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 #define unlink(file) DeleteFile(file)
167 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
168 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
170 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
177 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
178 #define MDB_PIDLOCK 1
180 #ifdef MDB_USE_POSIX_SEM
182 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
183 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
184 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
185 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
188 mdb_sem_wait(sem_t *sem)
191 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
196 /** Lock the reader mutex.
198 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
199 /** Unlock the reader mutex.
201 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
203 /** Lock the writer mutex.
204 * Only a single write transaction is allowed at a time. Other writers
205 * will block waiting for this mutex.
207 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
208 /** Unlock the writer mutex.
210 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
211 #endif /* MDB_USE_POSIX_SEM */
213 /** Get the error code for the last failed system function.
215 #define ErrCode() errno
217 /** An abstraction for a file handle.
218 * On POSIX systems file handles are small integers. On Windows
219 * they're opaque pointers.
223 /** A value for an invalid file handle.
224 * Mainly used to initialize file variables and signify that they are
227 #define INVALID_HANDLE_VALUE (-1)
229 /** Get the size of a memory page for the system.
230 * This is the basic size that the platform's memory manager uses, and is
231 * fundamental to the use of memory-mapped files.
233 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
236 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
239 #define MNAME_LEN (sizeof(pthread_mutex_t))
245 /** A flag for opening a file and requesting synchronous data writes.
246 * This is only used when writing a meta page. It's not strictly needed;
247 * we could just do a normal write and then immediately perform a flush.
248 * But if this flag is available it saves us an extra system call.
250 * @note If O_DSYNC is undefined but exists in /usr/include,
251 * preferably set some compiler flag to get the definition.
252 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
255 # define MDB_DSYNC O_DSYNC
259 /** Function for flushing the data of a file. Define this to fsync
260 * if fdatasync() is not supported.
262 #ifndef MDB_FDATASYNC
263 # define MDB_FDATASYNC fdatasync
267 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
278 /** A page number in the database.
279 * Note that 64 bit page numbers are overkill, since pages themselves
280 * already represent 12-13 bits of addressable memory, and the OS will
281 * always limit applications to a maximum of 63 bits of address space.
283 * @note In the #MDB_node structure, we only store 48 bits of this value,
284 * which thus limits us to only 60 bits of addressable data.
286 typedef MDB_ID pgno_t;
288 /** A transaction ID.
289 * See struct MDB_txn.mt_txnid for details.
291 typedef MDB_ID txnid_t;
293 /** @defgroup debug Debug Macros
297 /** Enable debug output. Needs variable argument macros (a C99 feature).
298 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
299 * read from and written to the database (used for free space management).
305 static int mdb_debug;
306 static txnid_t mdb_debug_start;
308 /** Print a debug message with printf formatting.
309 * Requires double parenthesis around 2 or more args.
311 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
312 # define DPRINTF0(fmt, ...) \
313 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
315 # define DPRINTF(args) ((void) 0)
317 /** Print a debug string.
318 * The string is printed literally, with no format processing.
320 #define DPUTS(arg) DPRINTF(("%s", arg))
323 /** A default memory page size.
324 * The actual size is platform-dependent, but we use this for
325 * boot-strapping. We probably should not be using this any more.
326 * The #GET_PAGESIZE() macro is used to get the actual size.
328 * Note that we don't currently support Huge pages. On Linux,
329 * regular data files cannot use Huge pages, and in general
330 * Huge pages aren't actually pageable. We rely on the OS
331 * demand-pager to read our data and page it out when memory
332 * pressure from other processes is high. So until OSs have
333 * actual paging support for Huge pages, they're not viable.
335 #define MDB_PAGESIZE 4096
337 /** The minimum number of keys required in a database page.
338 * Setting this to a larger value will place a smaller bound on the
339 * maximum size of a data item. Data items larger than this size will
340 * be pushed into overflow pages instead of being stored directly in
341 * the B-tree node. This value used to default to 4. With a page size
342 * of 4096 bytes that meant that any item larger than 1024 bytes would
343 * go into an overflow page. That also meant that on average 2-3KB of
344 * each overflow page was wasted space. The value cannot be lower than
345 * 2 because then there would no longer be a tree structure. With this
346 * value, items larger than 2KB will go into overflow pages, and on
347 * average only 1KB will be wasted.
349 #define MDB_MINKEYS 2
351 /** A stamp that identifies a file as an MDB file.
352 * There's nothing special about this value other than that it is easily
353 * recognizable, and it will reflect any byte order mismatches.
355 #define MDB_MAGIC 0xBEEFC0DE
357 /** The version number for a database's datafile format. */
358 #define MDB_DATA_VERSION 1
359 /** The version number for a database's lockfile format. */
360 #define MDB_LOCK_VERSION 1
362 /** @brief The maximum size of a key in the database.
364 * The library rejects bigger keys, and cannot deal with records
365 * with bigger keys stored by a library with bigger max keysize.
367 * We require that keys all fit onto a regular page. This limit
368 * could be raised a bit further if needed; to something just
369 * under #MDB_PAGESIZE / #MDB_MINKEYS.
371 * Note that data items in an #MDB_DUPSORT database are actually keys
372 * of a subDB, so they're also limited to this size.
374 #ifndef MDB_MAXKEYSIZE
375 #define MDB_MAXKEYSIZE 511
378 /** @brief The maximum size of a data item.
380 * We only store a 32 bit value for node sizes.
382 #define MAXDATASIZE 0xffffffffUL
387 * This is used for printing a hex dump of a key's contents.
389 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
390 /** Display a key in hex.
392 * Invoke a function to display a key in hex.
394 #define DKEY(x) mdb_dkey(x, kbuf)
396 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
400 /** An invalid page number.
401 * Mainly used to denote an empty tree.
403 #define P_INVALID (~(pgno_t)0)
405 /** Test if the flags \b f are set in a flag word \b w. */
406 #define F_ISSET(w, f) (((w) & (f)) == (f))
408 /** Used for offsets within a single page.
409 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
412 typedef uint16_t indx_t;
414 /** Default size of memory map.
415 * This is certainly too small for any actual applications. Apps should always set
416 * the size explicitly using #mdb_env_set_mapsize().
418 #define DEFAULT_MAPSIZE 1048576
420 /** @defgroup readers Reader Lock Table
421 * Readers don't acquire any locks for their data access. Instead, they
422 * simply record their transaction ID in the reader table. The reader
423 * mutex is needed just to find an empty slot in the reader table. The
424 * slot's address is saved in thread-specific data so that subsequent read
425 * transactions started by the same thread need no further locking to proceed.
427 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
429 * No reader table is used if the database is on a read-only filesystem.
431 * Since the database uses multi-version concurrency control, readers don't
432 * actually need any locking. This table is used to keep track of which
433 * readers are using data from which old transactions, so that we'll know
434 * when a particular old transaction is no longer in use. Old transactions
435 * that have discarded any data pages can then have those pages reclaimed
436 * for use by a later write transaction.
438 * The lock table is constructed such that reader slots are aligned with the
439 * processor's cache line size. Any slot is only ever used by one thread.
440 * This alignment guarantees that there will be no contention or cache
441 * thrashing as threads update their own slot info, and also eliminates
442 * any need for locking when accessing a slot.
444 * A writer thread will scan every slot in the table to determine the oldest
445 * outstanding reader transaction. Any freed pages older than this will be
446 * reclaimed by the writer. The writer doesn't use any locks when scanning
447 * this table. This means that there's no guarantee that the writer will
448 * see the most up-to-date reader info, but that's not required for correct
449 * operation - all we need is to know the upper bound on the oldest reader,
450 * we don't care at all about the newest reader. So the only consequence of
451 * reading stale information here is that old pages might hang around a
452 * while longer before being reclaimed. That's actually good anyway, because
453 * the longer we delay reclaiming old pages, the more likely it is that a
454 * string of contiguous pages can be found after coalescing old pages from
455 * many old transactions together.
458 /** Number of slots in the reader table.
459 * This value was chosen somewhat arbitrarily. 126 readers plus a
460 * couple mutexes fit exactly into 8KB on my development machine.
461 * Applications should set the table size using #mdb_env_set_maxreaders().
463 #define DEFAULT_READERS 126
465 /** The size of a CPU cache line in bytes. We want our lock structures
466 * aligned to this size to avoid false cache line sharing in the
468 * This value works for most CPUs. For Itanium this should be 128.
474 /** The information we store in a single slot of the reader table.
475 * In addition to a transaction ID, we also record the process and
476 * thread ID that owns a slot, so that we can detect stale information,
477 * e.g. threads or processes that went away without cleaning up.
478 * @note We currently don't check for stale records. We simply re-init
479 * the table when we know that we're the only process opening the
482 typedef struct MDB_rxbody {
483 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
484 * Multiple readers that start at the same time will probably have the
485 * same ID here. Again, it's not important to exclude them from
486 * anything; all we need to know is which version of the DB they
487 * started from so we can avoid overwriting any data used in that
488 * particular version.
491 /** The process ID of the process owning this reader txn. */
493 /** The thread ID of the thread owning this txn. */
497 /** The actual reader record, with cacheline padding. */
498 typedef struct MDB_reader {
501 /** shorthand for mrb_txnid */
502 #define mr_txnid mru.mrx.mrb_txnid
503 #define mr_pid mru.mrx.mrb_pid
504 #define mr_tid mru.mrx.mrb_tid
505 /** cache line alignment */
506 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
510 /** The header for the reader table.
511 * The table resides in a memory-mapped file. (This is a different file
512 * than is used for the main database.)
514 * For POSIX the actual mutexes reside in the shared memory of this
515 * mapped file. On Windows, mutexes are named objects allocated by the
516 * kernel; we store the mutex names in this mapped file so that other
517 * processes can grab them. This same approach is also used on
518 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
519 * process-shared POSIX mutexes. For these cases where a named object
520 * is used, the object name is derived from a 64 bit FNV hash of the
521 * environment pathname. As such, naming collisions are extremely
522 * unlikely. If a collision occurs, the results are unpredictable.
524 typedef struct MDB_txbody {
525 /** Stamp identifying this as an MDB file. It must be set
528 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
530 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
531 char mtb_rmname[MNAME_LEN];
533 /** Mutex protecting access to this table.
534 * This is the reader lock that #LOCK_MUTEX_R acquires.
536 pthread_mutex_t mtb_mutex;
538 /** The ID of the last transaction committed to the database.
539 * This is recorded here only for convenience; the value can always
540 * be determined by reading the main database meta pages.
543 /** The number of slots that have been used in the reader table.
544 * This always records the maximum count, it is not decremented
545 * when readers release their slots.
547 unsigned mtb_numreaders;
550 /** The actual reader table definition. */
551 typedef struct MDB_txninfo {
554 #define mti_magic mt1.mtb.mtb_magic
555 #define mti_format mt1.mtb.mtb_format
556 #define mti_mutex mt1.mtb.mtb_mutex
557 #define mti_rmname mt1.mtb.mtb_rmname
558 #define mti_txnid mt1.mtb.mtb_txnid
559 #define mti_numreaders mt1.mtb.mtb_numreaders
560 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
563 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
564 char mt2_wmname[MNAME_LEN];
565 #define mti_wmname mt2.mt2_wmname
567 pthread_mutex_t mt2_wmutex;
568 #define mti_wmutex mt2.mt2_wmutex
570 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
572 MDB_reader mti_readers[1];
575 /** Lockfile format signature: version, features and field layout */
576 #define MDB_LOCK_FORMAT \
578 ((MDB_LOCK_VERSION) \
579 /* Flags which describe functionality */ \
580 + (((MDB_PIDLOCK) != 0) << 16)))
583 /** Common header for all page types.
584 * Overflow records occupy a number of contiguous pages with no
585 * headers on any page after the first.
587 typedef struct MDB_page {
588 #define mp_pgno mp_p.p_pgno
589 #define mp_next mp_p.p_next
591 pgno_t p_pgno; /**< page number */
592 void * p_next; /**< for in-memory list of freed structs */
595 /** @defgroup mdb_page Page Flags
597 * Flags for the page headers.
600 #define P_BRANCH 0x01 /**< branch page */
601 #define P_LEAF 0x02 /**< leaf page */
602 #define P_OVERFLOW 0x04 /**< overflow page */
603 #define P_META 0x08 /**< meta page */
604 #define P_DIRTY 0x10 /**< dirty page */
605 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
606 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
607 #define P_KEEP 0x8000 /**< leave this page alone during spill */
609 uint16_t mp_flags; /**< @ref mdb_page */
610 #define mp_lower mp_pb.pb.pb_lower
611 #define mp_upper mp_pb.pb.pb_upper
612 #define mp_pages mp_pb.pb_pages
615 indx_t pb_lower; /**< lower bound of free space */
616 indx_t pb_upper; /**< upper bound of free space */
618 uint32_t pb_pages; /**< number of overflow pages */
620 indx_t mp_ptrs[1]; /**< dynamic size */
623 /** Size of the page header, excluding dynamic data at the end */
624 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
626 /** Address of first usable data byte in a page, after the header */
627 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
629 /** Number of nodes on a page */
630 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
632 /** The amount of space remaining in the page */
633 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
635 /** The percentage of space used in the page, in tenths of a percent. */
636 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
637 ((env)->me_psize - PAGEHDRSZ))
638 /** The minimum page fill factor, in tenths of a percent.
639 * Pages emptier than this are candidates for merging.
641 #define FILL_THRESHOLD 250
643 /** Test if a page is a leaf page */
644 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
645 /** Test if a page is a LEAF2 page */
646 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
647 /** Test if a page is a branch page */
648 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
649 /** Test if a page is an overflow page */
650 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
651 /** Test if a page is a sub page */
652 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
654 /** The number of overflow pages needed to store the given size. */
655 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
657 /** Header for a single key/data pair within a page.
658 * We guarantee 2-byte alignment for nodes.
660 typedef struct MDB_node {
661 /** lo and hi are used for data size on leaf nodes and for
662 * child pgno on branch nodes. On 64 bit platforms, flags
663 * is also used for pgno. (Branch nodes have no flags).
664 * They are in host byte order in case that lets some
665 * accesses be optimized into a 32-bit word access.
667 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
668 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
669 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
670 /** @defgroup mdb_node Node Flags
672 * Flags for node headers.
675 #define F_BIGDATA 0x01 /**< data put on overflow page */
676 #define F_SUBDATA 0x02 /**< data is a sub-database */
677 #define F_DUPDATA 0x04 /**< data has duplicates */
679 /** valid flags for #mdb_node_add() */
680 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
683 unsigned short mn_flags; /**< @ref mdb_node */
684 unsigned short mn_ksize; /**< key size */
685 char mn_data[1]; /**< key and data are appended here */
688 /** Size of the node header, excluding dynamic data at the end */
689 #define NODESIZE offsetof(MDB_node, mn_data)
691 /** Bit position of top word in page number, for shifting mn_flags */
692 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
694 /** Size of a node in a branch page with a given key.
695 * This is just the node header plus the key, there is no data.
697 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
699 /** Size of a node in a leaf page with a given key and data.
700 * This is node header plus key plus data size.
702 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
704 /** Address of node \b i in page \b p */
705 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
707 /** Address of the key for the node */
708 #define NODEKEY(node) (void *)((node)->mn_data)
710 /** Address of the data for a node */
711 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
713 /** Get the page number pointed to by a branch node */
714 #define NODEPGNO(node) \
715 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
716 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
717 /** Set the page number in a branch node */
718 #define SETPGNO(node,pgno) do { \
719 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
720 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
722 /** Get the size of the data in a leaf node */
723 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
724 /** Set the size of the data for a leaf node */
725 #define SETDSZ(node,size) do { \
726 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
727 /** The size of a key in a node */
728 #define NODEKSZ(node) ((node)->mn_ksize)
730 /** Copy a page number from src to dst */
732 #define COPY_PGNO(dst,src) dst = src
734 #if SIZE_MAX > 4294967295UL
735 #define COPY_PGNO(dst,src) do { \
736 unsigned short *s, *d; \
737 s = (unsigned short *)&(src); \
738 d = (unsigned short *)&(dst); \
745 #define COPY_PGNO(dst,src) do { \
746 unsigned short *s, *d; \
747 s = (unsigned short *)&(src); \
748 d = (unsigned short *)&(dst); \
754 /** The address of a key in a LEAF2 page.
755 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
756 * There are no node headers, keys are stored contiguously.
758 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
760 /** Set the \b node's key into \b key, if requested. */
761 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
762 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
764 /** Information about a single database in the environment. */
765 typedef struct MDB_db {
766 uint32_t md_pad; /**< also ksize for LEAF2 pages */
767 uint16_t md_flags; /**< @ref mdb_dbi_open */
768 uint16_t md_depth; /**< depth of this tree */
769 pgno_t md_branch_pages; /**< number of internal pages */
770 pgno_t md_leaf_pages; /**< number of leaf pages */
771 pgno_t md_overflow_pages; /**< number of overflow pages */
772 size_t md_entries; /**< number of data items */
773 pgno_t md_root; /**< the root page of this tree */
776 /** mdb_dbi_open flags */
777 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
778 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
779 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
780 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
782 /** Handle for the DB used to track free pages. */
784 /** Handle for the default DB. */
787 /** Meta page content. */
788 typedef struct MDB_meta {
789 /** Stamp identifying this as an MDB file. It must be set
792 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
794 void *mm_address; /**< address for fixed mapping */
795 size_t mm_mapsize; /**< size of mmap region */
796 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
797 /** The size of pages used in this DB */
798 #define mm_psize mm_dbs[0].md_pad
799 /** Any persistent environment flags. @ref mdb_env */
800 #define mm_flags mm_dbs[0].md_flags
801 pgno_t mm_last_pg; /**< last used page in file */
802 txnid_t mm_txnid; /**< txnid that committed this page */
805 /** Buffer for a stack-allocated dirty page.
806 * The members define size and alignment, and silence type
807 * aliasing warnings. They are not used directly; that could
808 * mean incorrectly using several union members in parallel.
810 typedef union MDB_pagebuf {
811 char mb_raw[MDB_PAGESIZE];
814 char mm_pad[PAGEHDRSZ];
819 /** Auxiliary DB info.
820 * The information here is mostly static/read-only. There is
821 * only a single copy of this record in the environment.
823 typedef struct MDB_dbx {
824 MDB_val md_name; /**< name of the database */
825 MDB_cmp_func *md_cmp; /**< function for comparing keys */
826 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
827 MDB_rel_func *md_rel; /**< user relocate function */
828 void *md_relctx; /**< user-provided context for md_rel */
831 /** A database transaction.
832 * Every operation requires a transaction handle.
835 MDB_txn *mt_parent; /**< parent of a nested txn */
836 MDB_txn *mt_child; /**< nested txn under this txn */
837 pgno_t mt_next_pgno; /**< next unallocated page */
838 /** The ID of this transaction. IDs are integers incrementing from 1.
839 * Only committed write transactions increment the ID. If a transaction
840 * aborts, the ID may be re-used by the next writer.
843 MDB_env *mt_env; /**< the DB environment */
844 /** The list of pages that became unused during this transaction.
847 /** The sorted list of dirty pages we temporarily wrote to disk
848 * because the dirty list was full.
850 MDB_IDL mt_spill_pgs;
852 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
854 /** For read txns: This thread/txn's reader table slot, or NULL. */
857 /** Array of records for each DB known in the environment. */
859 /** Array of MDB_db records for each known DB */
861 /** @defgroup mt_dbflag Transaction DB Flags
865 #define DB_DIRTY 0x01 /**< DB was written in this txn */
866 #define DB_STALE 0x02 /**< DB record is older than txnID */
867 #define DB_NEW 0x04 /**< DB handle opened in this txn */
868 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
870 /** In write txns, array of cursors for each DB */
871 MDB_cursor **mt_cursors;
872 /** Array of flags for each DB */
873 unsigned char *mt_dbflags;
874 /** Number of DB records in use. This number only ever increments;
875 * we don't decrement it when individual DB handles are closed.
879 /** @defgroup mdb_txn Transaction Flags
883 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
884 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
885 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
886 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
888 unsigned int mt_flags; /**< @ref mdb_txn */
889 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
890 unsigned int mt_dirty_room;
891 /** Tracks which of the two meta pages was used at the start
892 * of this transaction.
894 unsigned int mt_toggle;
897 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
898 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
899 * raise this on a 64 bit machine.
901 #define CURSOR_STACK 32
905 /** Cursors are used for all DB operations */
907 /** Next cursor on this DB in this txn */
909 /** Backup of the original cursor if this cursor is a shadow */
910 MDB_cursor *mc_backup;
911 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
912 struct MDB_xcursor *mc_xcursor;
913 /** The transaction that owns this cursor */
915 /** The database handle this cursor operates on */
917 /** The database record for this cursor */
919 /** The database auxiliary record for this cursor */
921 /** The @ref mt_dbflag for this database */
922 unsigned char *mc_dbflag;
923 unsigned short mc_snum; /**< number of pushed pages */
924 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
925 /** @defgroup mdb_cursor Cursor Flags
927 * Cursor state flags.
930 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
931 #define C_EOF 0x02 /**< No more data */
932 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
933 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
934 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
936 unsigned int mc_flags; /**< @ref mdb_cursor */
937 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
938 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
941 /** Context for sorted-dup records.
942 * We could have gone to a fully recursive design, with arbitrarily
943 * deep nesting of sub-databases. But for now we only handle these
944 * levels - main DB, optional sub-DB, sorted-duplicate DB.
