys; i++) {
6614 if (mp->mp_ptrs[i] <= ptr)
6615 mp->mp_ptrs[i] -= delta;
6618 base = (char *)mp + mp->mp_upper;
6619 len = ptr - mp->mp_upper + NODESIZE;
6620 memmove(base - delta, base, len);
6621 mp->mp_upper -= delta;
6623 node = NODEPTR(mp, indx);
6626 /* But even if no shift was needed, update ksize */
6628 node->mn_ksize = key->mv_size;
6631 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6637 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6639 /** Move a node from csrc to cdst.
6642 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6649 unsigned short flags;
6653 /* Mark src and dst as dirty. */
6654 if ((rc = mdb_page_touch(csrc)) ||
6655 (rc = mdb_page_touch(cdst)))
6658 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6659 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6660 key.mv_size = csrc->mc_db->md_pad;
6661 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6663 data.mv_data = NULL;
6667 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6668 assert(!((long)srcnode&1));
6669 srcpg = NODEPGNO(srcnode);
6670 flags = srcnode->mn_flags;
6671 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6672 unsigned int snum = csrc->mc_snum;
6674 /* must find the lowest key below src */
6675 mdb_page_search_lowest(csrc);
6676 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6677 key.mv_size = csrc->mc_db->md_pad;
6678 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6680 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6681 key.mv_size = NODEKSZ(s2);
6682 key.mv_data = NODEKEY(s2);
6684 csrc->mc_snum = snum--;
6685 csrc->mc_top = snum;
6687 key.mv_size = NODEKSZ(srcnode);
6688 key.mv_data = NODEKEY(srcnode);
6690 data.mv_size = NODEDSZ(srcnode);
6691 data.mv_data = NODEDATA(srcnode);
6693 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6694 unsigned int snum = cdst->mc_snum;
6697 /* must find the lowest key below dst */
6698 mdb_page_search_lowest(cdst);
6699 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6700 bkey.mv_size = cdst->mc_db->md_pad;
6701 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6703 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6704 bkey.mv_size = NODEKSZ(s2);
6705 bkey.mv_data = NODEKEY(s2);
6707 cdst->mc_snum = snum--;
6708 cdst->mc_top = snum;
6709 mdb_cursor_copy(cdst, &mn);
6711 rc = mdb_update_key(&mn, &bkey);
6716 DPRINTF("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6717 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6718 csrc->mc_ki[csrc->mc_top],
6720 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6721 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6723 /* Add the node to the destination page.
6725 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6726 if (rc != MDB_SUCCESS)
6729 /* Delete the node from the source page.
6731 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6734 /* Adjust other cursors pointing to mp */
6735 MDB_cursor *m2, *m3;
6736 MDB_dbi dbi = csrc->mc_dbi;
6737 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6739 if (csrc->mc_flags & C_SUB)
6742 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6743 if (csrc->mc_flags & C_SUB)
6744 m3 = &m2->mc_xcursor->mx_cursor;
6747 if (m3 == csrc) continue;
6748 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6749 csrc->mc_ki[csrc->mc_top]) {
6750 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6751 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6756 /* Update the parent separators.
6758 if (csrc->mc_ki[csrc->mc_top] == 0) {
6759 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6760 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6761 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6763 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6764 key.mv_size = NODEKSZ(srcnode);
6765 key.mv_data = NODEKEY(srcnode);
6767 DPRINTF("update separator for source page %"Z"u to [%s]",
6768 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6769 mdb_cursor_copy(csrc, &mn);
6772 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6775 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6777 indx_t ix = csrc->mc_ki[csrc->mc_top];
6778 nullkey.mv_size = 0;
6779 csrc->mc_ki[csrc->mc_top] = 0;
6780 rc = mdb_update_key(csrc, &nullkey);
6781 csrc->mc_ki[csrc->mc_top] = ix;
6782 assert(rc == MDB_SUCCESS);
6786 if (cdst->mc_ki[cdst->mc_top] == 0) {
6787 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6788 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6789 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6791 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6792 key.mv_size = NODEKSZ(srcnode);
6793 key.mv_data = NODEKEY(srcnode);
6795 DPRINTF("update separator for destination page %"Z"u to [%s]",
6796 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6797 mdb_cursor_copy(cdst, &mn);
6800 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6803 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6805 indx_t ix = cdst->mc_ki[cdst->mc_top];
6806 nullkey.mv_size = 0;
6807 cdst->mc_ki[cdst->mc_top] = 0;
6808 rc = mdb_update_key(cdst, &nullkey);
6809 cdst->mc_ki[cdst->mc_top] = ix;
6810 assert(rc == MDB_SUCCESS);
6817 /** Merge one page into another.
6818 * The nodes from the page pointed to by \b csrc will
6819 * be copied to the page pointed to by \b cdst and then
6820 * the \b csrc page will be freed.
6821 * @param[in] csrc Cursor pointing to the source page.
6822 * @param[in] cdst Cursor pointing to the destination page.
6825 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6833 DPRINTF("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6834 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6836 assert(csrc->mc_snum > 1); /* can't merge root page */
6837 assert(cdst->mc_snum > 1);
6839 /* Mark dst as dirty. */
6840 if ((rc = mdb_page_touch(cdst)))
6843 /* Move all nodes from src to dst.
6845 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6846 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6847 key.mv_size = csrc->mc_db->md_pad;
6848 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6849 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6850 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6851 if (rc != MDB_SUCCESS)
6853 key.mv_data = (char *)key.mv_data + key.mv_size;
6856 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6857 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6858 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6859 unsigned int snum = csrc->mc_snum;
6861 /* must find the lowest key below src */
6862 mdb_page_search_lowest(csrc);
6863 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6864 key.mv_size = csrc->mc_db->md_pad;
6865 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6867 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6868 key.mv_size = NODEKSZ(s2);
6869 key.mv_data = NODEKEY(s2);
6871 csrc->mc_snum = snum--;
6872 csrc->mc_top = snum;
6874 key.mv_size = srcnode->mn_ksize;
6875 key.mv_data = NODEKEY(srcnode);
6878 data.mv_size = NODEDSZ(srcnode);
6879 data.mv_data = NODEDATA(srcnode);
6880 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6881 if (rc != MDB_SUCCESS)
6886 DPRINTF("dst page %"Z"u now has %u keys (%.1f%% filled)",
6887 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);
6889 /* Unlink the src page from parent and add to free list.
6891 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6892 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6895 rc = mdb_update_key(csrc, &key);
6901 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6902 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6905 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6906 csrc->mc_db->md_leaf_pages--;
6908 csrc->mc_db->md_branch_pages--;
6910 /* Adjust other cursors pointing to mp */
6911 MDB_cursor *m2, *m3;
6912 MDB_dbi dbi = csrc->mc_dbi;
6913 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6915 if (csrc->mc_flags & C_SUB)
6918 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6919 if (csrc->mc_flags & C_SUB)
6920 m3 = &m2->mc_xcursor->mx_cursor;
6923 if (m3 == csrc) continue;
6924 if (m3->mc_snum < csrc->mc_snum) continue;
6925 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6926 m3->mc_pg[csrc->mc_top] = mp;
6927 m3->mc_ki[csrc->mc_top] += nkeys;
6931 mdb_cursor_pop(csrc);
6933 return mdb_rebalance(csrc);
6936 /** Copy the contents of a cursor.
6937 * @param[in] csrc The cursor to copy from.
6938 * @param[out] cdst The cursor to copy to.
6941 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6945 cdst->mc_txn = csrc->mc_txn;
6946 cdst->mc_dbi = csrc->mc_dbi;
6947 cdst->mc_db = csrc->mc_db;
6948 cdst->mc_dbx = csrc->mc_dbx;
6949 cdst->mc_snum = csrc->mc_snum;
6950 cdst->mc_top = csrc->mc_top;
6951 cdst->mc_flags = csrc->mc_flags;
6953 for (i=0; i<csrc->mc_snum; i++) {
6954 cdst->mc_pg[i] = csrc->mc_pg[i];
6955 cdst->mc_ki[i] = csrc->mc_ki[i];
6959 /** Rebalance the tree after a delete operation.
6960 * @param[in] mc Cursor pointing to the page where rebalancing
6962 * @return 0 on success, non-zero on failure.