946 typedef struct MDB_xcursor {
947 /** A sub-cursor for traversing the Dup DB */
948 MDB_cursor mx_cursor;
949 /** The database record for this Dup DB */
951 /** The auxiliary DB record for this Dup DB */
953 /** The @ref mt_dbflag for this Dup DB */
954 unsigned char mx_dbflag;
957 /** State of FreeDB old pages, stored in the MDB_env */
958 typedef struct MDB_pgstate {
959 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
960 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
963 /** The database environment. */
965 HANDLE me_fd; /**< The main data file */
966 HANDLE me_lfd; /**< The lock file */
967 HANDLE me_mfd; /**< just for writing the meta pages */
968 /** Failed to update the meta page. Probably an I/O error. */
969 #define MDB_FATAL_ERROR 0x80000000U
970 /** Some fields are initialized. */
971 #define MDB_ENV_ACTIVE 0x20000000U
972 /** me_txkey is set */
973 #define MDB_ENV_TXKEY 0x10000000U
974 /** Have liveness lock in reader table */
975 #define MDB_LIVE_READER 0x08000000U
976 uint32_t me_flags; /**< @ref mdb_env */
977 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
978 unsigned int me_maxreaders; /**< size of the reader table */
979 unsigned int me_numreaders; /**< max numreaders set by this env */
980 MDB_dbi me_numdbs; /**< number of DBs opened */
981 MDB_dbi me_maxdbs; /**< size of the DB table */
982 pid_t me_pid; /**< process ID of this env */
983 char *me_path; /**< path to the DB files */
984 char *me_map; /**< the memory map of the data file */
985 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
986 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
987 MDB_txn *me_txn; /**< current write transaction */
988 size_t me_mapsize; /**< size of the data memory map */
989 off_t me_size; /**< current file size */
990 pgno_t me_maxpg; /**< me_mapsize / me_psize */
991 MDB_dbx *me_dbxs; /**< array of static DB info */
992 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
993 pthread_key_t me_txkey; /**< thread-key for readers */
994 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
995 # define me_pglast me_pgstate.mf_pglast
996 # define me_pghead me_pgstate.mf_pghead
997 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
998 /** IDL of pages that became unused in a write txn */
1000 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1001 MDB_ID2L me_dirty_list;
1002 /** Max number of freelist items that can fit in a single overflow page */
1004 /** Max size of a node on a page */
1005 unsigned int me_nodemax;
1007 int me_pidquery; /**< Used in OpenProcess */
1008 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1010 #elif defined(MDB_USE_POSIX_SEM)
1011 sem_t *me_rmutex; /* Shared mutexes are not supported */
1016 /** Nested transaction */
1017 typedef struct MDB_ntxn {
1018 MDB_txn mnt_txn; /* the transaction */
1019 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
1022 /** max number of pages to commit in one writev() call */
1023 #define MDB_COMMIT_PAGES 64
1024 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1025 #undef MDB_COMMIT_PAGES
1026 #define MDB_COMMIT_PAGES IOV_MAX
1029 /* max bytes to write in one call */
1030 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1032 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1033 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1034 static int mdb_page_touch(MDB_cursor *mc);
1036 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1037 static int mdb_page_search_root(MDB_cursor *mc,
1038 MDB_val *key, int modify);
1039 #define MDB_PS_MODIFY 1
1040 #define MDB_PS_ROOTONLY 2
1041 static int mdb_page_search(MDB_cursor *mc,
1042 MDB_val *key, int flags);
1043 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1045 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1046 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1047 pgno_t newpgno, unsigned int nflags);
1049 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1050 static int mdb_env_pick_meta(const MDB_env *env);
1051 static int mdb_env_write_meta(MDB_txn *txn);
1052 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1053 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1055 static void mdb_env_close0(MDB_env *env, int excl);
1057 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1058 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1059 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1060 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1061 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1062 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1063 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1064 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1065 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1067 static int mdb_rebalance(MDB_cursor *mc);
1068 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1070 static void mdb_cursor_pop(MDB_cursor *mc);
1071 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1073 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1074 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1075 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1076 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1077 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1079 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1080 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1082 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1083 static void mdb_xcursor_init0(MDB_cursor *mc);
1084 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1086 static int mdb_drop0(MDB_cursor *mc, int subs);
1087 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1090 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1094 static SECURITY_DESCRIPTOR mdb_null_sd;
1095 static SECURITY_ATTRIBUTES mdb_all_sa;
1096 static int mdb_sec_inited;
1099 /** Return the library version info. */
1101 mdb_version(int *major, int *minor, int *patch)
1103 if (major) *major = MDB_VERSION_MAJOR;
1104 if (minor) *minor = MDB_VERSION_MINOR;
1105 if (patch) *patch = MDB_VERSION_PATCH;
1106 return MDB_VERSION_STRING;
1109 /** Table of descriptions for MDB @ref errors */
1110 static char *const mdb_errstr[] = {
1111 "MDB_KEYEXIST: Key/data pair already exists",
1112 "MDB_NOTFOUND: No matching key/data pair found",
1113 "MDB_PAGE_NOTFOUND: Requested page not found",
1114 "MDB_CORRUPTED: Located page was wrong type",
1115 "MDB_PANIC: Update of meta page failed",
1116 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1117 "MDB_INVALID: File is not an MDB file",
1118 "MDB_MAP_FULL: Environment mapsize limit reached",
1119 "MDB_DBS_FULL: Environment maxdbs limit reached",
1120 "MDB_READERS_FULL: Environment maxreaders limit reached",
1121 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1122 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1123 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1124 "MDB_PAGE_FULL: Internal error - page has no more space",
1125 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1126 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1127 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1128 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1129 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1133 mdb_strerror(int err)
1137 return ("Successful return: 0");
1139 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1140 i = err - MDB_KEYEXIST;
1141 return mdb_errstr[i];
1144 return strerror(err);
1148 /** Display a key in hexadecimal and return the address of the result.
1149 * @param[in] key the key to display
1150 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1151 * @return The key in hexadecimal form.
1154 mdb_dkey(MDB_val *key, char *buf)
1157 unsigned char *c = key->mv_data;
1163 if (key->mv_size > MDB_MAXKEYSIZE)
1164 return "MDB_MAXKEYSIZE";
1165 /* may want to make this a dynamic check: if the key is mostly
1166 * printable characters, print it as-is instead of converting to hex.
1170 for (i=0; i<key->mv_size; i++)
1171 ptr += sprintf(ptr, "%02x", *c++);
1173 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1178 /** Display all the keys in the page. */
1180 mdb_page_list(MDB_page *mp)
1183 unsigned int i, nkeys, nsize;
1187 nkeys = NUMKEYS(mp);
1188 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1189 for (i=0; i<nkeys; i++) {
1190 node = NODEPTR(mp, i);
1191 key.mv_size = node->mn_ksize;
1192 key.mv_data = node->mn_data;
1193 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1194 if (IS_BRANCH(mp)) {
1195 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1198 if (F_ISSET(node->mn_flags, F_BIGDATA))
1199 nsize += sizeof(pgno_t);
1201 nsize += NODEDSZ(node);
1202 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1208 mdb_cursor_chk(MDB_cursor *mc)
1214 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1215 for (i=0; i<mc->mc_top; i++) {
1217 node = NODEPTR(mp, mc->mc_ki[i]);
1218 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1221 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1227 /** Count all the pages in each DB and in the freelist
1228 * and make sure it matches the actual number of pages
1231 static void mdb_audit(MDB_txn *txn)
1235 MDB_ID freecount, count;
1240 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1241 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1242 freecount += *(MDB_ID *)data.mv_data;
1245 for (i = 0; i<txn->mt_numdbs; i++) {
1247 mdb_cursor_init(&mc, txn, i, &mx);
1248 if (txn->mt_dbs[i].md_root == P_INVALID)
1250 count += txn->mt_dbs[i].md_branch_pages +
1251 txn->mt_dbs[i].md_leaf_pages +
1252 txn->mt_dbs[i].md_overflow_pages;
1253 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1254 mdb_page_search(&mc, NULL, 0);
1258 mp = mc.mc_pg[mc.mc_top];
1259 for (j=0; j<NUMKEYS(mp); j++) {
1260 MDB_node *leaf = NODEPTR(mp, j);
1261 if (leaf->mn_flags & F_SUBDATA) {
1263 memcpy(&db, NODEDATA(leaf), sizeof(db));
1264 count += db.md_branch_pages + db.md_leaf_pages +
1265 db.md_overflow_pages;
1269 while (mdb_cursor_sibling(&mc, 1) == 0);
1272 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1273 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1274 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1280 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1282 return txn->mt_dbxs[dbi].md_cmp(a, b);
1286 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1288 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1291 /** Allocate memory for a page.
1292 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1295 mdb_page_malloc(MDB_txn *txn, unsigned num)
1297 MDB_env *env = txn->mt_env;
1298 MDB_page *ret = env->me_dpages;
1299 size_t sz = env->me_psize;
1302 VGMEMP_ALLOC(env, ret, sz);
1303 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1304 env->me_dpages = ret->mp_next;
1310 if ((ret = malloc(sz)) != NULL) {
1311 VGMEMP_ALLOC(env, ret, sz);
1316 /** Free a single page.
1317 * Saves single pages to a list, for future reuse.
1318 * (This is not used for multi-page overflow pages.)
1321 mdb_page_free(MDB_env *env, MDB_page *mp)
1323 mp->mp_next = env->me_dpages;
1324 VGMEMP_FREE(env, mp);
1325 env->me_dpages = mp;
1328 /* Free a dirty page */
1330 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1332 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1333 mdb_page_free(env, dp);
1335 /* large pages just get freed directly */
1336 VGMEMP_FREE(env, dp);
1341 /** Return all dirty pages to dpage list */
1343 mdb_dlist_free(MDB_txn *txn)
1345 MDB_env *env = txn->mt_env;
1346 MDB_ID2L dl = txn->mt_u.dirty_list;
1347 unsigned i, n = dl[0].mid;
1349 for (i = 1; i <= n; i++) {
1350 mdb_dpage_free(env, dl[i].mptr);
1355 /* Set or clear P_KEEP in non-overflow, non-sub pages in this txn's cursors.
1356 * @param[in] mc A cursor handle for the current operation.
1357 * @param[in] pflags Flags of the pages to update:
1358 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1361 mdb_cursorpages_mark(MDB_cursor *mc, unsigned pflags)
1363 MDB_txn *txn = mc->mc_txn;
1368 if (mc->mc_flags & C_UNTRACK)
1369 mc = NULL; /* will find mc in mt_cursors */
1370 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1371 for (; mc; mc=mc->mc_next) {
1372 for (m3 = mc; m3->mc_flags & C_INITIALIZED; m3 = &mx->mx_cursor) {
1373 for (j=0; j<m3->mc_snum; j++)
1374 if ((m3->mc_pg[j]->mp_flags & (P_SUBP|P_DIRTY|P_KEEP))
1376 m3->mc_pg[j]->mp_flags ^= P_KEEP;
1377 mx = m3->mc_xcursor;
1387 static int mdb_page_flush(MDB_txn *txn);
1389 /** Spill pages from the dirty list back to disk.
1390 * This is intended to prevent running into #MDB_TXN_FULL situations,
1391 * but note that they may still occur in a few cases:
1392 * 1) pages in #MDB_DUPSORT sub-DBs are never spilled, so if there
1393 * are too many of these dirtied in one txn, the txn may still get
1395 * 2) child txns may run out of space if their parents dirtied a
1396 * lot of pages and never spilled them. TODO: we probably should do
1397 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1398 * the parent's dirty_room is below a given threshold.
1399 * 3) our estimate of the txn size could be too small. At the
1400 * moment this seems unlikely.
1402 * Otherwise, if not using nested txns, it is expected that apps will
1403 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1404 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1405 * If the txn never references them again, they can be left alone.
1406 * If the txn only reads them, they can be used without any fuss.
1407 * If the txn writes them again, they can be dirtied immediately without
1408 * going thru all of the work of #mdb_page_touch(). Such references are
1409 * handled by #mdb_page_unspill().
1411 * Also note, we never spill DB root pages, nor pages of active cursors,
1412 * because we'll need these back again soon anyway. And in nested txns,
1413 * we can't spill a page in a child txn if it was already spilled in a
1414 * parent txn. That would alter the parent txns' data even though
1415 * the child hasn't committed yet, and we'd have no way to undo it if
1416 * the child aborted.
1418 * @param[in] m0 cursor A cursor handle identifying the transaction and
1419 * database for which we are checking space.
1420 * @param[in] key For a put operation, the key being stored.
1421 * @param[in] data For a put operation, the data being stored.
1422 * @return 0 on success, non-zero on failure.
1425 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1427 MDB_txn *txn = m0->mc_txn;
1429 MDB_ID2L dl = txn->mt_u.dirty_list;
1433 if (m0->mc_flags & C_SUB)
1436 /* Estimate how much space this op will take */
1437 i = m0->mc_db->md_depth;
1438 /* Named DBs also dirty the main DB */
1439 if (m0->mc_dbi > MAIN_DBI)
1440 i += txn->mt_dbs[MAIN_DBI].md_depth;
1441 /* For puts, roughly factor in the key+data size */
1443 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1444 i += i; /* double it for good measure */
1446 if (txn->mt_dirty_room > i)
1449 if (!txn->mt_spill_pgs) {
1450 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1451 if (!txn->mt_spill_pgs)
1455 /* Mark all the dirty root pages we want to preserve */
1456 for (i=0; i<txn->mt_numdbs; i++) {
1457 if (txn->mt_dbflags[i] & DB_DIRTY) {
1458 pgno_t pgno = txn->mt_dbs[i].md_root;
1459 if (pgno == P_INVALID)
1461 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1463 if ((dp->mp_flags & P_DIRTY) && level <= 1)
1464 dp->mp_flags |= P_KEEP;
1468 /* Preserve pages used by cursors */
1469 mdb_cursorpages_mark(m0, P_DIRTY);
1471 /* Save the page IDs of all the pages we're flushing */
1472 for (i=1; i<=dl[0].mid; i++) {
1474 if (dp->mp_flags & P_KEEP)
1476 /* Can't spill twice, make sure it's not already in a parent's
1479 if (txn->mt_parent) {
1481 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1482 if (tx2->mt_spill_pgs) {
1483 j = mdb_midl_search(tx2->mt_spill_pgs, dl[i].mid);
1484 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == dl[i].mid) {
1485 dp->mp_flags |= P_KEEP;
1493 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, dl[i].mid)))
1496 mdb_midl_sort(txn->mt_spill_pgs);
1498 rc = mdb_page_flush(txn);
1500 mdb_cursorpages_mark(m0, P_DIRTY|P_KEEP);
1504 if (txn->mt_parent) {
1506 pgno_t pgno = dl[i].mid;
1507 txn->mt_dirty_room = txn->mt_parent->mt_dirty_room - dl[0].mid;
1508 /* dirty pages that are dirty in an ancestor don't
1509 * count against this txn's dirty_room.
1511 for (i=1; i<=dl[0].mid; i++) {
1512 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1513 j = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1514 if (j <= tx2->mt_u.dirty_list[0].mid &&
1515 tx2->mt_u.dirty_list[j].mid == pgno) {
1516 txn->mt_dirty_room++;
1522 txn->mt_dirty_room = MDB_IDL_UM_MAX - dl[0].mid;
1524 txn->mt_flags |= MDB_TXN_SPILLS;
1526 txn->mt_flags |= MDB_TXN_ERROR;
1531 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1533 mdb_find_oldest(MDB_txn *txn)
1536 txnid_t mr, oldest = txn->mt_txnid - 1;
1537 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1538 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1548 /** Add a page to the txn's dirty list */
1550 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1553 int (*insert)(MDB_ID2L, MDB_ID2 *);
1555 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1556 insert = mdb_mid2l_append;
1558 insert = mdb_mid2l_insert;
1560 mid.mid = mp->mp_pgno;
1562 insert(txn->mt_u.dirty_list, &mid);
1563 txn->mt_dirty_room--;
1566 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1567 * me_pghead and mt_next_pgno.
1569 * If there are free pages available from older transactions, they
1570 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1571 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1572 * and move me_pglast to say which records were consumed. Only this
1573 * function can create me_pghead and move me_pglast/mt_next_pgno.
1574 * @param[in] mc cursor A cursor handle identifying the transaction and
1575 * database for which we are allocating.
1576 * @param[in] num the number of pages to allocate.
1577 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1578 * will always be satisfied by a single contiguous chunk of memory.
1579 * @return 0 on success, non-zero on failure.
1582 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1584 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1585 /* Get at most <Max_retries> more freeDB records once me_pghead
1586 * has enough pages. If not enough, use new pages from the map.
1587 * If <Paranoid> and mc is updating the freeDB, only get new
1588 * records if me_pghead is empty. Then the freelist cannot play
1589 * catch-up with itself by growing while trying to save it.
1591 enum { Paranoid = 1, Max_retries = 500 };
1593 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1595 int rc, n2 = num-1, retry = Max_retries;
1596 MDB_txn *txn = mc->mc_txn;
1597 MDB_env *env = txn->mt_env;
1598 pgno_t pgno, *mop = env->me_pghead;
1599 unsigned i, j, k, mop_len = mop ? mop[0] : 0;
1601 txnid_t oldest = 0, last;
1607 /* If our dirty list is already full, we can't do anything */
1608 if (txn->mt_dirty_room == 0)
1609 return MDB_TXN_FULL;
1611 for (op = MDB_FIRST;; op = MDB_NEXT) {
1614 pgno_t *idl, old_id, new_id;
1616 /* Seek a big enough contiguous page range. Prefer
1617 * pages at the tail, just truncating the list.
1619 if (mop_len >= (unsigned)num) {
1623 if (mop[i-n2] == pgno+n2)
1625 } while (--i >= (unsigned)num);
1626 if (Max_retries < INT_MAX && --retry < 0)
1630 if (op == MDB_FIRST) { /* 1st iteration */
1631 /* Prepare to fetch more and coalesce */
1632 oldest = mdb_find_oldest(txn);
1633 last = env->me_pglast;
1634 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1637 key.mv_data = &last; /* will look up last+1 */
1638 key.mv_size = sizeof(last);
1640 if (Paranoid && mc->mc_dbi == FREE_DBI)
1643 if (Paranoid && retry < 0 && mop_len)
1647 /* Do not fetch more if the record will be too recent */
1650 rc = mdb_cursor_get(&m2, &key, NULL, op);
1652 if (rc == MDB_NOTFOUND)
1656 last = *(txnid_t*)key.mv_data;
1659 np = m2.mc_pg[m2.mc_top];
1660 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1661 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1664 idl = (MDB_ID *) data.mv_data;
1667 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1670 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1672 mop = env->me_pghead;
1674 env->me_pglast = last;
1676 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1677 last, txn->mt_dbs[FREE_DBI].md_root, i));
1679 DPRINTF(("IDL %"Z"u", idl[k]));
1681 /* Merge in descending sorted order */
1684 mop[0] = (pgno_t)-1;
1688 for (; old_id < new_id; old_id = mop[--j])
1695 /* Use new pages from the map when nothing suitable in the freeDB */
1697 pgno = txn->mt_next_pgno;
1698 if (pgno + num >= env->me_maxpg) {
1699 DPUTS("DB size maxed out");
1700 return MDB_MAP_FULL;
1704 if (env->me_flags & MDB_WRITEMAP) {
1705 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1707 if (!(np = mdb_page_malloc(txn, num)))
1711 mop[0] = mop_len -= num;
1712 /* Move any stragglers down */
1713 for (j = i-num; j < mop_len; )
1714 mop[++j] = mop[++i];
1716 txn->mt_next_pgno = pgno + num;
1719 mdb_page_dirty(txn, np);
1725 /** Copy the used portions of a non-overflow page.
1726 * @param[in] dst page to copy into
1727 * @param[in] src page to copy from
1728 * @param[in] psize size of a page
1731 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1733 enum { Align = sizeof(pgno_t) };
1734 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1736 /* If page isn't full, just copy the used portion. Adjust
1737 * alignment so memcpy may copy words instead of bytes.
1739 if ((unused &= -Align) && !IS_LEAF2(src)) {
1741 memcpy(dst, src, (lower + (Align-1)) & -Align);
1742 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1745 memcpy(dst, src, psize - unused);
1749 /** Pull a page off the txn's spill list, if present.
1750 * If a page being referenced was spilled to disk in this txn, bring
1751 * it back and make it dirty/writable again.
1752 * @param[in] tx0 the transaction handle.
1753 * @param[in] mp the page being referenced.
1754 * @param[out] ret the writable page, if any. ret is unchanged if
1755 * mp wasn't spilled.
1758 mdb_page_unspill(MDB_txn *tx0, MDB_page *mp, MDB_page **ret)
1760 MDB_env *env = tx0->mt_env;
1763 pgno_t pgno = mp->mp_pgno;
1765 for (txn = tx0; txn; txn=txn->mt_parent) {
1766 if (!txn->mt_spill_pgs)
1768 x = mdb_midl_search(txn->mt_spill_pgs, pgno);
1769 if (x <= txn->mt_spill_pgs[0] && txn->mt_spill_pgs[x] == pgno) {
1772 if (IS_OVERFLOW(mp))
1776 if (env->me_flags & MDB_WRITEMAP) {
1779 np = mdb_page_malloc(txn, num);
1783 memcpy(np, mp, num * env->me_psize);
1785 mdb_page_copy(np, mp, env->me_psize);
1788 /* If in current txn, this page is no longer spilled */
1789 for (; x < txn->mt_spill_pgs[0]; x++)
1790 txn->mt_spill_pgs[x] = txn->mt_spill_pgs[x+1];
1791 txn->mt_spill_pgs[0]--;
1792 } /* otherwise, if belonging to a parent txn, the
1793 * page remains spilled until child commits
1796 if (txn->mt_parent) {
1798 /* If this page is also in a parent's dirty list, then
1799 * it's already accounted in dirty_room, and we need to
1800 * cancel out the decrement that mdb_page_dirty does.
1802 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1803 x = mdb_mid2l_search(tx2->mt_u.dirty_list, pgno);
1804 if (x <= tx2->mt_u.dirty_list[0].mid &&
1805 tx2->mt_u.dirty_list[x].mid == pgno) {
1806 txn->mt_dirty_room++;
1811 mdb_page_dirty(tx0, np);
1812 np->mp_flags |= P_DIRTY;
1820 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1821 * @param[in] mc cursor pointing to the page to be touched
1822 * @return 0 on success, non-zero on failure.