6965 mdb_rebalance(MDB_cursor *mc)
6969 unsigned int ptop, minkeys;
6972 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6976 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6977 DPRINTF("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
6978 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6979 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6983 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6984 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6987 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6988 DPRINTF("no need to rebalance page %"Z"u, above fill threshold",
6994 if (mc->mc_snum < 2) {
6995 MDB_page *mp = mc->mc_pg[0];
6997 DPUTS("Can't rebalance a subpage, ignoring");
7000 if (NUMKEYS(mp) == 0) {
7001 DPUTS("tree is completely empty");
7002 mc->mc_db->md_root = P_INVALID;
7003 mc->mc_db->md_depth = 0;
7004 mc->mc_db->md_leaf_pages = 0;
7005 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7008 /* Adjust cursors pointing to mp */
7012 MDB_cursor *m2, *m3;
7013 MDB_dbi dbi = mc->mc_dbi;
7015 if (mc->mc_flags & C_SUB)
7018 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7019 if (mc->mc_flags & C_SUB)
7020 m3 = &m2->mc_xcursor->mx_cursor;
7023 if (m3->mc_snum < mc->mc_snum) continue;
7024 if (m3->mc_pg[0] == mp) {
7030 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7031 DPUTS("collapsing root page!");
7032 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7035 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7036 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7039 mc->mc_db->md_depth--;
7040 mc->mc_db->md_branch_pages--;
7041 mc->mc_ki[0] = mc->mc_ki[1];
7043 /* Adjust other cursors pointing to mp */
7044 MDB_cursor *m2, *m3;
7045 MDB_dbi dbi = mc->mc_dbi;
7047 if (mc->mc_flags & C_SUB)
7050 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7051 if (mc->mc_flags & C_SUB)
7052 m3 = &m2->mc_xcursor->mx_cursor;
7055 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7056 if (m3->mc_pg[0] == mp) {
7057 m3->mc_pg[0] = mc->mc_pg[0];
7060 m3->mc_ki[0] = m3->mc_ki[1];
7065 DPUTS("root page doesn't need rebalancing");
7069 /* The parent (branch page) must have at least 2 pointers,
7070 * otherwise the tree is invalid.
7072 ptop = mc->mc_top-1;
7073 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7075 /* Leaf page fill factor is below the threshold.
7076 * Try to move keys from left or right neighbor, or
7077 * merge with a neighbor page.
7082 mdb_cursor_copy(mc, &mn);
7083 mn.mc_xcursor = NULL;
7085 if (mc->mc_ki[ptop] == 0) {
7086 /* We're the leftmost leaf in our parent.
7088 DPUTS("reading right neighbor");
7090 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7091 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7094 mn.mc_ki[mn.mc_top] = 0;
7095 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7097 /* There is at least one neighbor to the left.
7099 DPUTS("reading left neighbor");
7101 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7102 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7105 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7106 mc->mc_ki[mc->mc_top] = 0;
7109 DPRINTF("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7110 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);
7112 /* If the neighbor page is above threshold and has enough keys,
7113 * move one key from it. Otherwise we should try to merge them.
7114 * (A branch page must never have less than 2 keys.)
7116 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7117 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7118 return mdb_node_move(&mn, mc);
7120 if (mc->mc_ki[ptop] == 0)
7121 rc = mdb_page_merge(&mn, mc);
7123 rc = mdb_page_merge(mc, &mn);
7124 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7129 /** Complete a delete operation started by #mdb_cursor_del(). */
7131 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7137 mp = mc->mc_pg[mc->mc_top];
7138 ki = mc->mc_ki[mc->mc_top];
7140 /* add overflow pages to free list */
7141 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7145 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7146 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7147 (rc = mdb_ovpage_free(mc, omp)))
7150 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7151 mc->mc_db->md_entries--;
7152 rc = mdb_rebalance(mc);
7153 if (rc != MDB_SUCCESS)
7154 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7155 /* if mc points past last node in page, invalidate */
7156 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7157 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7160 /* Adjust other cursors pointing to mp */
7163 MDB_dbi dbi = mc->mc_dbi;
7165 mp = mc->mc_pg[mc->mc_top];
7166 nkeys = NUMKEYS(mp);
7167 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7170 if (!(m2->mc_flags & C_INITIALIZED))
7172 if (m2->mc_pg[mc->mc_top] == mp) {
7173 if (m2->mc_ki[mc->mc_top] > ki)
7174 m2->mc_ki[mc->mc_top]--;
7175 if (m2->mc_ki[mc->mc_top] >= nkeys)
7176 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
7185 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7186 MDB_val *key, MDB_val *data)
7191 MDB_val rdata, *xdata;
7195 assert(key != NULL);
7197 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
7199 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7202 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7206 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7207 return MDB_BAD_VALSIZE;
7210 mdb_cursor_init(&mc, txn, dbi, &mx);
7213 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7214 /* must ignore any data */
7225 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7227 /* let mdb_page_split know about this cursor if needed:
7228 * delete will trigger a rebalance; if it needs to move
7229 * a node from one page to another, it will have to
7230 * update the parent's separator key(s). If the new sepkey
7231 * is larger than the current one, the parent page may
7232 * run out of space, triggering a split. We need this
7233 * cursor to be consistent until the end of the rebalance.
7235 mc.mc_flags |= C_UNTRACK;
7236 mc.mc_next = txn->mt_cursors[dbi];
7237 txn->mt_cursors[dbi] = &mc;
7238 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7239 txn->mt_cursors[dbi] = mc.mc_next;
7244 /** Split a page and insert a new node.
7245 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7246 * The cursor will be updated to point to the actual page and index where
7247 * the node got inserted after the split.
7248 * @param[in] newkey The key for the newly inserted node.
7249 * @param[in] newdata The data for the newly inserted node.
7250 * @param[in] newpgno The page number, if the new node is a branch node.
7251 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7252 * @return 0 on success, non-zero on failure.
7255 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7256 unsigned int nflags)
7259 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7262 unsigned int i, j, split_indx, nkeys, pmax;
7264 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7266 MDB_page *mp, *rp, *pp;
7271 mp = mc->mc_pg[mc->mc_top];
7272 newindx = mc->mc_ki[mc->mc_top];
7274 DPRINTF("-----> splitting %s page %"Z"u and adding [%s] at index %i",
7275 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7276 DKEY(newkey), mc->mc_ki[mc->mc_top]);
7278 /* Create a right sibling. */
7279 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7281 DPRINTF("new right sibling: page %"Z"u", rp->mp_pgno);
7283 if (mc->mc_snum < 2) {
7284 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7286 /* shift current top to make room for new parent */
7287 mc->mc_pg[1] = mc->mc_pg[0];
7288 mc->mc_ki[1] = mc->mc_ki[0];
7291 mc->mc_db->md_root = pp->mp_pgno;
7292 DPRINTF("root split! new root = %"Z"u", pp->mp_pgno);
7293 mc->mc_db->md_depth++;
7296 /* Add left (implicit) pointer. */
7297 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7298 /* undo the pre-push */
7299 mc->mc_pg[0] = mc->mc_pg[1];
7300 mc->mc_ki[0] = mc->mc_ki[1];
7301 mc->mc_db->md_root = mp->mp_pgno;
7302 mc->mc_db->md_depth--;
7309 ptop = mc->mc_top-1;
7310 DPRINTF("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno);
7313 mc->mc_flags |= C_SPLITTING;
7314 mdb_cursor_copy(mc, &mn);
7315 mn.mc_pg[mn.mc_top] = rp;
7316 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7318 if (nflags & MDB_APPEND) {
7319 mn.mc_ki[mn.mc_top] = 0;
7321 split_indx = newindx;
7326 nkeys = NUMKEYS(mp);
7327 split_indx = nkeys / 2;
7328 if (newindx < split_indx)
7334 unsigned int lsize, rsize, ksize;
7335 /* Move half of the keys to the right sibling */
7337 x = mc->mc_ki[mc->mc_top] - split_indx;
7338 ksize = mc->mc_db->md_pad;
7339 split = LEAF2KEY(mp, split_indx, ksize);
7340 rsize = (nkeys - split_indx) * ksize;
7341 lsize = (nkeys - split_indx) * sizeof(indx_t);
7342 mp->mp_lower -= lsize;
7343 rp->mp_lower += lsize;
7344 mp->mp_upper += rsize - lsize;
7345 rp->mp_upper -= rsize - lsize;
7346 sepkey.mv_size = ksize;
7347 if (newindx == split_indx) {
7348 sepkey.mv_data = newkey->mv_data;
7350 sepkey.mv_data = split;
7353 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7354 memcpy(rp->mp_ptrs, split, rsize);
7355 sepkey.mv_data = rp->mp_ptrs;
7356 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7357 memcpy(ins, newkey->mv_data, ksize);
7358 mp->mp_lower += sizeof(indx_t);
7359 mp->mp_upper -= ksize - sizeof(indx_t);
7362 memcpy(rp->mp_ptrs, split, x * ksize);
7363 ins = LEAF2KEY(rp, x, ksize);
7364 memcpy(ins, newkey->mv_data, ksize);
7365 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7366 rp->mp_lower += sizeof(indx_t);
7367 rp->mp_upper -= ksize - sizeof(indx_t);
7368 mc->mc_ki[mc->mc_top] = x;
7369 mc->mc_pg[mc->mc_top] = rp;
7374 /* For leaf pages, check the split point based on what
7375 * fits where, since otherwise mdb_node_add can fail.