1825 mdb_page_touch(MDB_cursor *mc)
1827 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1828 MDB_txn *txn = mc->mc_txn;
1829 MDB_cursor *m2, *m3;
1834 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1835 if (txn->mt_flags & MDB_TXN_SPILLS) {
1837 rc = mdb_page_unspill(txn, mp, &np);
1843 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1844 (rc = mdb_page_alloc(mc, 1, &np)))
1847 DPRINTF(("touched db %u page %"Z"u -> %"Z"u", mc->mc_dbi,mp->mp_pgno,pgno));
1848 assert(mp->mp_pgno != pgno);
1849 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1850 /* Update the parent page, if any, to point to the new page */
1852 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1853 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1854 SETPGNO(node, pgno);
1856 mc->mc_db->md_root = pgno;
1858 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1859 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1861 /* If txn has a parent, make sure the page is in our
1865 unsigned x = mdb_mid2l_search(dl, pgno);
1866 if (x <= dl[0].mid && dl[x].mid == pgno) {
1867 if (mp != dl[x].mptr) { /* bad cursor? */
1868 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1869 return MDB_CORRUPTED;
1874 assert(dl[0].mid < MDB_IDL_UM_MAX);
1876 np = mdb_page_malloc(txn, 1);
1881 mdb_mid2l_insert(dl, &mid);
1886 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1888 np->mp_flags |= P_DIRTY;
1891 /* Adjust cursors pointing to mp */
1892 mc->mc_pg[mc->mc_top] = np;
1894 if (mc->mc_flags & C_SUB) {
1896 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1897 m3 = &m2->mc_xcursor->mx_cursor;
1898 if (m3->mc_snum < mc->mc_snum) continue;
1899 if (m3->mc_pg[mc->mc_top] == mp)
1900 m3->mc_pg[mc->mc_top] = np;
1903 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1904 if (m2->mc_snum < mc->mc_snum) continue;
1905 if (m2->mc_pg[mc->mc_top] == mp) {
1906 m2->mc_pg[mc->mc_top] = np;
1907 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1908 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1910 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1911 if (!(leaf->mn_flags & F_SUBDATA))
1912 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1921 mdb_env_sync(MDB_env *env, int force)
1924 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1925 if (env->me_flags & MDB_WRITEMAP) {
1926 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1927 ? MS_ASYNC : MS_SYNC;
1928 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1931 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1935 if (MDB_FDATASYNC(env->me_fd))
1942 /** Back up parent txn's cursors, then grab the originals for tracking */
1944 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1946 MDB_cursor *mc, *bk;
1951 for (i = src->mt_numdbs; --i >= 0; ) {
1952 if ((mc = src->mt_cursors[i]) != NULL) {
1953 size = sizeof(MDB_cursor);
1955 size += sizeof(MDB_xcursor);
1956 for (; mc; mc = bk->mc_next) {
1962 mc->mc_db = &dst->mt_dbs[i];
1963 /* Kill pointers into src - and dst to reduce abuse: The
1964 * user may not use mc until dst ends. Otherwise we'd...
1966 mc->mc_txn = NULL; /* ...set this to dst */
1967 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
1968 if ((mx = mc->mc_xcursor) != NULL) {
1969 *(MDB_xcursor *)(bk+1) = *mx;
1970 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
1972 mc->mc_next = dst->mt_cursors[i];
1973 dst->mt_cursors[i] = mc;
1980 /** Close this write txn's cursors, give parent txn's cursors back to parent.
1981 * @param[in] txn the transaction handle.
1982 * @param[in] merge true to keep changes to parent cursors, false to revert.
1983 * @return 0 on success, non-zero on failure.
1986 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1988 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
1992 for (i = txn->mt_numdbs; --i >= 0; ) {
1993 for (mc = cursors[i]; mc; mc = next) {
1995 if ((bk = mc->mc_backup) != NULL) {
1997 /* Commit changes to parent txn */
1998 mc->mc_next = bk->mc_next;
1999 mc->mc_backup = bk->mc_backup;
2000 mc->mc_txn = bk->mc_txn;
2001 mc->mc_db = bk->mc_db;
2002 mc->mc_dbflag = bk->mc_dbflag;
2003 if ((mx = mc->mc_xcursor) != NULL)
2004 mx->mx_cursor.mc_txn = bk->mc_txn;
2006 /* Abort nested txn */
2008 if ((mx = mc->mc_xcursor) != NULL)
2009 *mx = *(MDB_xcursor *)(bk+1);
2013 /* Only malloced cursors are permanently tracked. */
2021 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2024 mdb_txn_reset0(MDB_txn *txn, const char *act);
2026 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2032 Pidset = F_SETLK, Pidcheck = F_GETLK
2036 /** Set or check a pid lock. Set returns 0 on success.
2037 * Check returns 0 if the process is certainly dead, nonzero if it may
2038 * be alive (the lock exists or an error happened so we do not know).
2040 * On Windows Pidset is a no-op, we merely check for the existence
2041 * of the process with the given pid. On POSIX we use a single byte
2042 * lock on the lockfile, set at an offset equal to the pid.
2045 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, pid_t pid)
2047 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2050 if (op == Pidcheck) {
2051 h = OpenProcess(env->me_pidquery, FALSE, pid);
2052 /* No documented "no such process" code, but other program use this: */
2054 return ErrCode() != ERROR_INVALID_PARAMETER;
2055 /* A process exists until all handles to it close. Has it exited? */
2056 ret = WaitForSingleObject(h, 0) != 0;
2063 struct flock lock_info;
2064 memset(&lock_info, 0, sizeof(lock_info));
2065 lock_info.l_type = F_WRLCK;
2066 lock_info.l_whence = SEEK_SET;
2067 lock_info.l_start = pid;
2068 lock_info.l_len = 1;
2069 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2070 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2072 } else if ((rc = ErrCode()) == EINTR) {
2080 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2081 * @param[in] txn the transaction handle to initialize
2082 * @return 0 on success, non-zero on failure.
2085 mdb_txn_renew0(MDB_txn *txn)
2087 MDB_env *env = txn->mt_env;
2090 int rc, new_notls = 0;
2093 txn->mt_numdbs = env->me_numdbs;
2094 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2096 if (txn->mt_flags & MDB_TXN_RDONLY) {
2097 if (!env->me_txns) {
2098 i = mdb_env_pick_meta(env);
2099 txn->mt_txnid = env->me_metas[i]->mm_txnid;
2100 txn->mt_u.reader = NULL;
2102 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2103 pthread_getspecific(env->me_txkey);
2105 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2106 return MDB_BAD_RSLOT;
2108 pid_t pid = env->me_pid;
2109 pthread_t tid = pthread_self();
2111 if (!(env->me_flags & MDB_LIVE_READER)) {
2112 rc = mdb_reader_pid(env, Pidset, pid);
2114 UNLOCK_MUTEX_R(env);
2117 env->me_flags |= MDB_LIVE_READER;
2121 for (i=0; i<env->me_txns->mti_numreaders; i++)
2122 if (env->me_txns->mti_readers[i].mr_pid == 0)
2124 if (i == env->me_maxreaders) {
2125 UNLOCK_MUTEX_R(env);
2126 return MDB_READERS_FULL;
2128 env->me_txns->mti_readers[i].mr_pid = pid;
2129 env->me_txns->mti_readers[i].mr_tid = tid;
2130 if (i >= env->me_txns->mti_numreaders)
2131 env->me_txns->mti_numreaders = i+1;
2132 /* Save numreaders for un-mutexed mdb_env_close() */
2133 env->me_numreaders = env->me_txns->mti_numreaders;
2134 UNLOCK_MUTEX_R(env);
2135 r = &env->me_txns->mti_readers[i];
2136 new_notls = (env->me_flags & MDB_NOTLS);
2137 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2142 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
2143 txn->mt_u.reader = r;
2145 txn->mt_toggle = txn->mt_txnid & 1;
2149 txn->mt_txnid = env->me_txns->mti_txnid;
2150 txn->mt_toggle = txn->mt_txnid & 1;
2153 if (txn->mt_txnid == mdb_debug_start)
2156 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2157 txn->mt_u.dirty_list = env->me_dirty_list;
2158 txn->mt_u.dirty_list[0].mid = 0;
2159 txn->mt_free_pgs = env->me_free_pgs;
2160 txn->mt_free_pgs[0] = 0;
2161 txn->mt_spill_pgs = NULL;
2165 /* Copy the DB info and flags */
2166 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
2168 /* Moved to here to avoid a data race in read TXNs */
2169 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
2171 for (i=2; i<txn->mt_numdbs; i++) {
2172 x = env->me_dbflags[i];
2173 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2174 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2176 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2178 if (env->me_maxpg < txn->mt_next_pgno) {
2179 mdb_txn_reset0(txn, "renew0-mapfail");
2181 txn->mt_u.reader->mr_pid = 0;
2182 txn->mt_u.reader = NULL;
2184 return MDB_MAP_RESIZED;
2191 mdb_txn_renew(MDB_txn *txn)
2195 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2198 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2199 DPUTS("environment had fatal error, must shutdown!");
2203 rc = mdb_txn_renew0(txn);
2204 if (rc == MDB_SUCCESS) {
2205 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2206 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2207 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2213 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2217 int rc, size, tsize = sizeof(MDB_txn);
2219 if (env->me_flags & MDB_FATAL_ERROR) {
2220 DPUTS("environment had fatal error, must shutdown!");
2223 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2226 /* Nested transactions: Max 1 child, write txns only, no writemap */
2227 if (parent->mt_child ||
2228 (flags & MDB_RDONLY) ||
2229 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2230 (env->me_flags & MDB_WRITEMAP))
2232 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2234 tsize = sizeof(MDB_ntxn);
2236 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2237 if (!(flags & MDB_RDONLY))
2238 size += env->me_maxdbs * sizeof(MDB_cursor *);
2240 if ((txn = calloc(1, size)) == NULL) {
2241 DPRINTF(("calloc: %s", strerror(ErrCode())));
2244 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2245 if (flags & MDB_RDONLY) {
2246 txn->mt_flags |= MDB_TXN_RDONLY;
2247 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2249 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2250 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2256 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2257 if (!txn->mt_u.dirty_list ||
2258 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2260 free(txn->mt_u.dirty_list);
2264 txn->mt_txnid = parent->mt_txnid;
2265 txn->mt_toggle = parent->mt_toggle;
2266 txn->mt_dirty_room = parent->mt_dirty_room;
2267 txn->mt_u.dirty_list[0].mid = 0;
2268 txn->mt_spill_pgs = NULL;
2269 txn->mt_next_pgno = parent->mt_next_pgno;
2270 parent->mt_child = txn;
2271 txn->mt_parent = parent;
2272 txn->mt_numdbs = parent->mt_numdbs;
2273 txn->mt_flags = parent->mt_flags;
2274 txn->mt_dbxs = parent->mt_dbxs;
2275 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2276 /* Copy parent's mt_dbflags, but clear DB_NEW */
2277 for (i=0; i<txn->mt_numdbs; i++)
2278 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2280 ntxn = (MDB_ntxn *)txn;
2281 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2282 if (env->me_pghead) {
2283 size = MDB_IDL_SIZEOF(env->me_pghead);
2284 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2286 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2291 rc = mdb_cursor_shadow(parent, txn);
2293 mdb_txn_reset0(txn, "beginchild-fail");
2295 rc = mdb_txn_renew0(txn);
2301 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2302 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2303 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2310 mdb_txn_env(MDB_txn *txn)
2312 if(!txn) return NULL;
2316 /** Export or close DBI handles opened in this txn. */
2318 mdb_dbis_update(MDB_txn *txn, int keep)
2321 MDB_dbi n = txn->mt_numdbs;
2322 MDB_env *env = txn->mt_env;
2323 unsigned char *tdbflags = txn->mt_dbflags;
2325 for (i = n; --i >= 2;) {
2326 if (tdbflags[i] & DB_NEW) {
2328 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2330 char *ptr = env->me_dbxs[i].md_name.mv_data;
2331 env->me_dbxs[i].md_name.mv_data = NULL;
2332 env->me_dbxs[i].md_name.mv_size = 0;
2333 env->me_dbflags[i] = 0;
2338 if (keep && env->me_numdbs < n)
2342 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2343 * May be called twice for readonly txns: First reset it, then abort.
2344 * @param[in] txn the transaction handle to reset
2345 * @param[in] act why the transaction is being reset
2348 mdb_txn_reset0(MDB_txn *txn, const char *act)
2350 MDB_env *env = txn->mt_env;
2352 /* Close any DBI handles opened in this txn */
2353 mdb_dbis_update(txn, 0);
2355 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2356 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2357 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2359 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2360 if (txn->mt_u.reader) {
2361 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2362 if (!(env->me_flags & MDB_NOTLS))
2363 txn->mt_u.reader = NULL; /* txn does not own reader */
2365 txn->mt_numdbs = 0; /* close nothing if called again */
2366 txn->mt_dbxs = NULL; /* mark txn as reset */
2368 mdb_cursors_close(txn, 0);
2370 if (!(env->me_flags & MDB_WRITEMAP)) {
2371 mdb_dlist_free(txn);
2373 mdb_midl_free(env->me_pghead);
2375 if (txn->mt_parent) {
2376 txn->mt_parent->mt_child = NULL;
2377 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2378 mdb_midl_free(txn->mt_free_pgs);
2379 mdb_midl_free(txn->mt_spill_pgs);
2380 free(txn->mt_u.dirty_list);
2384 if (mdb_midl_shrink(&txn->mt_free_pgs))
2385 env->me_free_pgs = txn->mt_free_pgs;
2386 env->me_pghead = NULL;
2390 /* The writer mutex was locked in mdb_txn_begin. */
2391 UNLOCK_MUTEX_W(env);
2396 mdb_txn_reset(MDB_txn *txn)
2401 /* This call is only valid for read-only txns */
2402 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2405 mdb_txn_reset0(txn, "reset");
2409 mdb_txn_abort(MDB_txn *txn)
2415 mdb_txn_abort(txn->mt_child);
2417 mdb_txn_reset0(txn, "abort");
2418 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2419 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2420 txn->mt_u.reader->mr_pid = 0;
2425 /** Save the freelist as of this transaction to the freeDB.
2426 * This changes the freelist. Keep trying until it stabilizes.
2429 mdb_freelist_save(MDB_txn *txn)
2431 /* env->me_pghead[] can grow and shrink during this call.
2432 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2433 * Page numbers cannot disappear from txn->mt_free_pgs[].
2436 MDB_env *env = txn->mt_env;
2437 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2438 txnid_t pglast = 0, head_id = 0;
2439 pgno_t freecnt = 0, *free_pgs, *mop;
2440 ssize_t head_room = 0, total_room = 0, mop_len;
2442 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2444 if (env->me_pghead) {
2445 /* Make sure first page of freeDB is touched and on freelist */
2446 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2447 if (rc && rc != MDB_NOTFOUND)
2452 /* Come back here after each Put() in case freelist changed */
2455 /* If using records from freeDB which we have not yet
2456 * deleted, delete them and any we reserved for me_pghead.
2458 while (pglast < env->me_pglast) {
2459 rc = mdb_cursor_first(&mc, &key, NULL);
2462 pglast = head_id = *(txnid_t *)key.mv_data;
2463 total_room = head_room = 0;
2464 assert(pglast <= env->me_pglast);
2465 rc = mdb_cursor_del(&mc, 0);
2470 /* Save the IDL of pages freed by this txn, to a single record */
2471 if (freecnt < txn->mt_free_pgs[0]) {
2473 /* Make sure last page of freeDB is touched and on freelist */
2474 key.mv_size = MDB_MAXKEYSIZE+1;
2476 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2477 if (rc && rc != MDB_NOTFOUND)
2480 free_pgs = txn->mt_free_pgs;
2481 /* Write to last page of freeDB */
2482 key.mv_size = sizeof(txn->mt_txnid);
2483 key.mv_data = &txn->mt_txnid;
2485 freecnt = free_pgs[0];
2486 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2487 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2490 /* Retry if mt_free_pgs[] grew during the Put() */
2491 free_pgs = txn->mt_free_pgs;
2492 } while (freecnt < free_pgs[0]);
2493 mdb_midl_sort(free_pgs);
2494 memcpy(data.mv_data, free_pgs, data.mv_size);
2497 unsigned int i = free_pgs[0];
2498 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2499 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2501 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2507 mop = env->me_pghead;
2508 mop_len = mop ? mop[0] : 0;
2510 /* Reserve records for me_pghead[]. Split it if multi-page,
2511 * to avoid searching freeDB for a page range. Use keys in
2512 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2514 if (total_room >= mop_len) {
2515 if (total_room == mop_len || --more < 0)
2517 } else if (head_room >= maxfree_1pg && head_id > 1) {
2518 /* Keep current record (overflow page), add a new one */
2522 /* (Re)write {key = head_id, IDL length = head_room} */
2523 total_room -= head_room;
2524 head_room = mop_len - total_room;
2525 if (head_room > maxfree_1pg && head_id > 1) {
2526 /* Overflow multi-page for part of me_pghead */
2527 head_room /= head_id; /* amortize page sizes */
2528 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2529 } else if (head_room < 0) {
2530 /* Rare case, not bothering to delete this record */
2533 key.mv_size = sizeof(head_id);
2534 key.mv_data = &head_id;
2535 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2536 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2539 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2540 total_room += head_room;
2543 /* Fill in the reserved, touched me_pghead records */
2549 rc = mdb_cursor_first(&mc, &key, &data);
2550 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2551 unsigned flags = MDB_CURRENT;
2552 txnid_t id = *(txnid_t *)key.mv_data;
2553 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2556 assert(len >= 0 && id <= env->me_pglast);
2558 if (len > mop_len) {
2560 data.mv_size = (len + 1) * sizeof(MDB_ID);
2563 data.mv_data = mop -= len;
2566 rc = mdb_cursor_put(&mc, &key, &data, flags);
2568 if (rc || !(mop_len -= len))
2575 /** Flush dirty pages to the map, after clearing their dirty flag.