7377 * This check is only needed when the data items are
7378 * relatively large, such that being off by one will
7379 * make the difference between success or failure.
7381 * It's also relevant if a page happens to be laid out
7382 * such that one half of its nodes are all "small" and
7383 * the other half of its nodes are "large." If the new
7384 * item is also "large" and falls on the half with
7385 * "large" nodes, it also may not fit.
7388 unsigned int psize, nsize;
7389 /* Maximum free space in an empty page */
7390 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7391 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7392 if ((nkeys < 20) || (nsize > pmax/16)) {
7393 if (newindx <= split_indx) {
7396 for (i=0; i<split_indx; i++) {
7397 node = NODEPTR(mp, i);
7398 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7399 if (F_ISSET(node->mn_flags, F_BIGDATA))
7400 psize += sizeof(pgno_t);
7402 psize += NODEDSZ(node);
7406 split_indx = newindx;
7417 for (i=nkeys-1; i>=split_indx; i--) {
7418 node = NODEPTR(mp, i);
7419 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7420 if (F_ISSET(node->mn_flags, F_BIGDATA))
7421 psize += sizeof(pgno_t);
7423 psize += NODEDSZ(node);
7427 split_indx = newindx;
7438 /* First find the separating key between the split pages.
7439 * The case where newindx == split_indx is ambiguous; the
7440 * new item could go to the new page or stay on the original
7441 * page. If newpos == 1 it goes to the new page.
7443 if (newindx == split_indx && newpos) {
7444 sepkey.mv_size = newkey->mv_size;
7445 sepkey.mv_data = newkey->mv_data;
7447 node = NODEPTR(mp, split_indx);
7448 sepkey.mv_size = node->mn_ksize;
7449 sepkey.mv_data = NODEKEY(node);
7453 DPRINTF("separator is [%s]", DKEY(&sepkey));
7455 /* Copy separator key to the parent.
7457 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7461 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7464 if (mn.mc_snum == mc->mc_snum) {
7465 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7466 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7467 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7468 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7473 /* Right page might now have changed parent.
7474 * Check if left page also changed parent.
7476 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7477 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7478 for (i=0; i<ptop; i++) {
7479 mc->mc_pg[i] = mn.mc_pg[i];
7480 mc->mc_ki[i] = mn.mc_ki[i];
7482 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7483 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7487 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7490 mc->mc_flags ^= C_SPLITTING;
7491 if (rc != MDB_SUCCESS) {
7494 if (nflags & MDB_APPEND) {
7495 mc->mc_pg[mc->mc_top] = rp;
7496 mc->mc_ki[mc->mc_top] = 0;
7497 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7500 for (i=0; i<mc->mc_top; i++)
7501 mc->mc_ki[i] = mn.mc_ki[i];
7508 /* Move half of the keys to the right sibling. */
7510 /* grab a page to hold a temporary copy */
7511 copy = mdb_page_malloc(mc->mc_txn, 1);
7515 copy->mp_pgno = mp->mp_pgno;
7516 copy->mp_flags = mp->mp_flags;
7517 copy->mp_lower = PAGEHDRSZ;
7518 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7519 mc->mc_pg[mc->mc_top] = copy;
7520 for (i = j = 0; i <= nkeys; j++) {
7521 if (i == split_indx) {
7522 /* Insert in right sibling. */
7523 /* Reset insert index for right sibling. */
7524 if (i != newindx || (newpos ^ ins_new)) {
7526 mc->mc_pg[mc->mc_top] = rp;
7530 if (i == newindx && !ins_new) {
7531 /* Insert the original entry that caused the split. */
7532 rkey.mv_data = newkey->mv_data;
7533 rkey.mv_size = newkey->mv_size;
7542 /* Update index for the new key. */
7543 mc->mc_ki[mc->mc_top] = j;
7544 } else if (i == nkeys) {
7547 node = NODEPTR(mp, i);
7548 rkey.mv_data = NODEKEY(node);
7549 rkey.mv_size = node->mn_ksize;
7551 xdata.mv_data = NODEDATA(node);
7552 xdata.mv_size = NODEDSZ(node);
7555 pgno = NODEPGNO(node);
7556 flags = node->mn_flags;
7561 if (!IS_LEAF(mp) && j == 0) {
7562 /* First branch index doesn't need key data. */
7566 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7570 nkeys = NUMKEYS(copy);
7571 for (i=0; i<nkeys; i++)
7572 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7573 mp->mp_lower = copy->mp_lower;
7574 mp->mp_upper = copy->mp_upper;
7575 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7576 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7578 /* reset back to original page */
7579 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7580 mc->mc_pg[mc->mc_top] = mp;
7581 if (nflags & MDB_RESERVE) {
7582 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7583 if (!(node->mn_flags & F_BIGDATA))
7584 newdata->mv_data = NODEDATA(node);
7588 /* Make sure mc_ki is still valid.