2578 mdb_page_flush(MDB_txn *txn)
2580 MDB_env *env = txn->mt_env;
2581 MDB_ID2L dl = txn->mt_u.dirty_list;
2582 unsigned psize = env->me_psize, j;
2583 int i, pagecount = dl[0].mid, rc;
2584 size_t size = 0, pos = 0;
2586 MDB_page *dp = NULL;
2590 struct iovec iov[MDB_COMMIT_PAGES];
2591 ssize_t wpos = 0, wsize = 0, wres;
2592 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2597 if (env->me_flags & MDB_WRITEMAP) {
2598 /* Clear dirty flags */
2599 for (i=1; i<=pagecount; i++) {
2601 /* Don't flush this page yet */
2602 if (dp->mp_flags & P_KEEP) {
2603 dp->mp_flags ^= P_KEEP;
2607 dp->mp_flags &= ~P_DIRTY;
2613 /* Write the pages */
2615 if (i <= pagecount) {
2617 /* Don't flush this page yet */
2618 if (dp->mp_flags & P_KEEP) {
2619 dp->mp_flags ^= P_KEEP;
2624 /* clear dirty flag */
2625 dp->mp_flags &= ~P_DIRTY;
2628 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2633 /* Windows actually supports scatter/gather I/O, but only on
2634 * unbuffered file handles. Since we're relying on the OS page
2635 * cache for all our data, that's self-defeating. So we just
2636 * write pages one at a time. We use the ov structure to set
2637 * the write offset, to at least save the overhead of a Seek
2640 DPRINTF(("committing page %"Z"u", pgno));
2641 memset(&ov, 0, sizeof(ov));
2642 ov.Offset = pos & 0xffffffff;
2643 ov.OffsetHigh = pos >> 16 >> 16;
2644 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2646 DPRINTF(("WriteFile: %d", rc));
2650 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2651 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2653 /* Write previous page(s) */
2654 #ifdef MDB_USE_PWRITEV
2655 wres = pwritev(env->me_fd, iov, n, wpos);
2658 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2660 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2662 DPRINTF(("lseek: %s", strerror(rc)));
2665 wres = writev(env->me_fd, iov, n);
2668 if (wres != wsize) {
2671 DPRINTF(("Write error: %s", strerror(rc)));
2673 rc = EIO; /* TODO: Use which error code? */
2674 DPUTS("short write, filesystem full?");
2685 DPRINTF(("committing page %"Z"u", pgno));
2686 next_pos = pos + size;
2687 iov[n].iov_len = size;
2688 iov[n].iov_base = (char *)dp;
2695 for (i=1; i<=pagecount; i++) {
2697 /* This is a page we skipped above */
2700 dl[j].mid = dp->mp_pgno;
2703 mdb_dpage_free(env, dp);
2711 mdb_txn_commit(MDB_txn *txn)
2717 assert(txn != NULL);
2718 assert(txn->mt_env != NULL);
2720 if (txn->mt_child) {
2721 rc = mdb_txn_commit(txn->mt_child);
2722 txn->mt_child = NULL;
2729 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2730 mdb_dbis_update(txn, 1);
2731 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2736 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2737 DPUTS("error flag is set, can't commit");
2739 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2744 if (txn->mt_parent) {
2745 MDB_txn *parent = txn->mt_parent;
2749 /* Append our free list to parent's */
2750 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2753 mdb_midl_free(txn->mt_free_pgs);
2755 parent->mt_next_pgno = txn->mt_next_pgno;
2756 parent->mt_flags = txn->mt_flags;
2758 /* Merge our cursors into parent's and close them */
2759 mdb_cursors_close(txn, 1);
2761 /* Update parent's DB table. */
2762 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2763 parent->mt_numdbs = txn->mt_numdbs;
2764 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2765 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2766 for (i=2; i<txn->mt_numdbs; i++) {
2767 /* preserve parent's DB_NEW status */
2768 x = parent->mt_dbflags[i] & DB_NEW;
2769 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2772 dst = parent->mt_u.dirty_list;
2773 src = txn->mt_u.dirty_list;
2774 /* Remove anything in our dirty list from parent's spill list */
2775 if (parent->mt_spill_pgs) {
2776 x = parent->mt_spill_pgs[0];
2778 /* zero out our dirty pages in parent spill list */
2779 for (i=1; i<=src[0].mid; i++) {
2780 if (src[i].mid < parent->mt_spill_pgs[x])
2782 if (src[i].mid > parent->mt_spill_pgs[x]) {
2788 parent->mt_spill_pgs[x] = 0;
2791 /* OK, we had a few hits, squash zeros from the spill list */
2792 if (len < parent->mt_spill_pgs[0]) {
2794 for (y=1; y<=parent->mt_spill_pgs[0]; y++) {
2795 if (parent->mt_spill_pgs[y]) {
2797 parent->mt_spill_pgs[x] = parent->mt_spill_pgs[y];
2802 parent->mt_spill_pgs[0] = len;
2805 /* Find len = length of merging our dirty list with parent's */
2807 dst[0].mid = 0; /* simplify loops */
2808 if (parent->mt_parent) {
2809 len = x + src[0].mid;
2810 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2811 for (i = x; y && i; y--) {
2812 pgno_t yp = src[y].mid;
2813 while (yp < dst[i].mid)
2815 if (yp == dst[i].mid) {
2820 } else { /* Simplify the above for single-ancestor case */
2821 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2823 /* Merge our dirty list with parent's */
2825 for (i = len; y; dst[i--] = src[y--]) {
2826 pgno_t yp = src[y].mid;
2827 while (yp < dst[x].mid)
2828 dst[i--] = dst[x--];
2829 if (yp == dst[x].mid)
2830 free(dst[x--].mptr);
2834 free(txn->mt_u.dirty_list);
2835 parent->mt_dirty_room = txn->mt_dirty_room;
2836 if (txn->mt_spill_pgs) {
2837 if (parent->mt_spill_pgs) {
2838 mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2839 mdb_midl_free(txn->mt_spill_pgs);
2840 mdb_midl_sort(parent->mt_spill_pgs);
2842 parent->mt_spill_pgs = txn->mt_spill_pgs;
2846 parent->mt_child = NULL;
2847 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2852 if (txn != env->me_txn) {
2853 DPUTS("attempt to commit unknown transaction");
2858 mdb_cursors_close(txn, 0);
2860 if (!txn->mt_u.dirty_list[0].mid &&
2861 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2864 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2865 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2867 /* Update DB root pointers */
2868 if (txn->mt_numdbs > 2) {
2872 data.mv_size = sizeof(MDB_db);
2874 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2875 for (i = 2; i < txn->mt_numdbs; i++) {
2876 if (txn->mt_dbflags[i] & DB_DIRTY) {
2877 data.mv_data = &txn->mt_dbs[i];
2878 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2885 rc = mdb_freelist_save(txn);
2889 mdb_midl_free(env->me_pghead);
2890 env->me_pghead = NULL;
2891 if (mdb_midl_shrink(&txn->mt_free_pgs))
2892 env->me_free_pgs = txn->mt_free_pgs;
2898 if ((rc = mdb_page_flush(txn)) ||
2899 (rc = mdb_env_sync(env, 0)) ||
2900 (rc = mdb_env_write_meta(txn)))
2906 mdb_dbis_update(txn, 1);
2908 UNLOCK_MUTEX_W(env);
2918 /** Read the environment parameters of a DB environment before
2919 * mapping it into memory.
2920 * @param[in] env the environment handle
2921 * @param[out] meta address of where to store the meta information
2922 * @return 0 on success, non-zero on failure.
2925 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2932 /* We don't know the page size yet, so use a minimum value.
2933 * Read both meta pages so we can use the latest one.
2936 for (i=off=0; i<2; i++, off = meta->mm_psize) {
2940 memset(&ov, 0, sizeof(ov));
2942 rc = ReadFile(env->me_fd,&pbuf,MDB_PAGESIZE,&len,&ov) ? (int)len : -1;
2943 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
2946 rc = pread(env->me_fd, &pbuf, MDB_PAGESIZE, off);
2948 if (rc != MDB_PAGESIZE) {
2949 if (rc == 0 && off == 0)
2951 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
2952 DPRINTF(("read: %s", mdb_strerror(rc)));
2956 p = (MDB_page *)&pbuf;
2958 if (!F_ISSET(p->mp_flags, P_META)) {
2959 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
2964 if (m->mm_magic != MDB_MAGIC) {
2965 DPUTS("meta has invalid magic");
2969 if (m->mm_version != MDB_DATA_VERSION) {
2970 DPRINTF(("database is version %u, expected version %u",
2971 m->mm_version, MDB_DATA_VERSION));
2972 return MDB_VERSION_MISMATCH;
2975 if (off == 0 || m->mm_txnid > meta->mm_txnid)
2981 /** Write the environment parameters of a freshly created DB environment.
2982 * @param[in] env the environment handle
2983 * @param[out] meta address of where to store the meta information
2984 * @return 0 on success, non-zero on failure.
2987 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2995 memset(&ov, 0, sizeof(ov));
2996 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
2998 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3001 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3002 len = pwrite(fd, ptr, size, pos); \
3003 rc = (len >= 0); } while(0)
3006 DPUTS("writing new meta page");
3008 GET_PAGESIZE(psize);
3010 meta->mm_magic = MDB_MAGIC;
3011 meta->mm_version = MDB_DATA_VERSION;
3012 meta->mm_mapsize = env->me_mapsize;
3013 meta->mm_psize = psize;
3014 meta->mm_last_pg = 1;
3015 meta->mm_flags = env->me_flags & 0xffff;
3016 meta->mm_flags |= MDB_INTEGERKEY;
3017 meta->mm_dbs[0].md_root = P_INVALID;
3018 meta->mm_dbs[1].md_root = P_INVALID;
3020 p = calloc(2, psize);
3022 p->mp_flags = P_META;
3023 *(MDB_meta *)METADATA(p) = *meta;
3025 q = (MDB_page *)((char *)p + psize);
3027 q->mp_flags = P_META;
3028 *(MDB_meta *)METADATA(q) = *meta;
3030 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3033 else if ((unsigned) len == psize * 2)
3041 /** Update the environment info to commit a transaction.
3042 * @param[in] txn the transaction that's being committed
3043 * @return 0 on success, non-zero on failure.
3046 mdb_env_write_meta(MDB_txn *txn)
3049 MDB_meta meta, metab, *mp;
3051 int rc, len, toggle;
3060 assert(txn != NULL);
3061 assert(txn->mt_env != NULL);
3063 toggle = !txn->mt_toggle;
3064 DPRINTF(("writing meta page %d for root page %"Z"u",
3065 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3068 mp = env->me_metas[toggle];
3070 if (env->me_flags & MDB_WRITEMAP) {
3071 /* Persist any increases of mapsize config */
3072 if (env->me_mapsize > mp->mm_mapsize)
3073 mp->mm_mapsize = env->me_mapsize;
3074 mp->mm_dbs[0] = txn->mt_dbs[0];
3075 mp->mm_dbs[1] = txn->mt_dbs[1];
3076 mp->mm_last_pg = txn->mt_next_pgno - 1;
3077 mp->mm_txnid = txn->mt_txnid;
3078 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3079 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3082 ptr += env->me_psize;
3083 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
3090 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3091 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3093 ptr = (char *)&meta;
3094 if (env->me_mapsize > mp->mm_mapsize) {
3095 /* Persist any increases of mapsize config */
3096 meta.mm_mapsize = env->me_mapsize;
3097 off = offsetof(MDB_meta, mm_mapsize);
3099 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3101 len = sizeof(MDB_meta) - off;
3104 meta.mm_dbs[0] = txn->mt_dbs[0];
3105 meta.mm_dbs[1] = txn->mt_dbs[1];
3106 meta.mm_last_pg = txn->mt_next_pgno - 1;
3107 meta.mm_txnid = txn->mt_txnid;
3110 off += env->me_psize;
3113 /* Write to the SYNC fd */
3114 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3115 env->me_fd : env->me_mfd;
3118 memset(&ov, 0, sizeof(ov));
3120 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3124 rc = pwrite(mfd, ptr, len, off);
3127 rc = rc < 0 ? ErrCode() : EIO;
3128 DPUTS("write failed, disk error?");
3129 /* On a failure, the pagecache still contains the new data.
3130 * Write some old data back, to prevent it from being used.
3131 * Use the non-SYNC fd; we know it will fail anyway.
3133 meta.mm_last_pg = metab.mm_last_pg;
3134 meta.mm_txnid = metab.mm_txnid;
3136 memset(&ov, 0, sizeof(ov));
3138 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3140 r2 = pwrite(env->me_fd, ptr, len, off);
3143 env->me_flags |= MDB_FATAL_ERROR;
3147 /* Memory ordering issues are irrelevant; since the entire writer
3148 * is wrapped by wmutex, all of these changes will become visible
3149 * after the wmutex is unlocked. Since the DB is multi-version,
3150 * readers will get consistent data regardless of how fresh or
3151 * how stale their view of these values is.
3153 env->me_txns->mti_txnid = txn->mt_txnid;
3158 /** Check both meta pages to see which one is newer.
3159 * @param[in] env the environment handle
3160 * @return meta toggle (0 or 1).
3163 mdb_env_pick_meta(const MDB_env *env)
3165 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3169 mdb_env_create(MDB_env **env)
3173 e = calloc(1, sizeof(MDB_env));
3177 e->me_maxreaders = DEFAULT_READERS;
3178 e->me_maxdbs = e->me_numdbs = 2;
3179 e->me_fd = INVALID_HANDLE_VALUE;
3180 e->me_lfd = INVALID_HANDLE_VALUE;
3181 e->me_mfd = INVALID_HANDLE_VALUE;
3182 #ifdef MDB_USE_POSIX_SEM
3183 e->me_rmutex = SEM_FAILED;
3184 e->me_wmutex = SEM_FAILED;
3186 e->me_pid = getpid();
3187 VGMEMP_CREATE(e,0,0);
3193 mdb_env_set_mapsize(MDB_env *env, size_t size)
3197 env->me_mapsize = size;
3199 env->me_maxpg = env->me_mapsize / env->me_psize;
3204 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3208 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3213 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3215 if (env->me_map || readers < 1)
3217 env->me_maxreaders = readers;
3222 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3224 if (!env || !readers)
3226 *readers = env->me_maxreaders;
3230 /** Further setup required for opening an MDB environment
3233 mdb_env_open2(MDB_env *env)
3235 unsigned int flags = env->me_flags;
3243 memset(&meta, 0, sizeof(meta));
3245 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3248 DPUTS("new mdbenv");
3252 /* Was a mapsize configured? */
3253 if (!env->me_mapsize) {
3254 /* If this is a new environment, take the default,
3255 * else use the size recorded in the existing env.
3257 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3258 } else if (env->me_mapsize < meta.mm_mapsize) {
3259 /* If the configured size is smaller, make sure it's
3260 * still big enough. Silently round up to minimum if not.
3262 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3263 if (env->me_mapsize < minsize)
3264 env->me_mapsize = minsize;
3271 LONG sizelo, sizehi;
3272 sizelo = env->me_mapsize & 0xffffffff;
3273 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3275 /* See if we should use QueryLimited */
3277 if ((rc & 0xff) > 5)
3278 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3280 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3282 /* Windows won't create mappings for zero length files.
3283 * Just allocate the maxsize right now.
3286 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3287 || !SetEndOfFile(env->me_fd)
3288 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3291 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3292 PAGE_READWRITE : PAGE_READONLY,
3293 sizehi, sizelo, NULL);
3296 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3297 FILE_MAP_WRITE : FILE_MAP_READ,
3298 0, 0, env->me_mapsize, meta.mm_address);
3299 rc = env->me_map ? 0 : ErrCode();
3307 if (flags & MDB_WRITEMAP) {
3309 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3312 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
3314 if (env->me_map == MAP_FAILED) {
3318 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3320 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3322 #ifdef POSIX_MADV_RANDOM
3323 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3324 #endif /* POSIX_MADV_RANDOM */
3325 #endif /* MADV_RANDOM */
3329 if (flags & MDB_FIXEDMAP)
3330 meta.mm_address = env->me_map;
3331 i = mdb_env_init_meta(env, &meta);
3332 if (i != MDB_SUCCESS) {
3335 } else if (meta.mm_address && env->me_map != meta.mm_address) {
3336 /* Can happen because the address argument to mmap() is just a
3337 * hint. mmap() can pick another, e.g. if the range is in use.
3338 * The MAP_FIXED flag would prevent that, but then mmap could
3339 * instead unmap existing pages to make room for the new map.
3341 return EBUSY; /* TODO: Make a new MDB_* error code? */
3343 env->me_psize = meta.mm_psize;
3344 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3345 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
3347 env->me_maxpg = env->me_mapsize / env->me_psize;
3349 p = (MDB_page *)env->me_map;
3350 env->me_metas[0] = METADATA(p);
3351 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
3355 int toggle = mdb_env_pick_meta(env);
3356 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3358 DPRINTF(("opened database version %u, pagesize %u",
3359 env->me_metas[0]->mm_version, env->me_psize));
3360 DPRINTF(("using meta page %d", toggle));
3361 DPRINTF(("depth: %u", db->md_depth));
3362 DPRINTF(("entries: %"Z"u", db->md_entries));
3363 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3364 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3365 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3366 DPRINTF(("root: %"Z"u", db->md_root));
3374 /** Release a reader thread's slot in the reader lock table.
3375 * This function is called automatically when a thread exits.
3376 * @param[in] ptr This points to the slot in the reader lock table.
3379 mdb_env_reader_dest(void *ptr)
3381 MDB_reader *reader = ptr;
3387 /** Junk for arranging thread-specific callbacks on Windows. This is
3388 * necessarily platform and compiler-specific. Windows supports up
3389 * to 1088 keys. Let's assume nobody opens more than 64 environments
3390 * in a single process, for now. They can override this if needed.
3392 #ifndef MAX_TLS_KEYS
3393 #define MAX_TLS_KEYS 64
3395 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3396 static int mdb_tls_nkeys;
3398 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3402 case DLL_PROCESS_ATTACH: break;
3403 case DLL_THREAD_ATTACH: break;
3404 case DLL_THREAD_DETACH:
3405 for (i=0; i<mdb_tls_nkeys; i++) {
3406 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3407 mdb_env_reader_dest(r);
3410 case DLL_PROCESS_DETACH: break;
3415 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3417 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3421 /* Force some symbol references.
3422 * _tls_used forces the linker to create the TLS directory if not already done
3423 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3425 #pragma comment(linker, "/INCLUDE:_tls_used")
3426 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3427 #pragma const_seg(".CRT$XLB")
3428 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3429 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3432 #pragma comment(linker, "/INCLUDE:__tls_used")
3433 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3434 #pragma data_seg(".CRT$XLB")
3435 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3437 #endif /* WIN 32/64 */
3438 #endif /* !__GNUC__ */
3441 /** Downgrade the exclusive lock on the region back to shared */
3443 mdb_env_share_locks(MDB_env *env, int *excl)
3445 int rc = 0, toggle = mdb_env_pick_meta(env);
3447 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3452 /* First acquire a shared lock. The Unlock will
3453 * then release the existing exclusive lock.
3455 memset(&ov, 0, sizeof(ov));
3456 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3459 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3465 struct flock lock_info;
3466 /* The shared lock replaces the existing lock */
3467 memset((void *)&lock_info, 0, sizeof(lock_info));
3468 lock_info.l_type = F_RDLCK;
3469 lock_info.l_whence = SEEK_SET;
3470 lock_info.l_start = 0;
3471 lock_info.l_len = 1;
3472 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3473 (rc = ErrCode()) == EINTR) ;
3474 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3481 /** Try to get exlusive lock, otherwise shared.
3482 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3485 mdb_env_excl_lock(MDB_env *env, int *excl)
3489 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3493 memset(&ov, 0, sizeof(ov));
3494 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3501 struct flock lock_info;
3502 memset((void *)&lock_info, 0, sizeof(lock_info));
3503 lock_info.l_type = F_WRLCK;
3504 lock_info.l_whence = SEEK_SET;
3505 lock_info.l_start = 0;
3506 lock_info.l_len = 1;
3507 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3508 (rc = ErrCode()) == EINTR) ;
3512 # ifdef MDB_USE_POSIX_SEM
3513 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3516 lock_info.l_type = F_RDLCK;
3517 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3518 (rc = ErrCode()) == EINTR) ;
3526 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3528 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3530 * @(#) $Revision: 5.1 $
3531 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3532 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3534 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3538 * Please do not copyright this code. This code is in the public domain.
3540 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3541 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3542 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3543 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3544 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3545 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3546 * PERFORMANCE OF THIS SOFTWARE.
3549 * chongo <Landon Curt Noll> /\oo/\
3550 * http://www.isthe.com/chongo/
3552 * Share and Enjoy! :-)
3555 typedef unsigned long long mdb_hash_t;
3556 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3558 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3559 * @param[in] val value to hash
3560 * @param[in] hval initial value for hash
3561 * @return 64 bit hash
3563 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3564 * hval arg on the first call.
3567 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3569 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3570 unsigned char *end = s + val->mv_size;
3572 * FNV-1a hash each octet of the string
3575 /* xor the bottom with the current octet */
3576 hval ^= (mdb_hash_t)*s++;
3578 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3579 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3580 (hval << 7) + (hval << 8) + (hval << 40);
3582 /* return our new hash value */
3586 /** Hash the string and output the encoded hash.
3587 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3588 * very short name limits. We don't care about the encoding being reversible,
3589 * we just want to preserve as many bits of the input as possible in a
3590 * small printable string.
3591 * @param[in] str string to hash
3592 * @param[out] encbuf an array of 11 chars to hold the hash
3594 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3597 mdb_pack85(unsigned long l, char *out)
3601 for (i=0; i<5; i++) {
3602 *out++ = mdb_a85[l % 85];
3608 mdb_hash_enc(MDB_val *val, char *encbuf)
3610 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3611 unsigned long *l = (unsigned long *)&h;
3613 mdb_pack85(l[0], encbuf);
3614 mdb_pack85(l[1], encbuf+5);
3619 /** Open and/or initialize the lock region for the environment.
3620 * @param[in] env The MDB environment.
3621 * @param[in] lpath The pathname of the file used for the lock region.
3622 * @param[in] mode The Unix permissions for the file, if we create it.
3623 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3624 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3625 * @return 0 on success, non-zero on failure.
3628 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3631 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3633 # define MDB_ERRCODE_ROFS EROFS
3634 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3635 # define MDB_CLOEXEC O_CLOEXEC
3638 # define MDB_CLOEXEC 0
3645 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3646 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3647 FILE_ATTRIBUTE_NORMAL, NULL);
3649 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3651 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3653 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3658 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3659 /* Lose record locks when exec*() */
3660 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3661 fcntl(env->me_lfd, F_SETFD, fdflags);
3664 if (!(env->me_flags & MDB_NOTLS)) {
3665 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3668 env->me_flags |= MDB_ENV_TXKEY;
3670 /* Windows TLS callbacks need help finding their TLS info. */
3671 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3675 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3679 /* Try to get exclusive lock. If we succeed, then
3680 * nobody is using the lock region and we should initialize it.
3682 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3685 size = GetFileSize(env->me_lfd, NULL);
3687 size = lseek(env->me_lfd, 0, SEEK_END);
3688 if (size == -1) goto fail_errno;
3690 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3691 if (size < rsize && *excl > 0) {
3693 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != rsize
3694 || !SetEndOfFile(env->me_lfd))
3697 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3701 size = rsize - sizeof(MDB_txninfo);
3702 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3707 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3709 if (!mh) goto fail_errno;
3710 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3712 if (!env->me_txns) goto fail_errno;
3714 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3716 if (m == MAP_FAILED) goto fail_errno;
3722 BY_HANDLE_FILE_INFORMATION stbuf;
3731 if (!mdb_sec_inited) {
3732 InitializeSecurityDescriptor(&mdb_null_sd,
3733 SECURITY_DESCRIPTOR_REVISION);
3734 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3735 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3736 mdb_all_sa.bInheritHandle = FALSE;
3737 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3740 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3741 idbuf.volume = stbuf.dwVolumeSerialNumber;
3742 idbuf.nhigh = stbuf.nFileIndexHigh;
3743 idbuf.nlow = stbuf.nFileIndexLow;
3744 val.mv_data = &idbuf;
3745 val.mv_size = sizeof(idbuf);
3746 mdb_hash_enc(&val, encbuf);
3747 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3748 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3749 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3750 if (!env->me_rmutex) goto fail_errno;
3751 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3752 if (!env->me_wmutex) goto fail_errno;
3753 #elif defined(MDB_USE_POSIX_SEM)
3762 #if defined(__NetBSD__)
3763 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3765 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3766 idbuf.dev = stbuf.st_dev;
3767 idbuf.ino = stbuf.st_ino;
3768 val.mv_data = &idbuf;
3769 val.mv_size = sizeof(idbuf);
3770 mdb_hash_enc(&val, encbuf);
3771 #ifdef MDB_SHORT_SEMNAMES
3772 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3774 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3775 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3776 /* Clean up after a previous run, if needed: Try to
3777 * remove both semaphores before doing anything else.