7590 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7591 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7592 for (i=0; i<ptop; i++) {
7593 mc->mc_pg[i] = mn.mc_pg[i];
7594 mc->mc_ki[i] = mn.mc_ki[i];
7596 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7597 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7601 /* return tmp page to freelist */
7602 mdb_page_free(mc->mc_txn->mt_env, copy);
7605 /* Adjust other cursors pointing to mp */
7606 MDB_cursor *m2, *m3;
7607 MDB_dbi dbi = mc->mc_dbi;
7608 int fixup = NUMKEYS(mp);
7610 if (mc->mc_flags & C_SUB)
7613 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7614 if (mc->mc_flags & C_SUB)
7615 m3 = &m2->mc_xcursor->mx_cursor;
7620 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7622 if (m3->mc_flags & C_SPLITTING)
7627 for (k=m3->mc_top; k>=0; k--) {
7628 m3->mc_ki[k+1] = m3->mc_ki[k];
7629 m3->mc_pg[k+1] = m3->mc_pg[k];
7631 if (m3->mc_ki[0] >= split_indx) {
7636 m3->mc_pg[0] = mc->mc_pg[0];
7640 if (m3->mc_pg[mc->mc_top] == mp) {
7641 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7642 m3->mc_ki[mc->mc_top]++;
7643 if (m3->mc_ki[mc->mc_top] >= fixup) {
7644 m3->mc_pg[mc->mc_top] = rp;
7645 m3->mc_ki[mc->mc_top] -= fixup;
7646 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7648 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7649 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7658 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7659 MDB_val *key, MDB_val *data, unsigned int flags)
7664 assert(key != NULL);
7665 assert(data != NULL);
7667 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7670 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7674 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7675 return MDB_BAD_VALSIZE;
7678 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7681 mdb_cursor_init(&mc, txn, dbi, &mx);
7682 return mdb_cursor_put(&mc, key, data, flags);
7686 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7688 if ((flag & CHANGEABLE) != flag)
7691 env->me_flags |= flag;
7693 env->me_flags &= ~flag;
7698 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7703 *arg = env->me_flags;
7708 mdb_env_get_path(MDB_env *env, const char **arg)
7713 *arg = env->me_path;
7717 /** Common code for #mdb_stat() and #mdb_env_stat().
7718 * @param[in] env the environment to operate in.
7719 * @param[in] db the #MDB_db record containing the stats to return.
7720 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7721 * @return 0, this function always succeeds.
7724 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7726 arg->ms_psize = env->me_psize;
7727 arg->ms_depth = db->md_depth;
7728 arg->ms_branch_pages = db->md_branch_pages;
7729 arg->ms_leaf_pages = db->md_leaf_pages;
7730 arg->ms_overflow_pages = db->md_overflow_pages;
7731 arg->ms_entries = db->md_entries;
7736 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7740 if (env == NULL || arg == NULL)
7743 toggle = mdb_env_pick_meta(env);
7745 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7749 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7753 if (env == NULL || arg == NULL)
7756 toggle = mdb_env_pick_meta(env);
7757 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7758 arg->me_mapsize = env->me_mapsize;
7759 arg->me_maxreaders = env->me_maxreaders;
7761 /* me_numreaders may be zero if this process never used any readers. Use
7762 * the shared numreader count if it exists.