3779 sem_unlink(env->me_txns->mti_rmname);
3780 sem_unlink(env->me_txns->mti_wmname);
3781 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3782 O_CREAT|O_EXCL, mode, 1);
3783 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3784 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3785 O_CREAT|O_EXCL, mode, 1);
3786 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3787 #else /* MDB_USE_POSIX_SEM */
3788 pthread_mutexattr_t mattr;
3790 if ((rc = pthread_mutexattr_init(&mattr))
3791 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3792 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3793 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3795 pthread_mutexattr_destroy(&mattr);
3796 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3798 env->me_txns->mti_magic = MDB_MAGIC;
3799 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3800 env->me_txns->mti_txnid = 0;
3801 env->me_txns->mti_numreaders = 0;
3804 if (env->me_txns->mti_magic != MDB_MAGIC) {
3805 DPUTS("lock region has invalid magic");
3809 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3810 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3811 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3812 rc = MDB_VERSION_MISMATCH;
3816 if (rc && rc != EACCES && rc != EAGAIN) {
3820 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3821 if (!env->me_rmutex) goto fail_errno;
3822 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3823 if (!env->me_wmutex) goto fail_errno;
3824 #elif defined(MDB_USE_POSIX_SEM)
3825 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3826 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3827 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3828 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3839 /** The name of the lock file in the DB environment */
3840 #define LOCKNAME "/lock.mdb"
3841 /** The name of the data file in the DB environment */
3842 #define DATANAME "/data.mdb"
3843 /** The suffix of the lock file when no subdir is used */
3844 #define LOCKSUFF "-lock"
3845 /** Only a subset of the @ref mdb_env flags can be changed
3846 * at runtime. Changing other flags requires closing the
3847 * environment and re-opening it with the new flags.
3849 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3850 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3853 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3855 int oflags, rc, len, excl = -1, rmlock = 0;
3856 char *lpath, *dpath;
3858 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3862 if (flags & MDB_NOSUBDIR) {
3863 rc = len + sizeof(LOCKSUFF) + len + 1;
3865 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3870 if (flags & MDB_NOSUBDIR) {
3871 dpath = lpath + len + sizeof(LOCKSUFF);
3872 sprintf(lpath, "%s" LOCKSUFF, path);
3873 strcpy(dpath, path);
3875 dpath = lpath + len + sizeof(LOCKNAME);
3876 sprintf(lpath, "%s" LOCKNAME, path);
3877 sprintf(dpath, "%s" DATANAME, path);
3881 flags |= env->me_flags;
3882 if (flags & MDB_RDONLY) {
3883 /* silently ignore WRITEMAP when we're only getting read access */
3884 flags &= ~MDB_WRITEMAP;
3886 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3887 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3890 env->me_flags = flags |= MDB_ENV_ACTIVE;
3894 env->me_path = strdup(path);
3895 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3896 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3897 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3902 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3907 #define MDB_NO_SUCH_FILE ERROR_FILE_NOT_FOUND
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 #define MDB_NO_SUCH_FILE ENOENT
3920 if (F_ISSET(flags, MDB_RDONLY))
3923 oflags = O_RDWR | O_CREAT;
3925 env->me_fd = open(dpath, oflags, mode);
3927 if (env->me_fd == INVALID_HANDLE_VALUE) {
3929 if (F_ISSET(flags, MDB_RDONLY) && rc == MDB_NO_SUCH_FILE)
3934 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3935 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3936 env->me_mfd = env->me_fd;
3938 /* Synchronous fd for meta writes. Needed even with
3939 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3942 env->me_mfd = CreateFile(dpath, oflags,
3943 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3944 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3946 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3948 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3953 DPRINTF(("opened dbenv %p", (void *) env));
3955 rc = mdb_env_share_locks(env, &excl);
3961 mdb_env_close0(env, excl);
3969 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3971 mdb_env_close0(MDB_env *env, int excl)
3975 if (!(env->me_flags & MDB_ENV_ACTIVE))
3978 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3979 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3980 free(env->me_dbxs[i].md_name.mv_data);
3982 free(env->me_dbflags);
3985 free(env->me_dirty_list);
3986 mdb_midl_free(env->me_free_pgs);
3988 if (env->me_flags & MDB_ENV_TXKEY) {
3989 pthread_key_delete(env->me_txkey);
3991 /* Delete our key from the global list */
3992 for (i=0; i<mdb_tls_nkeys; i++)
3993 if (mdb_tls_keys[i] == env->me_txkey) {
3994 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4002 munmap(env->me_map, env->me_mapsize);
4004 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4005 (void) close(env->me_mfd);
4006 if (env->me_fd != INVALID_HANDLE_VALUE)
4007 (void) close(env->me_fd);
4009 pid_t pid = env->me_pid;
4010 /* Clearing readers is done in this function because
4011 * me_txkey with its destructor must be disabled first.
4013 for (i = env->me_numreaders; --i >= 0; )
4014 if (env->me_txns->mti_readers[i].mr_pid == pid)
4015 env->me_txns->mti_readers[i].mr_pid = 0;
4017 if (env->me_rmutex) {
4018 CloseHandle(env->me_rmutex);
4019 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4021 /* Windows automatically destroys the mutexes when
4022 * the last handle closes.
4024 #elif defined(MDB_USE_POSIX_SEM)
4025 if (env->me_rmutex != SEM_FAILED) {
4026 sem_close(env->me_rmutex);
4027 if (env->me_wmutex != SEM_FAILED)
4028 sem_close(env->me_wmutex);
4029 /* If we have the filelock: If we are the
4030 * only remaining user, clean up semaphores.
4033 mdb_env_excl_lock(env, &excl);
4035 sem_unlink(env->me_txns->mti_rmname);
4036 sem_unlink(env->me_txns->mti_wmname);
4040 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4042 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4045 /* Unlock the lockfile. Windows would have unlocked it
4046 * after closing anyway, but not necessarily at once.
4048 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4051 (void) close(env->me_lfd);
4054 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4058 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4060 MDB_txn *txn = NULL;
4066 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4070 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4073 /* Do the lock/unlock of the reader mutex before starting the
4074 * write txn. Otherwise other read txns could block writers.
4076 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4081 /* We must start the actual read txn after blocking writers */
4082 mdb_txn_reset0(txn, "reset-stage1");
4084 /* Temporarily block writers until we snapshot the meta pages */
4087 rc = mdb_txn_renew0(txn);
4089 UNLOCK_MUTEX_W(env);
4094 wsize = env->me_psize * 2;
4098 DO_WRITE(rc, fd, ptr, w2, len);
4102 } else if (len > 0) {
4108 /* Non-blocking or async handles are not supported */
4114 UNLOCK_MUTEX_W(env);
4119 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4121 if (wsize > MAX_WRITE)
4125 DO_WRITE(rc, fd, ptr, w2, len);
4129 } else if (len > 0) {
4146 mdb_env_copy(MDB_env *env, const char *path)
4150 HANDLE newfd = INVALID_HANDLE_VALUE;
4152 if (env->me_flags & MDB_NOSUBDIR) {
4153 lpath = (char *)path;
4156 len += sizeof(DATANAME);
4157 lpath = malloc(len);
4160 sprintf(lpath, "%s" DATANAME, path);
4163 /* The destination path must exist, but the destination file must not.
4164 * We don't want the OS to cache the writes, since the source data is
4165 * already in the OS cache.
4168 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4169 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4171 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
4177 if (newfd == INVALID_HANDLE_VALUE) {
4182 #ifdef F_NOCACHE /* __APPLE__ */
4183 rc = fcntl(newfd, F_NOCACHE, 1);
4190 rc = mdb_env_copyfd(env, newfd);
4193 if (!(env->me_flags & MDB_NOSUBDIR))
4195 if (newfd != INVALID_HANDLE_VALUE)
4196 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4203 mdb_env_close(MDB_env *env)
4210 VGMEMP_DESTROY(env);
4211 while ((dp = env->me_dpages) != NULL) {
4212 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4213 env->me_dpages = dp->mp_next;
4217 mdb_env_close0(env, 0);
4221 /** Compare two items pointing at aligned size_t's */
4223 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4225 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4226 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4229 /** Compare two items pointing at aligned int's */
4231 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4233 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4234 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4237 /** Compare two items pointing at ints of unknown alignment.
4238 * Nodes and keys are guaranteed to be 2-byte aligned.
4241 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4243 #if BYTE_ORDER == LITTLE_ENDIAN
4244 unsigned short *u, *c;
4247 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4248 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4251 } while(!x && u > (unsigned short *)a->mv_data);
4254 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4258 /** Compare two items lexically */
4260 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4267 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4273 diff = memcmp(a->mv_data, b->mv_data, len);
4274 return diff ? diff : len_diff<0 ? -1 : len_diff;
4277 /** Compare two items in reverse byte order */
4279 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4281 const unsigned char *p1, *p2, *p1_lim;
4285 p1_lim = (const unsigned char *)a->mv_data;
4286 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4287 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4289 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4295 while (p1 > p1_lim) {
4296 diff = *--p1 - *--p2;
4300 return len_diff<0 ? -1 : len_diff;
4303 /** Search for key within a page, using binary search.
4304 * Returns the smallest entry larger or equal to the key.
4305 * If exactp is non-null, stores whether the found entry was an exact match
4306 * in *exactp (1 or 0).
4307 * Updates the cursor index with the index of the found entry.
4308 * If no entry larger or equal to the key is found, returns NULL.
4311 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4313 unsigned int i = 0, nkeys;
4316 MDB_page *mp = mc->mc_pg[mc->mc_top];
4317 MDB_node *node = NULL;
4322 nkeys = NUMKEYS(mp);
4327 COPY_PGNO(pgno, mp->mp_pgno);
4328 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4329 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4336 low = IS_LEAF(mp) ? 0 : 1;
4338 cmp = mc->mc_dbx->md_cmp;
4340 /* Branch pages have no data, so if using integer keys,
4341 * alignment is guaranteed. Use faster mdb_cmp_int.
4343 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4344 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4351 nodekey.mv_size = mc->mc_db->md_pad;
4352 node = NODEPTR(mp, 0); /* fake */
4353 while (low <= high) {
4354 i = (low + high) >> 1;
4355 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4356 rc = cmp(key, &nodekey);
4357 DPRINTF(("found leaf index %u [%s], rc = %i",
4358 i, DKEY(&nodekey), rc));
4367 while (low <= high) {
4368 i = (low + high) >> 1;
4370 node = NODEPTR(mp, i);
4371 nodekey.mv_size = NODEKSZ(node);
4372 nodekey.mv_data = NODEKEY(node);
4374 rc = cmp(key, &nodekey);
4377 DPRINTF(("found leaf index %u [%s], rc = %i",
4378 i, DKEY(&nodekey), rc));
4380 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4381 i, DKEY(&nodekey), NODEPGNO(node), rc));
4392 if (rc > 0) { /* Found entry is less than the key. */
4393 i++; /* Skip to get the smallest entry larger than key. */
4395 node = NODEPTR(mp, i);
4398 *exactp = (rc == 0);
4399 /* store the key index */
4400 mc->mc_ki[mc->mc_top] = i;
4402 /* There is no entry larger or equal to the key. */
4405 /* nodeptr is fake for LEAF2 */
4411 mdb_cursor_adjust(MDB_cursor *mc, func)
4415 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4416 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4423 /** Pop a page off the top of the cursor's stack. */
4425 mdb_cursor_pop(MDB_cursor *mc)
4429 MDB_page *top = mc->mc_pg[mc->mc_top];
4435 DPRINTF(("popped page %"Z"u off db %u cursor %p", top->mp_pgno,
4436 mc->mc_dbi, (void *) mc));
4440 /** Push a page onto the top of the cursor's stack. */
4442 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4444 DPRINTF(("pushing page %"Z"u on db %u cursor %p", mp->mp_pgno,
4445 mc->mc_dbi, (void *) mc));
4447 if (mc->mc_snum >= CURSOR_STACK) {
4448 assert(mc->mc_snum < CURSOR_STACK);
4449 return MDB_CURSOR_FULL;
4452 mc->mc_top = mc->mc_snum++;
4453 mc->mc_pg[mc->mc_top] = mp;
4454 mc->mc_ki[mc->mc_top] = 0;
4459 /** Find the address of the page corresponding to a given page number.
4460 * @param[in] txn the transaction for this access.
4461 * @param[in] pgno the page number for the page to retrieve.
4462 * @param[out] ret address of a pointer where the page's address will be stored.
4463 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4464 * @return 0 on success, non-zero on failure.
4467 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4469 MDB_env *env = txn->mt_env;
4473 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4477 MDB_ID2L dl = tx2->mt_u.dirty_list;
4479 /* Spilled pages were dirtied in this txn and flushed
4480 * because the dirty list got full. Bring this page
4481 * back in from the map (but don't unspill it here,
4482 * leave that unless page_touch happens again).
4484 if (tx2->mt_spill_pgs) {
4485 x = mdb_midl_search(tx2->mt_spill_pgs, pgno);
4486 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pgno) {
4487 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4492 unsigned x = mdb_mid2l_search(dl, pgno);
4493 if (x <= dl[0].mid && dl[x].mid == pgno) {
4499 } while ((tx2 = tx2->mt_parent) != NULL);
4502 if (pgno < txn->mt_next_pgno) {
4504 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4506 DPRINTF(("page %"Z"u not found", pgno));
4508 return MDB_PAGE_NOTFOUND;
4518 /** Search for the page a given key should be in.
4519 * Pushes parent pages on the cursor stack. This function continues a
4520 * search on a cursor that has already been initialized. (Usually by
4521 * #mdb_page_search() but also by #mdb_node_move().)
4522 * @param[in,out] mc the cursor for this operation.
4523 * @param[in] key the key to search for. If NULL, search for the lowest
4524 * page. (This is used by #mdb_cursor_first().)
4525 * @param[in] modify If true, visited pages are updated with new page numbers.
4526 * @return 0 on success, non-zero on failure.
4529 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4531 MDB_page *mp = mc->mc_pg[mc->mc_top];
4536 while (IS_BRANCH(mp)) {
4540 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4541 assert(NUMKEYS(mp) > 1);
4542 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4544 if (key == NULL) /* Initialize cursor to first page. */
4546 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4547 /* cursor to last page */
4551 node = mdb_node_search(mc, key, &exact);
4553 i = NUMKEYS(mp) - 1;
4555 i = mc->mc_ki[mc->mc_top];
4564 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4565 assert(i < NUMKEYS(mp));
4566 node = NODEPTR(mp, i);
4568 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4571 mc->mc_ki[mc->mc_top] = i;
4572 if ((rc = mdb_cursor_push(mc, mp)))
4576 if ((rc = mdb_page_touch(mc)) != 0)
4578 mp = mc->mc_pg[mc->mc_top];
4583 DPRINTF(("internal error, index points to a %02X page!?",
4585 return MDB_CORRUPTED;
4588 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4589 key ? DKEY(key) : NULL));
4590 mc->mc_flags |= C_INITIALIZED;
4591 mc->mc_flags &= ~C_EOF;
4596 /** Search for the lowest key under the current branch page.
4597 * This just bypasses a NUMKEYS check in the current page
4598 * before calling mdb_page_search_root(), because the callers
4599 * are all in situations where the current page is known to
4603 mdb_page_search_lowest(MDB_cursor *mc)
4605 MDB_page *mp = mc->mc_pg[mc->mc_top];
4606 MDB_node *node = NODEPTR(mp, 0);
4609 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4612 mc->mc_ki[mc->mc_top] = 0;
4613 if ((rc = mdb_cursor_push(mc, mp)))
4615 return mdb_page_search_root(mc, NULL, 0);
4618 /** Search for the page a given key should be in.
4619 * Pushes parent pages on the cursor stack. This function just sets up
4620 * the search; it finds the root page for \b mc's database and sets this
4621 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4622 * called to complete the search.
4623 * @param[in,out] mc the cursor for this operation.
4624 * @param[in] key the key to search for. If NULL, search for the lowest
4625 * page. (This is used by #mdb_cursor_first().)
4626 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4627 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4628 * @return 0 on success, non-zero on failure.
4631 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4636 /* Make sure the txn is still viable, then find the root from
4637 * the txn's db table.
4639 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4640 DPUTS("transaction has failed, must abort");
4643 /* Make sure we're using an up-to-date root */
4644 if (mc->mc_dbi > MAIN_DBI) {
4645 if ((*mc->mc_dbflag & DB_STALE) ||
4646 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4648 unsigned char dbflag = 0;
4649 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4650 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4653 if (*mc->mc_dbflag & DB_STALE) {
4657 MDB_node *leaf = mdb_node_search(&mc2,
4658 &mc->mc_dbx->md_name, &exact);
4660 return MDB_NOTFOUND;
4661 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4664 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4666 /* The txn may not know this DBI, or another process may
4667 * have dropped and recreated the DB with other flags.
4669 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4670 return MDB_INCOMPATIBLE;
4671 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4673 if (flags & MDB_PS_MODIFY)
4675 *mc->mc_dbflag &= ~DB_STALE;
4676 *mc->mc_dbflag |= dbflag;
4679 root = mc->mc_db->md_root;
4681 if (root == P_INVALID) { /* Tree is empty. */
4682 DPUTS("tree is empty");
4683 return MDB_NOTFOUND;
4688 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4689 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4695 DPRINTF(("db %u root page %"Z"u has flags 0x%X",
4696 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags));
4698 if (flags & MDB_PS_MODIFY) {
4699 if ((rc = mdb_page_touch(mc)))
4703 if (flags & MDB_PS_ROOTONLY)
4706 return mdb_page_search_root(mc, key, flags);
4710 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4712 MDB_txn *txn = mc->mc_txn;
4713 pgno_t pg = mp->mp_pgno;
4714 unsigned x = 0, ovpages = mp->mp_pages;
4715 MDB_env *env = txn->mt_env;
4716 MDB_IDL sl = txn->mt_spill_pgs;
4719 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4720 /* If the page is dirty or on the spill list we just acquired it,
4721 * so we should give it back to our current free list, if any.
4722 * Otherwise put it onto the list of pages we freed in this txn.
4724 * Won't create me_pghead: me_pglast must be inited along with it.
4725 * Unsupported in nested txns: They would need to hide the page
4726 * range in ancestor txns' dirty and spilled lists.
4728 if (env->me_pghead &&
4730 ((mp->mp_flags & P_DIRTY) ||
4731 (sl && (x = mdb_midl_search(sl, pg)) <= sl[0] && sl[x] == pg)))
4735 MDB_ID2 *dl, ix, iy;
4736 rc = mdb_midl_need(&env->me_pghead, ovpages);
4739 if (!(mp->mp_flags & P_DIRTY)) {
4740 /* This page is no longer spilled */
4741 for (; x < sl[0]; x++)
4746 /* Remove from dirty list */
4747 dl = txn->mt_u.dirty_list;
4749 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4757 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4758 txn->mt_flags |= MDB_TXN_ERROR;
4759 return MDB_CORRUPTED;
4762 if (!(env->me_flags & MDB_WRITEMAP))
4763 mdb_dpage_free(env, mp);
4765 /* Insert in me_pghead */
4766 mop = env->me_pghead;
4767 j = mop[0] + ovpages;
4768 for (i = mop[0]; i && mop[i] < pg; i--)
4774 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4778 mc->mc_db->md_overflow_pages -= ovpages;
4782 /** Return the data associated with a given node.
4783 * @param[in] txn The transaction for this operation.
4784 * @param[in] leaf The node being read.
4785 * @param[out] data Updated to point to the node's data.
4786 * @return 0 on success, non-zero on failure.
4789 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4791 MDB_page *omp; /* overflow page */
4795 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4796 data->mv_size = NODEDSZ(leaf);
4797 data->mv_data = NODEDATA(leaf);
4801 /* Read overflow data.
4803 data->mv_size = NODEDSZ(leaf);
4804 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4805 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4806 DPRINTF(("read overflow page %"Z"u failed", pgno));
4809 data->mv_data = METADATA(omp);
4815 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4816 MDB_val *key, MDB_val *data)
4825 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4827 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4830 if (txn->mt_flags & MDB_TXN_ERROR)
4833 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4834 return MDB_BAD_VALSIZE;
4837 mdb_cursor_init(&mc, txn, dbi, &mx);
4838 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4841 /** Find a sibling for a page.
4842 * Replaces the page at the top of the cursor's stack with the
4843 * specified sibling, if one exists.
4844 * @param[in] mc The cursor for this operation.
4845 * @param[in] move_right Non-zero if the right sibling is requested,
4846 * otherwise the left sibling.
4847 * @return 0 on success, non-zero on failure.
4850 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4856 if (mc->mc_snum < 2) {
4857 return MDB_NOTFOUND; /* root has no siblings */
4861 DPRINTF(("parent page is page %"Z"u, index %u",
4862 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
4864 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4865 : (mc->mc_ki[mc->mc_top] == 0)) {
4866 DPRINTF(("no more keys left, moving to %s sibling",
4867 move_right ? "right" : "left"));
4868 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4869 /* undo cursor_pop before returning */
4876 mc->mc_ki[mc->mc_top]++;
4878 mc->mc_ki[mc->mc_top]--;
4879 DPRINTF(("just moving to %s index key %u",
4880 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
4882 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4884 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4885 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4888 mdb_cursor_push(mc, mp);
4890 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4895 /** Move the cursor to the next data item. */
4897 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4903 if (mc->mc_flags & C_EOF) {
4904 return MDB_NOTFOUND;
4907 assert(mc->mc_flags & C_INITIALIZED);
4909 mp = mc->mc_pg[mc->mc_top];
4911 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4912 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4913 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4914 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4915 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4916 if (op != MDB_NEXT || rc != MDB_NOTFOUND)
4920 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4921 if (op == MDB_NEXT_DUP)
4922 return MDB_NOTFOUND;
4926 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
4928 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4929 DPUTS("=====> move to next sibling page");
4930 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
4931 mc->mc_flags |= C_EOF;
4934 mp = mc->mc_pg[mc->mc_top];
4935 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
4937 mc->mc_ki[mc->mc_top]++;
4939 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
4940 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
4943 key->mv_size = mc->mc_db->md_pad;
4944 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4948 assert(IS_LEAF(mp));
4949 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4951 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4952 mdb_xcursor_init1(mc, leaf);
4955 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4958 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4959 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4960 if (rc != MDB_SUCCESS)
4965 MDB_GET_KEY(leaf, key);
4969 /** Move the cursor to the previous data item. */
4971 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4977 assert(mc->mc_flags & C_INITIALIZED);
4979 mp = mc->mc_pg[mc->mc_top];
4981 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4982 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4983 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4984 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4985 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4986 if (op != MDB_PREV || rc != MDB_NOTFOUND)
4989 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
4990 if (op == MDB_PREV_DUP)
4991 return MDB_NOTFOUND;
4996 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
4998 if (mc->mc_ki[mc->mc_top] == 0) {
4999 DPUTS("=====> move to prev sibling page");
5000 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5003 mp = mc->mc_pg[mc->mc_top];
5004 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5005 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5007 mc->mc_ki[mc->mc_top]--;
5009 mc->mc_flags &= ~C_EOF;
5011 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5012 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5015 key->mv_size = mc->mc_db->md_pad;
5016 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5020 assert(IS_LEAF(mp));
5021 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5023 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5024 mdb_xcursor_init1(mc, leaf);
5027 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5030 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5031 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5032 if (rc != MDB_SUCCESS)
5037 MDB_GET_KEY(leaf, key);
5041 /** Set the cursor on a specific data item. */
5043 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5044 MDB_cursor_op op, int *exactp)
5048 MDB_node *leaf = NULL;
5053 assert(key->mv_size > 0);
5056 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5058 /* See if we're already on the right page */
5059 if (mc->mc_flags & C_INITIALIZED) {
5062 mp = mc->mc_pg[mc->mc_top];
5064 mc->mc_ki[mc->mc_top] = 0;
5065 return MDB_NOTFOUND;
5067 if (mp->mp_flags & P_LEAF2) {
5068 nodekey.mv_size = mc->mc_db->md_pad;
5069 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5071 leaf = NODEPTR(mp, 0);
5072 MDB_GET_KEY(leaf, &nodekey);
5074 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5076 /* Probably happens rarely, but first node on the page
5077 * was the one we wanted.
5079 mc->mc_ki[mc->mc_top] = 0;
5086 unsigned int nkeys = NUMKEYS(mp);
5088 if (mp->mp_flags & P_LEAF2) {
5089 nodekey.mv_data = LEAF2KEY(mp,
5090 nkeys-1, nodekey.mv_size);
5092 leaf = NODEPTR(mp, nkeys-1);
5093 MDB_GET_KEY(leaf, &nodekey);
5095 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5097 /* last node was the one we wanted */
5098 mc->mc_ki[mc->mc_top] = nkeys-1;
5104 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5105 /* This is definitely the right page, skip search_page */
5106 if (mp->mp_flags & P_LEAF2) {
5107 nodekey.mv_data = LEAF2KEY(mp,
5108 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5110 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5111 MDB_GET_KEY(leaf, &nodekey);
5113 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5115 /* current node was the one we wanted */
5125 /* If any parents have right-sibs, search.
5126 * Otherwise, there's nothing further.
5128 for (i=0; i<mc->mc_top; i++)
5130 NUMKEYS(mc->mc_pg[i])-1)
5132 if (i == mc->mc_top) {
5133 /* There are no other pages */
5134 mc->mc_ki[mc->mc_top] = nkeys;
5135 return MDB_NOTFOUND;
5139 /* There are no other pages */
5140 mc->mc_ki[mc->mc_top] = 0;
5141 return MDB_NOTFOUND;
5145 rc = mdb_page_search(mc, key, 0);
5146 if (rc != MDB_SUCCESS)
5149 mp = mc->mc_pg[mc->mc_top];
5150 assert(IS_LEAF(mp));
5153 leaf = mdb_node_search(mc, key, exactp);
5154 if (exactp != NULL && !*exactp) {
5155 /* MDB_SET specified and not an exact match. */
5156 return MDB_NOTFOUND;
5160 DPUTS("===> inexact leaf not found, goto sibling");
5161 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5162 return rc; /* no entries matched */
5163 mp = mc->mc_pg[mc->mc_top];
5164 assert(IS_LEAF(mp));
5165 leaf = NODEPTR(mp, 0);
5169 mc->mc_flags |= C_INITIALIZED;
5170 mc->mc_flags &= ~C_EOF;
5173 key->mv_size = mc->mc_db->md_pad;
5174 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5178 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5179 mdb_xcursor_init1(mc, leaf);
5182 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5183 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5184 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5187 if (op == MDB_GET_BOTH) {
5193 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5194 if (rc != MDB_SUCCESS)
5197 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5199 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5201 rc = mc->mc_dbx->md_dcmp(data, &d2);
5203 if (op == MDB_GET_BOTH || rc > 0)
5204 return MDB_NOTFOUND;
5209 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5210 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5215 /* The key already matches in all other cases */
5216 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5217 MDB_GET_KEY(leaf, key);
5218 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5223 /** Move the cursor to the first item in the database. */
5225 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5231 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5233 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5234 rc = mdb_page_search(mc, NULL, 0);
5235 if (rc != MDB_SUCCESS)
5238 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5240 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5241 mc->mc_flags |= C_INITIALIZED;
5242 mc->mc_flags &= ~C_EOF;
5244 mc->mc_ki[mc->mc_top] = 0;
5246 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5247 key->mv_size = mc->mc_db->md_pad;
5248 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5253 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5254 mdb_xcursor_init1(mc, leaf);
5255 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5259 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5263 MDB_GET_KEY(leaf, key);
5267 /** Move the cursor to the last item in the database. */
5269 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5275 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5277 if (!(mc->mc_flags & C_EOF)) {
5279 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5282 lkey.mv_size = MDB_MAXKEYSIZE+1;
5283 lkey.mv_data = NULL;
5284 rc = mdb_page_search(mc, &lkey, 0);
5285 if (rc != MDB_SUCCESS)
5288 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5291 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5292 mc->mc_flags |= C_INITIALIZED|C_EOF;
5293 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5295 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5296 key->mv_size = mc->mc_db->md_pad;
5297 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5302 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5303 mdb_xcursor_init1(mc, leaf);
5304 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5308 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5313 MDB_GET_KEY(leaf, key);
5318 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5323 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5327 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5331 case MDB_GET_CURRENT:
5332 if (!(mc->mc_flags & C_INITIALIZED)) {
5335 MDB_page *mp = mc->mc_pg[mc->mc_top];
5337 mc->mc_ki[mc->mc_top] = 0;
5343 key->mv_size = mc->mc_db->md_pad;
5344 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5346 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5347 MDB_GET_KEY(leaf, key);
5349 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5350 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5352 rc = mdb_node_read(mc->mc_txn, leaf, data);
5359 case MDB_GET_BOTH_RANGE:
5364 if (mc->mc_xcursor == NULL) {
5365 rc = MDB_INCOMPATIBLE;
5374 } else if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
5375 rc = MDB_BAD_VALSIZE;
5376 } else if (op == MDB_SET_RANGE)
5377 rc = mdb_cursor_set(mc, key, data, op, NULL);
5379 rc = mdb_cursor_set(mc, key, data, op, &exact);
5381 case MDB_GET_MULTIPLE:
5382 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5386 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5387 rc = MDB_INCOMPATIBLE;
5391 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5392 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5395 case MDB_NEXT_MULTIPLE:
5400 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5401 rc = MDB_INCOMPATIBLE;
5404 if (!(mc->mc_flags & C_INITIALIZED))
5405 rc = mdb_cursor_first(mc, key, data);
5407 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5408 if (rc == MDB_SUCCESS) {
5409 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5412 mx = &mc->mc_xcursor->mx_cursor;
5413 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5415 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5416 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5424 case MDB_NEXT_NODUP:
5425 if (!(mc->mc_flags & C_INITIALIZED))
5426 rc = mdb_cursor_first(mc, key, data);
5428 rc = mdb_cursor_next(mc, key, data, op);
5432 case MDB_PREV_NODUP:
5433 if (!(mc->mc_flags & C_INITIALIZED)) {
5434 rc = mdb_cursor_last(mc, key, data);
5437 mc->mc_flags |= C_INITIALIZED;
5438 mc->mc_ki[mc->mc_top]++;
5440 rc = mdb_cursor_prev(mc, key, data, op);
5443 rc = mdb_cursor_first(mc, key, data);
5446 mfunc = mdb_cursor_first;
5448 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5452 if (mc->mc_xcursor == NULL) {
5453 rc = MDB_INCOMPATIBLE;
5456 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5460 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5463 rc = mdb_cursor_last(mc, key, data);
5466 mfunc = mdb_cursor_last;
5469 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5477 /** Touch all the pages in the cursor stack.
5478 * Makes sure all the pages are writable, before attempting a write operation.
5479 * @param[in] mc The cursor to operate on.
5482 mdb_cursor_touch(MDB_cursor *mc)
5486 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5489 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5490 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5493 *mc->mc_dbflag |= DB_DIRTY;
5495 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5496 rc = mdb_page_touch(mc);
5500 mc->mc_top = mc->mc_snum-1;
5504 /** Do not spill pages to disk if txn is getting full, may fail instead */
5505 #define MDB_NOSPILL 0x8000
5508 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5511 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5512 MDB_node *leaf = NULL;
5513 MDB_val xdata, *rdata, dkey;
5516 int do_sub = 0, insert = 0;
5517 unsigned int mcount = 0, dcount = 0, nospill;
5521 char dbuf[MDB_MAXKEYSIZE+1];
5522 unsigned int nflags;
5525 /* Check this first so counter will always be zero on any
5528 if (flags & MDB_MULTIPLE) {
5529 dcount = data[1].mv_size;
5530 data[1].mv_size = 0;
5531 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5532 return MDB_INCOMPATIBLE;
5535 nospill = flags & MDB_NOSPILL;
5536 flags &= ~MDB_NOSPILL;
5538 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5539 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5541 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5542 return MDB_BAD_VALSIZE;
5544 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5545 return MDB_BAD_VALSIZE;
5547 #if SIZE_MAX > MAXDATASIZE
5548 if (data->mv_size > MAXDATASIZE)
5549 return MDB_BAD_VALSIZE;
5552 DPRINTF(("==> put db %u key [%s], size %"Z"u, data size %"Z"u",
5553 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size));
5557 if (flags == MDB_CURRENT) {
5558 if (!(mc->mc_flags & C_INITIALIZED))
5561 } else if (mc->mc_db->md_root == P_INVALID) {
5562 /* new database, cursor has nothing to point to */
5564 mc->mc_flags &= ~C_INITIALIZED;
5569 if (flags & MDB_APPEND) {
5571 rc = mdb_cursor_last(mc, &k2, &d2);
5573 rc = mc->mc_dbx->md_cmp(key, &k2);
5576 mc->mc_ki[mc->mc_top]++;
5578 /* new key is <= last key */
5583 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5585 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5586 DPRINTF(("duplicate key [%s]", DKEY(key)));
5588 return MDB_KEYEXIST;
5590 if (rc && rc != MDB_NOTFOUND)
5594 /* Cursor is positioned, check for room in the dirty list */
5596 if (flags & MDB_MULTIPLE) {
5598 xdata.mv_size = data->mv_size * dcount;
5602 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5606 if (rc == MDB_NO_ROOT) {
5608 /* new database, write a root leaf page */
5609 DPUTS("allocating new root leaf page");
5610 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5613 mdb_cursor_push(mc, np);
5614 mc->mc_db->md_root = np->mp_pgno;
5615 mc->mc_db->md_depth++;
5616 *mc->mc_dbflag |= DB_DIRTY;
5617 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5619 np->mp_flags |= P_LEAF2;
5620 mc->mc_flags |= C_INITIALIZED;
5622 /* make sure all cursor pages are writable */
5623 rc2 = mdb_cursor_touch(mc);
5628 /* The key already exists */
5629 if (rc == MDB_SUCCESS) {
5630 /* there's only a key anyway, so this is a no-op */
5631 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5632 unsigned int ksize = mc->mc_db->md_pad;
5633 if (key->mv_size != ksize)
5634 return MDB_BAD_VALSIZE;
5635 if (flags == MDB_CURRENT) {
5636 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5637 memcpy(ptr, key->mv_data, ksize);
5642 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5645 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5646 /* Was a single item before, must convert now */
5648 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5649 /* Just overwrite the current item */
5650 if (flags == MDB_CURRENT)
5653 dkey.mv_size = NODEDSZ(leaf);
5654 dkey.mv_data = NODEDATA(leaf);
5655 #if UINT_MAX < SIZE_MAX
5656 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5657 #ifdef MISALIGNED_OK
5658 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5660 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5663 /* if data matches, skip it */
5664 if (!mc->mc_dbx->md_dcmp(data, &dkey)) {
5665 if (flags & MDB_NODUPDATA)
5667 else if (flags & MDB_MULTIPLE)
5674 /* create a fake page for the dup items */
5675 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5676 dkey.mv_data = dbuf;
5677 fp = (MDB_page *)&pbuf;
5678 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5679 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5680 fp->mp_lower = PAGEHDRSZ;
5681 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5682 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5683 fp->mp_flags |= P_LEAF2;
5684 fp->mp_pad = data->mv_size;
5685 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5687 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5688 (dkey.mv_size & 1) + (data->mv_size & 1);
5690 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5693 xdata.mv_size = fp->mp_upper;
5698 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5699 /* See if we need to convert from fake page to subDB */
5701 unsigned int offset;
5705 fp = NODEDATA(leaf);
5706 if (flags == MDB_CURRENT) {
5708 fp->mp_flags |= P_DIRTY;
5709 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5710 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5714 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5715 offset = fp->mp_pad;
5716 if (SIZELEFT(fp) >= offset)
5718 offset *= 4; /* space for 4 more */
5720 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5722 offset += offset & 1;
5723 fp_flags = fp->mp_flags;
5724 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5725 offset >= mc->mc_txn->mt_env->me_nodemax) {
5726 /* yes, convert it */
5728 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5729 dummy.md_pad = fp->mp_pad;
5730 dummy.md_flags = MDB_DUPFIXED;
5731 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5732 dummy.md_flags |= MDB_INTEGERKEY;
5735 dummy.md_branch_pages = 0;
5736 dummy.md_leaf_pages = 1;
5737 dummy.md_overflow_pages = 0;
5738 dummy.md_entries = NUMKEYS(fp);
5740 xdata.mv_size = sizeof(MDB_db);
5741 xdata.mv_data = &dummy;
5742 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5744 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5745 flags |= F_DUPDATA|F_SUBDATA;
5746 dummy.md_root = mp->mp_pgno;
5747 fp_flags &= ~P_SUBP;
5749 /* no, just grow it */
5751 xdata.mv_size = NODEDSZ(leaf) + offset;
5752 xdata.mv_data = &pbuf;
5753 mp = (MDB_page *)&pbuf;
5754 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5757 mp->mp_flags = fp_flags | P_DIRTY;
5758 mp->mp_pad = fp->mp_pad;
5759 mp->mp_lower = fp->mp_lower;
5760 mp->mp_upper = fp->mp_upper + offset;
5762 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5764 nsize = NODEDSZ(leaf) - fp->mp_upper;
5765 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5766 for (i=0; i<NUMKEYS(fp); i++)
5767 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5769 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5773 /* data is on sub-DB, just store it */
5774 flags |= F_DUPDATA|F_SUBDATA;
5778 /* overflow page overwrites need special handling */
5779 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5782 unsigned psize = mc->mc_txn->mt_env->me_psize;
5783 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5785 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5786 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5788 ovpages = omp->mp_pages;
5790 /* Is the ov page large enough? */
5791 if (ovpages >= dpages) {
5792 if (!(omp->mp_flags & P_DIRTY) &&
5793 (level || (mc->mc_txn->mt_env->me_flags & MDB_WRITEMAP)))
5795 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5798 level = 0; /* dirty in this txn or clean */
5801 if (omp->mp_flags & P_DIRTY) {
5802 /* yes, overwrite it. Note in this case we don't
5803 * bother to try shrinking the page if the new data
5804 * is smaller than the overflow threshold.
5807 /* It is writable only in a parent txn */
5808 size_t sz = (size_t) psize * ovpages, off;
5809 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5815 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5816 if (!(flags & MDB_RESERVE)) {
5817 /* Copy end of page, adjusting alignment so
5818 * compiler may copy words instead of bytes.
5820 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5821 memcpy((size_t *)((char *)np + off),
5822 (size_t *)((char *)omp + off), sz - off);
5825 memcpy(np, omp, sz); /* Copy beginning of page */
5828 SETDSZ(leaf, data->mv_size);
5829 if (F_ISSET(flags, MDB_RESERVE))
5830 data->mv_data = METADATA(omp);
5832 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5836 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5838 } else if (NODEDSZ(leaf) == data->mv_size) {
5839 /* same size, just replace it. Note that we could
5840 * also reuse this node if the new data is smaller,
5841 * but instead we opt to shrink the node in that case.
5843 if (F_ISSET(flags, MDB_RESERVE))
5844 data->mv_data = NODEDATA(leaf);
5845 else if (data->mv_size)
5846 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5848 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5851 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5852 mc->mc_db->md_entries--;
5854 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5861 nflags = flags & NODE_ADD_FLAGS;
5862 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5863 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5864 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5865 nflags &= ~MDB_APPEND;
5867 nflags |= MDB_SPLIT_REPLACE;
5868 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5870 /* There is room already in this leaf page. */
5871 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5872 if (rc == 0 && !do_sub && insert) {
5873 /* Adjust other cursors pointing to mp */
5874 MDB_cursor *m2, *m3;
5875 MDB_dbi dbi = mc->mc_dbi;
5876 unsigned i = mc->mc_top;
5877 MDB_page *mp = mc->mc_pg[i];
5879 if (mc->mc_flags & C_SUB)
5882 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5883 if (mc->mc_flags & C_SUB)
5884 m3 = &m2->mc_xcursor->mx_cursor;
5887 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5888 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5895 if (rc != MDB_SUCCESS)
5896 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5898 /* Now store the actual data in the child DB. Note that we're
5899 * storing the user data in the keys field, so there are strict
5900 * size limits on dupdata. The actual data fields of the child
5901 * DB are all zero size.
5908 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5909 if (flags & MDB_CURRENT) {
5910 xflags = MDB_CURRENT|MDB_NOSPILL;
5912 mdb_xcursor_init1(mc, leaf);
5913 xflags = (flags & MDB_NODUPDATA) ?
5914 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
5916 /* converted, write the original data first */
5918 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5922 /* Adjust other cursors pointing to mp */
5924 unsigned i = mc->mc_top;
5925 MDB_page *mp = mc->mc_pg[i];
5927 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5928 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5929 if (!(m2->mc_flags & C_INITIALIZED)) continue;
5930 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5931 mdb_xcursor_init1(m2, leaf);
5935 /* we've done our job */
5938 if (flags & MDB_APPENDDUP)
5939 xflags |= MDB_APPEND;
5940 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5941 if (flags & F_SUBDATA) {
5942 void *db = NODEDATA(leaf);
5943 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5946 /* sub-writes might have failed so check rc again.
5947 * Don't increment count if we just replaced an existing item.
5949 if (!rc && !(flags & MDB_CURRENT))
5950 mc->mc_db->md_entries++;
5951 if (flags & MDB_MULTIPLE) {
5955 /* let caller know how many succeeded, if any */
5956 data[1].mv_size = mcount;
5957 if (mcount < dcount) {
5958 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5959 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5966 /* If we succeeded and the key didn't exist before, make sure
5967 * the cursor is marked valid.
5970 mc->mc_flags |= C_INITIALIZED;
5975 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5980 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5981 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5983 if (!(mc->mc_flags & C_INITIALIZED))
5986 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
5988 flags &= ~MDB_NOSPILL; /* TODO: Or change (flags != MDB_NODUPDATA) to ~(flags & MDB_NODUPDATA), not looking at the logic of that code just now */
5990 rc = mdb_cursor_touch(mc);
5994 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5996 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5997 if (!(flags & MDB_NODUPDATA)) {
5998 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5999 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6001 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6002 /* If sub-DB still has entries, we're done */
6003 if (mc->mc_xcursor->mx_db.md_entries) {
6004 if (leaf->mn_flags & F_SUBDATA) {
6005 /* update subDB info */
6006 void *db = NODEDATA(leaf);
6007 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6010 /* shrink fake page */
6011 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6012 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6013 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6014 /* fix other sub-DB cursors pointed at this fake page */
6015 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6016 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6017 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
6018 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6019 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6022 mc->mc_db->md_entries--;
6025 /* otherwise fall thru and delete the sub-DB */
6028 if (leaf->mn_flags & F_SUBDATA) {
6029 /* add all the child DB's pages to the free list */
6030 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6031 if (rc == MDB_SUCCESS) {
6032 mc->mc_db->md_entries -=
6033 mc->mc_xcursor->mx_db.md_entries;
6038 return mdb_cursor_del0(mc, leaf);
6041 /** Allocate and initialize new pages for a database.
6042 * @param[in] mc a cursor on the database being added to.
6043 * @param[in] flags flags defining what type of page is being allocated.
6044 * @param[in] num the number of pages to allocate. This is usually 1,
6045 * unless allocating overflow pages for a large record.
6046 * @param[out] mp Address of a page, or NULL on failure.
6047 * @return 0 on success, non-zero on failure.
6050 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6055 if ((rc = mdb_page_alloc(mc, num, &np)))
6057 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6058 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6059 np->mp_flags = flags | P_DIRTY;
6060 np->mp_lower = PAGEHDRSZ;
6061 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6064 mc->mc_db->md_branch_pages++;
6065 else if (IS_LEAF(np))
6066 mc->mc_db->md_leaf_pages++;
6067 else if (IS_OVERFLOW(np)) {
6068 mc->mc_db->md_overflow_pages += num;
6076 /** Calculate the size of a leaf node.