7764 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7766 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7767 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7771 /** Set the default comparison functions for a database.
7772 * Called immediately after a database is opened to set the defaults.
7773 * The user can then override them with #mdb_set_compare() or
7774 * #mdb_set_dupsort().
7775 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7776 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7779 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7781 uint16_t f = txn->mt_dbs[dbi].md_flags;
7783 txn->mt_dbxs[dbi].md_cmp =
7784 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7785 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7787 txn->mt_dbxs[dbi].md_dcmp =
7788 !(f & MDB_DUPSORT) ? 0 :
7789 ((f & MDB_INTEGERDUP)
7790 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7791 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7794 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7799 int rc, dbflag, exact;
7800 unsigned int unused = 0;
7803 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7804 mdb_default_cmp(txn, FREE_DBI);
7807 if ((flags & VALID_FLAGS) != flags)
7813 if (flags & PERSISTENT_FLAGS) {
7814 uint16_t f2 = flags & PERSISTENT_FLAGS;
7815 /* make sure flag changes get committed */
7816 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7817 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7818 txn->mt_flags |= MDB_TXN_DIRTY;
7821 mdb_default_cmp(txn, MAIN_DBI);
7825 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7826 mdb_default_cmp(txn, MAIN_DBI);
7829 /* Is the DB already open? */
7831 for (i=2; i<txn->mt_numdbs; i++) {
7832 if (!txn->mt_dbxs[i].md_name.mv_size) {
7833 /* Remember this free slot */
7834 if (!unused) unused = i;
7837 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7838 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7844 /* If no free slot and max hit, fail */
7845 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7846 return MDB_DBS_FULL;
7848 /* Cannot mix named databases with some mainDB flags */
7849 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7850 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7852 /* Find the DB info */
7853 dbflag = DB_NEW|DB_VALID;
7856 key.mv_data = (void *)name;
7857 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7858 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7859 if (rc == MDB_SUCCESS) {
7860 /* make sure this is actually a DB */
7861 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7862 if (!(node->mn_flags & F_SUBDATA))
7863 return MDB_INCOMPATIBLE;
7864 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7865 /* Create if requested */
7867 data.mv_size = sizeof(MDB_db);
7868 data.mv_data = &dummy;
7869 memset(&dummy, 0, sizeof(dummy));
7870 dummy.md_root = P_INVALID;
7871 dummy.md_flags = flags & PERSISTENT_FLAGS;
7872 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7876 /* OK, got info, add to table */
7877 if (rc == MDB_SUCCESS) {
7878 unsigned int slot = unused ? unused : txn->mt_numdbs;
7879 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7880 txn->mt_dbxs[slot].md_name.mv_size = len;
7881 txn->mt_dbxs[slot].md_rel = NULL;
7882 txn->mt_dbflags[slot] = dbflag;
7883 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7885 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7886 mdb_default_cmp(txn, slot);
7895 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7897 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7900 if (txn->mt_dbflags[dbi] & DB_STALE) {
7903 /* Stale, must read the DB's root. cursor_init does it for us. */
7904 mdb_cursor_init(&mc, txn, dbi, &mx);
7906 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7909 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7912 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7914 ptr = env->me_dbxs[dbi].md_name.mv_data;
7915 env->me_dbxs[dbi].md_name.mv_data = NULL;
7916 env->me_dbxs[dbi].md_name.mv_size = 0;
7917 env->me_dbflags[dbi] = 0;
7921 int mdb_dbi_flags(MDB_env *env, MDB_dbi dbi, unsigned int *flags)
7923 /* We could return the flags for the FREE_DBI too but what's the point? */
7924 if (dbi < MAIN_DBI || dbi >= env->me_numdbs)
7926 *flags = env->me_dbflags[dbi];
7930 /** Add all the DB's pages to the free list.
7931 * @param[in] mc Cursor on the DB to free.
7932 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7933 * @return 0 on success, non-zero on failure.