6077 * The size depends on the environment's page size; if a data item
6078 * is too large it will be put onto an overflow page and the node
6079 * size will only include the key and not the data. Sizes are always
6080 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6081 * of the #MDB_node headers.
6082 * @param[in] env The environment handle.
6083 * @param[in] key The key for the node.
6084 * @param[in] data The data for the node.
6085 * @return The number of bytes needed to store the node.
6088 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6092 sz = LEAFSIZE(key, data);
6093 if (sz >= env->me_nodemax) {
6094 /* put on overflow page */
6095 sz -= data->mv_size - sizeof(pgno_t);
6099 return sz + sizeof(indx_t);
6102 /** Calculate the size of a branch node.
6103 * The size should depend on the environment's page size but since
6104 * we currently don't support spilling large keys onto overflow
6105 * pages, it's simply the size of the #MDB_node header plus the
6106 * size of the key. Sizes are always rounded up to an even number
6107 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6108 * @param[in] env The environment handle.
6109 * @param[in] key The key for the node.
6110 * @return The number of bytes needed to store the node.
6113 mdb_branch_size(MDB_env *env, MDB_val *key)
6118 if (sz >= env->me_nodemax) {
6119 /* put on overflow page */
6120 /* not implemented */
6121 /* sz -= key->size - sizeof(pgno_t); */
6124 return sz + sizeof(indx_t);
6127 /** Add a node to the page pointed to by the cursor.
6128 * @param[in] mc The cursor for this operation.
6129 * @param[in] indx The index on the page where the new node should be added.
6130 * @param[in] key The key for the new node.
6131 * @param[in] data The data for the new node, if any.
6132 * @param[in] pgno The page number, if adding a branch node.
6133 * @param[in] flags Flags for the node.
6134 * @return 0 on success, non-zero on failure. Possible errors are:
6136 * <li>ENOMEM - failed to allocate overflow pages for the node.
6137 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6138 * should never happen since all callers already calculate the
6139 * page's free space before calling this function.
6143 mdb_node_add(MDB_cursor *mc, indx_t indx,
6144 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6147 size_t node_size = NODESIZE;
6150 MDB_page *mp = mc->mc_pg[mc->mc_top];
6151 MDB_page *ofp = NULL; /* overflow page */
6154 assert(mp->mp_upper >= mp->mp_lower);
6156 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6157 IS_LEAF(mp) ? "leaf" : "branch",
6158 IS_SUBP(mp) ? "sub-" : "",
6159 mp->mp_pgno, indx, data ? data->mv_size : 0,
6160 key ? key->mv_size : 0, key ? DKEY(key) : NULL));
6163 /* Move higher keys up one slot. */
6164 int ksize = mc->mc_db->md_pad, dif;
6165 char *ptr = LEAF2KEY(mp, indx, ksize);
6166 dif = NUMKEYS(mp) - indx;
6168 memmove(ptr+ksize, ptr, dif*ksize);
6169 /* insert new key */
6170 memcpy(ptr, key->mv_data, ksize);
6172 /* Just using these for counting */
6173 mp->mp_lower += sizeof(indx_t);
6174 mp->mp_upper -= ksize - sizeof(indx_t);
6179 node_size += key->mv_size;
6183 if (F_ISSET(flags, F_BIGDATA)) {
6184 /* Data already on overflow page. */
6185 node_size += sizeof(pgno_t);
6186 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
6187 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6189 /* Put data on overflow page. */
6190 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6191 data->mv_size, node_size+data->mv_size));
6192 node_size += sizeof(pgno_t);
6193 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6195 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6198 node_size += data->mv_size;
6201 node_size += node_size & 1;
6203 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
6204 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6205 mp->mp_pgno, NUMKEYS(mp)));
6206 DPRINTF(("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
6207 mp->mp_upper - mp->mp_lower));
6208 DPRINTF(("node size = %"Z"u", node_size));
6209 return MDB_PAGE_FULL;
6212 /* Move higher pointers up one slot. */
6213 for (i = NUMKEYS(mp); i > indx; i--)
6214 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6216 /* Adjust free space offsets. */
6217 ofs = mp->mp_upper - node_size;
6218 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6219 mp->mp_ptrs[indx] = ofs;
6221 mp->mp_lower += sizeof(indx_t);
6223 /* Write the node data. */
6224 node = NODEPTR(mp, indx);
6225 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6226 node->mn_flags = flags;
6228 SETDSZ(node,data->mv_size);
6233 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6238 if (F_ISSET(flags, F_BIGDATA))
6239 memcpy(node->mn_data + key->mv_size, data->mv_data,
6241 else if (F_ISSET(flags, MDB_RESERVE))
6242 data->mv_data = node->mn_data + key->mv_size;
6244 memcpy(node->mn_data + key->mv_size, data->mv_data,
6247 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6249 if (F_ISSET(flags, MDB_RESERVE))
6250 data->mv_data = METADATA(ofp);
6252 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6259 /** Delete the specified node from a page.
6260 * @param[in] mp The page to operate on.
6261 * @param[in] indx The index of the node to delete.
6262 * @param[in] ksize The size of a node. Only used if the page is
6263 * part of a #MDB_DUPFIXED database.
6266 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6269 indx_t i, j, numkeys, ptr;
6276 COPY_PGNO(pgno, mp->mp_pgno);
6277 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6278 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6281 assert(indx < NUMKEYS(mp));
6284 int x = NUMKEYS(mp) - 1 - indx;
6285 base = LEAF2KEY(mp, indx, ksize);
6287 memmove(base, base + ksize, x * ksize);
6288 mp->mp_lower -= sizeof(indx_t);
6289 mp->mp_upper += ksize - sizeof(indx_t);
6293 node = NODEPTR(mp, indx);
6294 sz = NODESIZE + node->mn_ksize;
6296 if (F_ISSET(node->mn_flags, F_BIGDATA))
6297 sz += sizeof(pgno_t);
6299 sz += NODEDSZ(node);
6303 ptr = mp->mp_ptrs[indx];
6304 numkeys = NUMKEYS(mp);
6305 for (i = j = 0; i < numkeys; i++) {
6307 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6308 if (mp->mp_ptrs[i] < ptr)
6309 mp->mp_ptrs[j] += sz;
6314 base = (char *)mp + mp->mp_upper;
6315 memmove(base + sz, base, ptr - mp->mp_upper);
6317 mp->mp_lower -= sizeof(indx_t);
6321 /** Compact the main page after deleting a node on a subpage.
6322 * @param[in] mp The main page to operate on.
6323 * @param[in] indx The index of the subpage on the main page.
6326 mdb_node_shrink(MDB_page *mp, indx_t indx)
6333 indx_t i, numkeys, ptr;
6335 node = NODEPTR(mp, indx);
6336 sp = (MDB_page *)NODEDATA(node);
6337 osize = NODEDSZ(node);
6339 delta = sp->mp_upper - sp->mp_lower;
6340 SETDSZ(node, osize - delta);
6341 xp = (MDB_page *)((char *)sp + delta);
6343 /* shift subpage upward */
6345 nsize = NUMKEYS(sp) * sp->mp_pad;
6346 memmove(METADATA(xp), METADATA(sp), nsize);
6349 nsize = osize - sp->mp_upper;
6350 numkeys = NUMKEYS(sp);
6351 for (i=numkeys-1; i>=0; i--)
6352 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6354 xp->mp_upper = sp->mp_lower;
6355 xp->mp_lower = sp->mp_lower;
6356 xp->mp_flags = sp->mp_flags;
6357 xp->mp_pad = sp->mp_pad;
6358 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6360 /* shift lower nodes upward */
6361 ptr = mp->mp_ptrs[indx];
6362 numkeys = NUMKEYS(mp);
6363 for (i = 0; i < numkeys; i++) {
6364 if (mp->mp_ptrs[i] <= ptr)
6365 mp->mp_ptrs[i] += delta;
6368 base = (char *)mp + mp->mp_upper;
6369 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6370 mp->mp_upper += delta;
6373 /** Initial setup of a sorted-dups cursor.
6374 * Sorted duplicates are implemented as a sub-database for the given key.
6375 * The duplicate data items are actually keys of the sub-database.
6376 * Operations on the duplicate data items are performed using a sub-cursor
6377 * initialized when the sub-database is first accessed. This function does
6378 * the preliminary setup of the sub-cursor, filling in the fields that
6379 * depend only on the parent DB.
6380 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6383 mdb_xcursor_init0(MDB_cursor *mc)
6385 MDB_xcursor *mx = mc->mc_xcursor;
6387 mx->mx_cursor.mc_xcursor = NULL;
6388 mx->mx_cursor.mc_txn = mc->mc_txn;
6389 mx->mx_cursor.mc_db = &mx->mx_db;
6390 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6391 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
6392 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6393 mx->mx_cursor.mc_snum = 0;
6394 mx->mx_cursor.mc_top = 0;
6395 mx->mx_cursor.mc_flags = C_SUB;
6396 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6397 mx->mx_dbx.md_dcmp = NULL;
6398 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6401 /** Final setup of a sorted-dups cursor.
6402 * Sets up the fields that depend on the data from the main cursor.
6403 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6404 * @param[in] node The data containing the #MDB_db record for the
6405 * sorted-dup database.
6408 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6410 MDB_xcursor *mx = mc->mc_xcursor;
6412 if (node->mn_flags & F_SUBDATA) {
6413 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6414 mx->mx_cursor.mc_pg[0] = 0;
6415 mx->mx_cursor.mc_snum = 0;
6416 mx->mx_cursor.mc_flags = C_SUB;
6418 MDB_page *fp = NODEDATA(node);
6419 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6420 mx->mx_db.md_flags = 0;
6421 mx->mx_db.md_depth = 1;
6422 mx->mx_db.md_branch_pages = 0;
6423 mx->mx_db.md_leaf_pages = 1;
6424 mx->mx_db.md_overflow_pages = 0;
6425 mx->mx_db.md_entries = NUMKEYS(fp);
6426 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6427 mx->mx_cursor.mc_snum = 1;
6428 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6429 mx->mx_cursor.mc_top = 0;
6430 mx->mx_cursor.mc_pg[0] = fp;
6431 mx->mx_cursor.mc_ki[0] = 0;
6432 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6433 mx->mx_db.md_flags = MDB_DUPFIXED;
6434 mx->mx_db.md_pad = fp->mp_pad;
6435 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6436 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6439 DPRINTF(("Sub-db %u for db %u root page %"Z"u", mx->mx_cursor.mc_dbi, mc->mc_dbi,
6440 mx->mx_db.md_root));
6441 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
6443 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
6444 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
6445 #if UINT_MAX < SIZE_MAX
6446 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6447 #ifdef MISALIGNED_OK
6448 mx->mx_dbx.md_cmp = mdb_cmp_long;
6450 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6455 /** Initialize a cursor for a given transaction and database. */
6457 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6460 mc->mc_backup = NULL;
6463 mc->mc_db = &txn->mt_dbs[dbi];
6464 mc->mc_dbx = &txn->mt_dbxs[dbi];
6465 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6470 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6472 mc->mc_xcursor = mx;
6473 mdb_xcursor_init0(mc);
6475 mc->mc_xcursor = NULL;
6477 if (*mc->mc_dbflag & DB_STALE) {
6478 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6483 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6486 size_t size = sizeof(MDB_cursor);
6488 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6491 if (txn->mt_flags & MDB_TXN_ERROR)
6494 /* Allow read access to the freelist */
6495 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6498 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6499 size += sizeof(MDB_xcursor);
6501 if ((mc = malloc(size)) != NULL) {
6502 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6503 if (txn->mt_cursors) {
6504 mc->mc_next = txn->mt_cursors[dbi];
6505 txn->mt_cursors[dbi] = mc;
6506 mc->mc_flags |= C_UNTRACK;
6518 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6520 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6523 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6526 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6530 /* Return the count of duplicate data items for the current key */
6532 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6536 if (mc == NULL || countp == NULL)
6539 if (mc->mc_xcursor == NULL)
6540 return MDB_INCOMPATIBLE;
6542 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6543 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6546 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6549 *countp = mc->mc_xcursor->mx_db.md_entries;
6555 mdb_cursor_close(MDB_cursor *mc)
6557 if (mc && !mc->mc_backup) {
6558 /* remove from txn, if tracked */
6559 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6560 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6561 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6563 *prev = mc->mc_next;
6570 mdb_cursor_txn(MDB_cursor *mc)
6572 if (!mc) return NULL;
6577 mdb_cursor_dbi(MDB_cursor *mc)
6583 /** Replace the key for a node with a new key.
6584 * @param[in] mc Cursor pointing to the node to operate on.
6585 * @param[in] key The new key to use.
6586 * @return 0 on success, non-zero on failure.
6589 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6596 indx_t ptr, i, numkeys, indx;
6599 indx = mc->mc_ki[mc->mc_top];
6600 mp = mc->mc_pg[mc->mc_top];
6601 node = NODEPTR(mp, indx);
6602 ptr = mp->mp_ptrs[indx];
6606 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6607 k2.mv_data = NODEKEY(node);
6608 k2.mv_size = node->mn_ksize;
6609 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6611 mdb_dkey(&k2, kbuf2),
6617 delta0 = delta = key->mv_size - node->mn_ksize;
6619 /* Must be 2-byte aligned. If new key is
6620 * shorter by 1, the shift will be skipped.
6622 delta += (delta & 1);
6624 if (delta > 0 && SIZELEFT(mp) < delta) {
6626 /* not enough space left, do a delete and split */
6627 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6628 pgno = NODEPGNO(node);
6629 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6630 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6633 numkeys = NUMKEYS(mp);
6634 for (i = 0; i < numkeys; i++) {
6635 if (mp->mp_ptrs[i] <= ptr)
6636 mp->mp_ptrs[i] -= delta;
6639 base = (char *)mp + mp->mp_upper;
6640 len = ptr - mp->mp_upper + NODESIZE;
6641 memmove(base - delta, base, len);
6642 mp->mp_upper -= delta;
6644 node = NODEPTR(mp, indx);
6647 /* But even if no shift was needed, update ksize */
6649 node->mn_ksize = key->mv_size;
6652 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6658 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6660 /** Move a node from csrc to cdst.
6663 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6670 unsigned short flags;
6674 /* Mark src and dst as dirty. */
6675 if ((rc = mdb_page_touch(csrc)) ||
6676 (rc = mdb_page_touch(cdst)))
6679 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6680 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6681 key.mv_size = csrc->mc_db->md_pad;
6682 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6684 data.mv_data = NULL;
6688 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6689 assert(!((long)srcnode&1));
6690 srcpg = NODEPGNO(srcnode);
6691 flags = srcnode->mn_flags;
6692 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6693 unsigned int snum = csrc->mc_snum;
6695 /* must find the lowest key below src */
6696 mdb_page_search_lowest(csrc);
6697 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6698 key.mv_size = csrc->mc_db->md_pad;
6699 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6701 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6702 key.mv_size = NODEKSZ(s2);
6703 key.mv_data = NODEKEY(s2);
6705 csrc->mc_snum = snum--;
6706 csrc->mc_top = snum;
6708 key.mv_size = NODEKSZ(srcnode);
6709 key.mv_data = NODEKEY(srcnode);
6711 data.mv_size = NODEDSZ(srcnode);
6712 data.mv_data = NODEDATA(srcnode);
6714 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6715 unsigned int snum = cdst->mc_snum;
6718 /* must find the lowest key below dst */
6719 mdb_page_search_lowest(cdst);
6720 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6721 bkey.mv_size = cdst->mc_db->md_pad;
6722 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6724 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6725 bkey.mv_size = NODEKSZ(s2);
6726 bkey.mv_data = NODEKEY(s2);
6728 cdst->mc_snum = snum--;
6729 cdst->mc_top = snum;
6730 mdb_cursor_copy(cdst, &mn);
6732 rc = mdb_update_key(&mn, &bkey);
6737 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6738 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6739 csrc->mc_ki[csrc->mc_top],
6741 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6742 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6744 /* Add the node to the destination page.
6746 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6747 if (rc != MDB_SUCCESS)
6750 /* Delete the node from the source page.
6752 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6755 /* Adjust other cursors pointing to mp */
6756 MDB_cursor *m2, *m3;
6757 MDB_dbi dbi = csrc->mc_dbi;
6758 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6760 if (csrc->mc_flags & C_SUB)
6763 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6764 if (csrc->mc_flags & C_SUB)
6765 m3 = &m2->mc_xcursor->mx_cursor;
6768 if (m3 == csrc) continue;
6769 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6770 csrc->mc_ki[csrc->mc_top]) {
6771 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6772 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6777 /* Update the parent separators.
6779 if (csrc->mc_ki[csrc->mc_top] == 0) {
6780 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6781 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6782 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6784 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6785 key.mv_size = NODEKSZ(srcnode);
6786 key.mv_data = NODEKEY(srcnode);
6788 DPRINTF(("update separator for source page %"Z"u to [%s]",
6789 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6790 mdb_cursor_copy(csrc, &mn);
6793 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6796 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6798 indx_t ix = csrc->mc_ki[csrc->mc_top];
6799 nullkey.mv_size = 0;
6800 csrc->mc_ki[csrc->mc_top] = 0;
6801 rc = mdb_update_key(csrc, &nullkey);
6802 csrc->mc_ki[csrc->mc_top] = ix;
6803 assert(rc == MDB_SUCCESS);
6807 if (cdst->mc_ki[cdst->mc_top] == 0) {
6808 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6809 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6810 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6812 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6813 key.mv_size = NODEKSZ(srcnode);
6814 key.mv_data = NODEKEY(srcnode);
6816 DPRINTF(("update separator for destination page %"Z"u to [%s]",
6817 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
6818 mdb_cursor_copy(cdst, &mn);
6821 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6824 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6826 indx_t ix = cdst->mc_ki[cdst->mc_top];
6827 nullkey.mv_size = 0;
6828 cdst->mc_ki[cdst->mc_top] = 0;
6829 rc = mdb_update_key(cdst, &nullkey);
6830 cdst->mc_ki[cdst->mc_top] = ix;
6831 assert(rc == MDB_SUCCESS);
6838 /** Merge one page into another.
6839 * The nodes from the page pointed to by \b csrc will
6840 * be copied to the page pointed to by \b cdst and then
6841 * the \b csrc page will be freed.
6842 * @param[in] csrc Cursor pointing to the source page.
6843 * @param[in] cdst Cursor pointing to the destination page.
6846 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6854 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6855 cdst->mc_pg[cdst->mc_top]->mp_pgno));
6857 assert(csrc->mc_snum > 1); /* can't merge root page */
6858 assert(cdst->mc_snum > 1);
6860 /* Mark dst as dirty. */
6861 if ((rc = mdb_page_touch(cdst)))
6864 /* Move all nodes from src to dst.
6866 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6867 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6868 key.mv_size = csrc->mc_db->md_pad;
6869 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6870 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6871 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6872 if (rc != MDB_SUCCESS)
6874 key.mv_data = (char *)key.mv_data + key.mv_size;
6877 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6878 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6879 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6880 unsigned int snum = csrc->mc_snum;
6882 /* must find the lowest key below src */
6883 mdb_page_search_lowest(csrc);
6884 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6885 key.mv_size = csrc->mc_db->md_pad;
6886 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6888 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6889 key.mv_size = NODEKSZ(s2);
6890 key.mv_data = NODEKEY(s2);
6892 csrc->mc_snum = snum--;
6893 csrc->mc_top = snum;
6895 key.mv_size = srcnode->mn_ksize;
6896 key.mv_data = NODEKEY(srcnode);
6899 data.mv_size = NODEDSZ(srcnode);
6900 data.mv_data = NODEDATA(srcnode);
6901 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6902 if (rc != MDB_SUCCESS)
6907 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
6908 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
6909 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
6911 /* Unlink the src page from parent and add to free list.
6913 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6914 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6917 rc = mdb_update_key(csrc, &key);
6923 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6924 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6927 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6928 csrc->mc_db->md_leaf_pages--;
6930 csrc->mc_db->md_branch_pages--;
6932 /* Adjust other cursors pointing to mp */
6933 MDB_cursor *m2, *m3;
6934 MDB_dbi dbi = csrc->mc_dbi;
6935 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6937 if (csrc->mc_flags & C_SUB)
6940 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6941 if (csrc->mc_flags & C_SUB)
6942 m3 = &m2->mc_xcursor->mx_cursor;
6945 if (m3 == csrc) continue;
6946 if (m3->mc_snum < csrc->mc_snum) continue;
6947 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6948 m3->mc_pg[csrc->mc_top] = mp;
6949 m3->mc_ki[csrc->mc_top] += nkeys;
6953 mdb_cursor_pop(csrc);
6955 return mdb_rebalance(csrc);
6958 /** Copy the contents of a cursor.
6959 * @param[in] csrc The cursor to copy from.
6960 * @param[out] cdst The cursor to copy to.
6963 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6967 cdst->mc_txn = csrc->mc_txn;
6968 cdst->mc_dbi = csrc->mc_dbi;
6969 cdst->mc_db = csrc->mc_db;
6970 cdst->mc_dbx = csrc->mc_dbx;
6971 cdst->mc_snum = csrc->mc_snum;
6972 cdst->mc_top = csrc->mc_top;
6973 cdst->mc_flags = csrc->mc_flags;
6975 for (i=0; i<csrc->mc_snum; i++) {
6976 cdst->mc_pg[i] = csrc->mc_pg[i];
6977 cdst->mc_ki[i] = csrc->mc_ki[i];
6981 /** Rebalance the tree after a delete operation.
6982 * @param[in] mc Cursor pointing to the page where rebalancing
6984 * @return 0 on success, non-zero on failure.