7936 mdb_drop0(MDB_cursor *mc, int subs)
7940 rc = mdb_page_search(mc, NULL, 0);
7941 if (rc == MDB_SUCCESS) {
7942 MDB_txn *txn = mc->mc_txn;
7947 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7948 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7951 mdb_cursor_copy(mc, &mx);
7952 while (mc->mc_snum > 0) {
7953 MDB_page *mp = mc->mc_pg[mc->mc_top];
7954 unsigned n = NUMKEYS(mp);
7956 for (i=0; i<n; i++) {
7957 ni = NODEPTR(mp, i);
7958 if (ni->mn_flags & F_BIGDATA) {
7961 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7962 rc = mdb_page_get(txn, pg, &omp, NULL);
7965 assert(IS_OVERFLOW(omp));
7966 rc = mdb_midl_append_range(&txn->mt_free_pgs,
7970 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7971 mdb_xcursor_init1(mc, ni);
7972 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7978 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
7980 for (i=0; i<n; i++) {
7982 ni = NODEPTR(mp, i);
7985 mdb_midl_xappend(txn->mt_free_pgs, pg);
7990 mc->mc_ki[mc->mc_top] = i;
7991 rc = mdb_cursor_sibling(mc, 1);
7993 /* no more siblings, go back to beginning
7994 * of previous level.
7998 for (i=1; i<mc->mc_snum; i++) {
8000 mc->mc_pg[i] = mx.mc_pg[i];
8005 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8006 } else if (rc == MDB_NOTFOUND) {
8012 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8014 MDB_cursor *mc, *m2;
8017 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8020 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8023 rc = mdb_cursor_open(txn, dbi, &mc);
8027 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8028 /* Invalidate the dropped DB's cursors */
8029 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8030 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8034 /* Can't delete the main DB */
8035 if (del && dbi > MAIN_DBI) {
8036 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8038 txn->mt_dbflags[dbi] = DB_STALE;
8039 mdb_dbi_close(txn->mt_env, dbi);
8042 /* reset the DB record, mark it dirty */
8043 txn->mt_dbflags[dbi] |= DB_DIRTY;
8044 txn->mt_dbs[dbi].md_depth = 0;
8045 txn->mt_dbs[dbi].md_branch_pages = 0;
8046 txn->mt_dbs[dbi].md_leaf_pages = 0;
8047 txn->mt_dbs[dbi].md_overflow_pages = 0;
8048 txn->mt_dbs[dbi].md_entries = 0;
8049 txn->mt_dbs[dbi].md_root = P_INVALID;
8051 txn->mt_flags |= MDB_TXN_DIRTY;
8054 mdb_cursor_close(mc);
8058 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8060 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8063 txn->mt_dbxs[dbi].md_cmp = cmp;
8067 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8069 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8072 txn->mt_dbxs[dbi].md_dcmp = cmp;
8076 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8078 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8081 txn->mt_dbxs[dbi].md_rel = rel;
8085 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8087 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8090 txn->mt_dbxs[dbi].md_relctx = ctx;
8094 int mdb_env_get_maxkeysize(MDB_env *env)
8096 return MDB_MAXKEYSIZE;
8099 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8101 unsigned int i, rdrs;
8108 if (!env->me_txns) {
8109 return func("(no reader locks)\n", ctx);
8111 rdrs = env->me_txns->mti_numreaders;
8112 mr = env->me_txns->mti_readers;
8113 for (i=0; i<rdrs; i++) {
8118 if (mr[i].mr_txnid == (txnid_t)-1) {
8119 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8121 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8125 func(" pid thread txnid\n", ctx);
8127 rc = func(buf, ctx);
8133 func("(no active readers)\n", ctx);
8138 /* insert pid into list if not already present.
8139 * return -1 if already present.
8141 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8143 /* binary search of pid in list */
8145 unsigned cursor = 1;
8147 unsigned n = ids[0];
8150 unsigned pivot = n >> 1;
8151 cursor = base + pivot + 1;
8152 val = pid - ids[cursor];
8157 } else if ( val > 0 ) {
8162 /* found, so it's a duplicate */
8171 for (n = ids[0]; n > cursor; n--)
8177 int mdb_reader_check(MDB_env *env, int *dead)
8179 unsigned int i, j, rdrs;
8190 rdrs = env->me_txns->mti_numreaders;
8191 pids = malloc((rdrs+1) * sizeof(pid_t));
8195 mr = env->me_txns->mti_readers;
8197 for (i=0; i<rdrs; i++) {
8198 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8200 if (mdb_pid_insert(pids, pid) == 0) {
8201 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8203 /* Recheck, a new process may have reused pid */
8204 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8205 for (j=i; j<rdrs; j++)
8206 if (mr[j].mr_pid == pid) {
8211 UNLOCK_MUTEX_R(env);