6987 mdb_rebalance(MDB_cursor *mc)
6991 unsigned int ptop, minkeys;
6994 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6998 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6999 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7000 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7001 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7002 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7006 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7007 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7010 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7011 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7017 if (mc->mc_snum < 2) {
7018 MDB_page *mp = mc->mc_pg[0];
7020 DPUTS("Can't rebalance a subpage, ignoring");
7023 if (NUMKEYS(mp) == 0) {
7024 DPUTS("tree is completely empty");
7025 mc->mc_db->md_root = P_INVALID;
7026 mc->mc_db->md_depth = 0;
7027 mc->mc_db->md_leaf_pages = 0;
7028 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7031 /* Adjust cursors pointing to mp */
7035 MDB_cursor *m2, *m3;
7036 MDB_dbi dbi = mc->mc_dbi;
7038 if (mc->mc_flags & C_SUB)
7041 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7042 if (mc->mc_flags & C_SUB)
7043 m3 = &m2->mc_xcursor->mx_cursor;
7046 if (m3->mc_snum < mc->mc_snum) continue;
7047 if (m3->mc_pg[0] == mp) {
7053 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7054 DPUTS("collapsing root page!");
7055 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7058 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7059 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7062 mc->mc_db->md_depth--;
7063 mc->mc_db->md_branch_pages--;
7064 mc->mc_ki[0] = mc->mc_ki[1];
7066 /* Adjust other cursors pointing to mp */
7067 MDB_cursor *m2, *m3;
7068 MDB_dbi dbi = mc->mc_dbi;
7070 if (mc->mc_flags & C_SUB)
7073 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7074 if (mc->mc_flags & C_SUB)
7075 m3 = &m2->mc_xcursor->mx_cursor;
7078 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7079 if (m3->mc_pg[0] == mp) {
7080 m3->mc_pg[0] = mc->mc_pg[0];
7083 m3->mc_ki[0] = m3->mc_ki[1];
7088 DPUTS("root page doesn't need rebalancing");
7092 /* The parent (branch page) must have at least 2 pointers,
7093 * otherwise the tree is invalid.
7095 ptop = mc->mc_top-1;
7096 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7098 /* Leaf page fill factor is below the threshold.
7099 * Try to move keys from left or right neighbor, or
7100 * merge with a neighbor page.
7105 mdb_cursor_copy(mc, &mn);
7106 mn.mc_xcursor = NULL;
7108 if (mc->mc_ki[ptop] == 0) {
7109 /* We're the leftmost leaf in our parent.
7111 DPUTS("reading right neighbor");
7113 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7114 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7117 mn.mc_ki[mn.mc_top] = 0;
7118 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7120 /* There is at least one neighbor to the left.
7122 DPUTS("reading left neighbor");
7124 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7125 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7128 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7129 mc->mc_ki[mc->mc_top] = 0;
7132 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7133 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7134 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7136 /* If the neighbor page is above threshold and has enough keys,
7137 * move one key from it. Otherwise we should try to merge them.
7138 * (A branch page must never have less than 2 keys.)
7140 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7141 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7142 return mdb_node_move(&mn, mc);
7144 if (mc->mc_ki[ptop] == 0)
7145 rc = mdb_page_merge(&mn, mc);
7147 rc = mdb_page_merge(mc, &mn);
7148 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7153 /** Complete a delete operation started by #mdb_cursor_del(). */
7155 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7161 mp = mc->mc_pg[mc->mc_top];
7162 ki = mc->mc_ki[mc->mc_top];
7164 /* add overflow pages to free list */
7165 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7169 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7170 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7171 (rc = mdb_ovpage_free(mc, omp)))
7174 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7175 mc->mc_db->md_entries--;
7176 rc = mdb_rebalance(mc);
7177 if (rc != MDB_SUCCESS)
7178 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7179 /* if mc points past last node in page, invalidate */
7180 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
7181 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7184 /* Adjust other cursors pointing to mp */
7187 MDB_dbi dbi = mc->mc_dbi;
7189 mp = mc->mc_pg[mc->mc_top];
7190 nkeys = NUMKEYS(mp);
7191 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7194 if (!(m2->mc_flags & C_INITIALIZED))
7196 if (m2->mc_pg[mc->mc_top] == mp) {
7197 if (m2->mc_ki[mc->mc_top] > ki)
7198 m2->mc_ki[mc->mc_top]--;
7199 if (m2->mc_ki[mc->mc_top] >= nkeys)
7200 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
7209 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7210 MDB_val *key, MDB_val *data)
7215 MDB_val rdata, *xdata;
7219 assert(key != NULL);
7221 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7223 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7226 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7227 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7229 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7230 return MDB_BAD_VALSIZE;
7233 mdb_cursor_init(&mc, txn, dbi, &mx);
7236 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7237 /* must ignore any data */
7248 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7250 /* let mdb_page_split know about this cursor if needed:
7251 * delete will trigger a rebalance; if it needs to move
7252 * a node from one page to another, it will have to
7253 * update the parent's separator key(s). If the new sepkey
7254 * is larger than the current one, the parent page may
7255 * run out of space, triggering a split. We need this
7256 * cursor to be consistent until the end of the rebalance.
7258 mc.mc_flags |= C_UNTRACK;
7259 mc.mc_next = txn->mt_cursors[dbi];
7260 txn->mt_cursors[dbi] = &mc;
7261 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7262 txn->mt_cursors[dbi] = mc.mc_next;
7267 /** Split a page and insert a new node.
7268 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7269 * The cursor will be updated to point to the actual page and index where
7270 * the node got inserted after the split.
7271 * @param[in] newkey The key for the newly inserted node.
7272 * @param[in] newdata The data for the newly inserted node.
7273 * @param[in] newpgno The page number, if the new node is a branch node.
7274 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7275 * @return 0 on success, non-zero on failure.
7278 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7279 unsigned int nflags)
7282 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
7285 unsigned int i, j, split_indx, nkeys, pmax;
7287 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7289 MDB_page *mp, *rp, *pp;
7294 mp = mc->mc_pg[mc->mc_top];
7295 newindx = mc->mc_ki[mc->mc_top];
7297 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i",
7298 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7299 DKEY(newkey), mc->mc_ki[mc->mc_top]));
7301 /* Create a right sibling. */
7302 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7304 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7306 if (mc->mc_snum < 2) {
7307 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7309 /* shift current top to make room for new parent */
7310 mc->mc_pg[1] = mc->mc_pg[0];
7311 mc->mc_ki[1] = mc->mc_ki[0];
7314 mc->mc_db->md_root = pp->mp_pgno;
7315 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7316 mc->mc_db->md_depth++;
7319 /* Add left (implicit) pointer. */
7320 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7321 /* undo the pre-push */
7322 mc->mc_pg[0] = mc->mc_pg[1];
7323 mc->mc_ki[0] = mc->mc_ki[1];
7324 mc->mc_db->md_root = mp->mp_pgno;
7325 mc->mc_db->md_depth--;
7332 ptop = mc->mc_top-1;
7333 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7336 mc->mc_flags |= C_SPLITTING;
7337 mdb_cursor_copy(mc, &mn);
7338 mn.mc_pg[mn.mc_top] = rp;
7339 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7341 if (nflags & MDB_APPEND) {
7342 mn.mc_ki[mn.mc_top] = 0;
7344 split_indx = newindx;
7349 nkeys = NUMKEYS(mp);
7350 split_indx = nkeys / 2;
7351 if (newindx < split_indx)
7357 unsigned int lsize, rsize, ksize;
7358 /* Move half of the keys to the right sibling */
7360 x = mc->mc_ki[mc->mc_top] - split_indx;
7361 ksize = mc->mc_db->md_pad;
7362 split = LEAF2KEY(mp, split_indx, ksize);
7363 rsize = (nkeys - split_indx) * ksize;
7364 lsize = (nkeys - split_indx) * sizeof(indx_t);
7365 mp->mp_lower -= lsize;
7366 rp->mp_lower += lsize;
7367 mp->mp_upper += rsize - lsize;
7368 rp->mp_upper -= rsize - lsize;
7369 sepkey.mv_size = ksize;
7370 if (newindx == split_indx) {
7371 sepkey.mv_data = newkey->mv_data;
7373 sepkey.mv_data = split;
7376 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7377 memcpy(rp->mp_ptrs, split, rsize);
7378 sepkey.mv_data = rp->mp_ptrs;
7379 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7380 memcpy(ins, newkey->mv_data, ksize);
7381 mp->mp_lower += sizeof(indx_t);
7382 mp->mp_upper -= ksize - sizeof(indx_t);
7385 memcpy(rp->mp_ptrs, split, x * ksize);
7386 ins = LEAF2KEY(rp, x, ksize);
7387 memcpy(ins, newkey->mv_data, ksize);
7388 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7389 rp->mp_lower += sizeof(indx_t);
7390 rp->mp_upper -= ksize - sizeof(indx_t);
7391 mc->mc_ki[mc->mc_top] = x;
7392 mc->mc_pg[mc->mc_top] = rp;
7397 /* For leaf pages, check the split point based on what
7398 * fits where, since otherwise mdb_node_add can fail.
7400 * This check is only needed when the data items are
7401 * relatively large, such that being off by one will
7402 * make the difference between success or failure.
7404 * It's also relevant if a page happens to be laid out
7405 * such that one half of its nodes are all "small" and
7406 * the other half of its nodes are "large." If the new
7407 * item is also "large" and falls on the half with
7408 * "large" nodes, it also may not fit.
7411 unsigned int psize, nsize;
7412 /* Maximum free space in an empty page */
7413 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
7414 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
7415 if ((nkeys < 20) || (nsize > pmax/16)) {
7416 if (newindx <= split_indx) {
7419 for (i=0; i<split_indx; i++) {
7420 node = NODEPTR(mp, i);
7421 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7422 if (F_ISSET(node->mn_flags, F_BIGDATA))
7423 psize += sizeof(pgno_t);
7425 psize += NODEDSZ(node);
7429 split_indx = newindx;
7440 for (i=nkeys-1; i>=split_indx; i--) {
7441 node = NODEPTR(mp, i);
7442 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7443 if (F_ISSET(node->mn_flags, F_BIGDATA))
7444 psize += sizeof(pgno_t);
7446 psize += NODEDSZ(node);
7450 split_indx = newindx;
7461 /* First find the separating key between the split pages.
7462 * The case where newindx == split_indx is ambiguous; the
7463 * new item could go to the new page or stay on the original
7464 * page. If newpos == 1 it goes to the new page.
7466 if (newindx == split_indx && newpos) {
7467 sepkey.mv_size = newkey->mv_size;
7468 sepkey.mv_data = newkey->mv_data;
7470 node = NODEPTR(mp, split_indx);
7471 sepkey.mv_size = node->mn_ksize;
7472 sepkey.mv_data = NODEKEY(node);
7476 DPRINTF(("separator is [%s]", DKEY(&sepkey)));
7478 /* Copy separator key to the parent.
7480 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
7484 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7487 if (mn.mc_snum == mc->mc_snum) {
7488 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7489 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7490 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7491 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7496 /* Right page might now have changed parent.
7497 * Check if left page also changed parent.
7499 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7500 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7501 for (i=0; i<ptop; i++) {
7502 mc->mc_pg[i] = mn.mc_pg[i];
7503 mc->mc_ki[i] = mn.mc_ki[i];
7505 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7506 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7510 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7513 mc->mc_flags ^= C_SPLITTING;
7514 if (rc != MDB_SUCCESS) {
7517 if (nflags & MDB_APPEND) {
7518 mc->mc_pg[mc->mc_top] = rp;
7519 mc->mc_ki[mc->mc_top] = 0;
7520 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7523 for (i=0; i<mc->mc_top; i++)
7524 mc->mc_ki[i] = mn.mc_ki[i];
7531 /* Move half of the keys to the right sibling. */
7533 /* grab a page to hold a temporary copy */
7534 copy = mdb_page_malloc(mc->mc_txn, 1);
7538 copy->mp_pgno = mp->mp_pgno;
7539 copy->mp_flags = mp->mp_flags;
7540 copy->mp_lower = PAGEHDRSZ;
7541 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
7542 mc->mc_pg[mc->mc_top] = copy;
7543 for (i = j = 0; i <= nkeys; j++) {
7544 if (i == split_indx) {
7545 /* Insert in right sibling. */
7546 /* Reset insert index for right sibling. */
7547 if (i != newindx || (newpos ^ ins_new)) {
7549 mc->mc_pg[mc->mc_top] = rp;
7553 if (i == newindx && !ins_new) {
7554 /* Insert the original entry that caused the split. */
7555 rkey.mv_data = newkey->mv_data;
7556 rkey.mv_size = newkey->mv_size;
7565 /* Update index for the new key. */
7566 mc->mc_ki[mc->mc_top] = j;
7567 } else if (i == nkeys) {
7570 node = NODEPTR(mp, i);
7571 rkey.mv_data = NODEKEY(node);
7572 rkey.mv_size = node->mn_ksize;
7574 xdata.mv_data = NODEDATA(node);
7575 xdata.mv_size = NODEDSZ(node);
7578 pgno = NODEPGNO(node);
7579 flags = node->mn_flags;
7584 if (!IS_LEAF(mp) && j == 0) {
7585 /* First branch index doesn't need key data. */
7589 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7593 nkeys = NUMKEYS(copy);
7594 for (i=0; i<nkeys; i++)
7595 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7596 mp->mp_lower = copy->mp_lower;
7597 mp->mp_upper = copy->mp_upper;
7598 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7599 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7601 /* reset back to original page */
7602 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7603 mc->mc_pg[mc->mc_top] = mp;
7604 if (nflags & MDB_RESERVE) {
7605 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7606 if (!(node->mn_flags & F_BIGDATA))
7607 newdata->mv_data = NODEDATA(node);
7611 /* Make sure mc_ki is still valid.
7613 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7614 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7615 for (i=0; i<ptop; i++) {
7616 mc->mc_pg[i] = mn.mc_pg[i];
7617 mc->mc_ki[i] = mn.mc_ki[i];
7619 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7620 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7624 /* return tmp page to freelist */
7625 mdb_page_free(mc->mc_txn->mt_env, copy);
7628 /* Adjust other cursors pointing to mp */
7629 MDB_cursor *m2, *m3;
7630 MDB_dbi dbi = mc->mc_dbi;
7631 int fixup = NUMKEYS(mp);
7633 if (mc->mc_flags & C_SUB)
7636 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7637 if (mc->mc_flags & C_SUB)
7638 m3 = &m2->mc_xcursor->mx_cursor;
7643 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7645 if (m3->mc_flags & C_SPLITTING)
7650 for (k=m3->mc_top; k>=0; k--) {
7651 m3->mc_ki[k+1] = m3->mc_ki[k];
7652 m3->mc_pg[k+1] = m3->mc_pg[k];
7654 if (m3->mc_ki[0] >= split_indx) {
7659 m3->mc_pg[0] = mc->mc_pg[0];
7663 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7664 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7665 m3->mc_ki[mc->mc_top]++;
7666 if (m3->mc_ki[mc->mc_top] >= fixup) {
7667 m3->mc_pg[mc->mc_top] = rp;
7668 m3->mc_ki[mc->mc_top] -= fixup;
7669 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7671 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7672 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7681 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7682 MDB_val *key, MDB_val *data, unsigned int flags)
7687 assert(key != NULL);
7688 assert(data != NULL);
7690 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7693 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7694 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7696 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7697 return MDB_BAD_VALSIZE;
7700 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7703 mdb_cursor_init(&mc, txn, dbi, &mx);
7704 return mdb_cursor_put(&mc, key, data, flags);
7708 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7710 if ((flag & CHANGEABLE) != flag)
7713 env->me_flags |= flag;
7715 env->me_flags &= ~flag;
7720 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7725 *arg = env->me_flags;
7730 mdb_env_get_path(MDB_env *env, const char **arg)
7735 *arg = env->me_path;
7739 /** Common code for #mdb_stat() and #mdb_env_stat().
7740 * @param[in] env the environment to operate in.
7741 * @param[in] db the #MDB_db record containing the stats to return.
7742 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7743 * @return 0, this function always succeeds.
7746 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7748 arg->ms_psize = env->me_psize;
7749 arg->ms_depth = db->md_depth;
7750 arg->ms_branch_pages = db->md_branch_pages;
7751 arg->ms_leaf_pages = db->md_leaf_pages;
7752 arg->ms_overflow_pages = db->md_overflow_pages;
7753 arg->ms_entries = db->md_entries;
7758 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7762 if (env == NULL || arg == NULL)
7765 toggle = mdb_env_pick_meta(env);
7767 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7771 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7775 if (env == NULL || arg == NULL)
7778 toggle = mdb_env_pick_meta(env);
7779 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7780 arg->me_mapsize = env->me_mapsize;
7781 arg->me_maxreaders = env->me_maxreaders;
7783 /* me_numreaders may be zero if this process never used any readers. Use
7784 * the shared numreader count if it exists.
7786 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7788 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7789 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7793 /** Set the default comparison functions for a database.
7794 * Called immediately after a database is opened to set the defaults.
7795 * The user can then override them with #mdb_set_compare() or
7796 * #mdb_set_dupsort().
7797 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7798 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7801 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7803 uint16_t f = txn->mt_dbs[dbi].md_flags;
7805 txn->mt_dbxs[dbi].md_cmp =
7806 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7807 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7809 txn->mt_dbxs[dbi].md_dcmp =
7810 !(f & MDB_DUPSORT) ? 0 :
7811 ((f & MDB_INTEGERDUP)
7812 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7813 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7816 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7821 int rc, dbflag, exact;
7822 unsigned int unused = 0;
7825 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7826 mdb_default_cmp(txn, FREE_DBI);
7829 if ((flags & VALID_FLAGS) != flags)
7831 if (txn->mt_flags & MDB_TXN_ERROR)
7837 if (flags & PERSISTENT_FLAGS) {
7838 uint16_t f2 = flags & PERSISTENT_FLAGS;
7839 /* make sure flag changes get committed */
7840 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7841 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7842 txn->mt_flags |= MDB_TXN_DIRTY;
7845 mdb_default_cmp(txn, MAIN_DBI);
7849 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7850 mdb_default_cmp(txn, MAIN_DBI);
7853 /* Is the DB already open? */
7855 for (i=2; i<txn->mt_numdbs; i++) {
7856 if (!txn->mt_dbxs[i].md_name.mv_size) {
7857 /* Remember this free slot */
7858 if (!unused) unused = i;
7861 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7862 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7868 /* If no free slot and max hit, fail */
7869 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7870 return MDB_DBS_FULL;
7872 /* Cannot mix named databases with some mainDB flags */
7873 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7874 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7876 /* Find the DB info */
7877 dbflag = DB_NEW|DB_VALID;
7880 key.mv_data = (void *)name;
7881 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7882 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7883 if (rc == MDB_SUCCESS) {
7884 /* make sure this is actually a DB */
7885 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7886 if (!(node->mn_flags & F_SUBDATA))
7887 return MDB_INCOMPATIBLE;
7888 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7889 /* Create if requested */
7891 data.mv_size = sizeof(MDB_db);
7892 data.mv_data = &dummy;
7893 memset(&dummy, 0, sizeof(dummy));
7894 dummy.md_root = P_INVALID;
7895 dummy.md_flags = flags & PERSISTENT_FLAGS;
7896 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7900 /* OK, got info, add to table */
7901 if (rc == MDB_SUCCESS) {
7902 unsigned int slot = unused ? unused : txn->mt_numdbs;
7903 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7904 txn->mt_dbxs[slot].md_name.mv_size = len;
7905 txn->mt_dbxs[slot].md_rel = NULL;
7906 txn->mt_dbflags[slot] = dbflag;
7907 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7909 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7910 mdb_default_cmp(txn, slot);
7919 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7921 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7924 if (txn->mt_dbflags[dbi] & DB_STALE) {
7927 /* Stale, must read the DB's root. cursor_init does it for us. */
7928 mdb_cursor_init(&mc, txn, dbi, &mx);
7930 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7933 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7936 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7938 ptr = env->me_dbxs[dbi].md_name.mv_data;
7939 env->me_dbxs[dbi].md_name.mv_data = NULL;
7940 env->me_dbxs[dbi].md_name.mv_size = 0;
7941 env->me_dbflags[dbi] = 0;
7945 int mdb_dbi_flags(MDB_env *env, MDB_dbi dbi, unsigned int *flags)
7947 /* We could return the flags for the FREE_DBI too but what's the point? */
7948 if (dbi < MAIN_DBI || dbi >= env->me_numdbs)
7950 *flags = env->me_dbflags[dbi];
7954 /** Add all the DB's pages to the free list.
7955 * @param[in] mc Cursor on the DB to free.
7956 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7957 * @return 0 on success, non-zero on failure.
7960 mdb_drop0(MDB_cursor *mc, int subs)
7964 rc = mdb_page_search(mc, NULL, 0);
7965 if (rc == MDB_SUCCESS) {
7966 MDB_txn *txn = mc->mc_txn;
7971 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7972 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7975 mdb_cursor_copy(mc, &mx);
7976 while (mc->mc_snum > 0) {
7977 MDB_page *mp = mc->mc_pg[mc->mc_top];
7978 unsigned n = NUMKEYS(mp);
7980 for (i=0; i<n; i++) {
7981 ni = NODEPTR(mp, i);
7982 if (ni->mn_flags & F_BIGDATA) {
7985 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7986 rc = mdb_page_get(txn, pg, &omp, NULL);
7989 assert(IS_OVERFLOW(omp));
7990 rc = mdb_midl_append_range(&txn->mt_free_pgs,
7994 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7995 mdb_xcursor_init1(mc, ni);
7996 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8002 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8004 for (i=0; i<n; i++) {
8006 ni = NODEPTR(mp, i);
8009 mdb_midl_xappend(txn->mt_free_pgs, pg);
8014 mc->mc_ki[mc->mc_top] = i;
8015 rc = mdb_cursor_sibling(mc, 1);
8017 /* no more siblings, go back to beginning
8018 * of previous level.
8022 for (i=1; i<mc->mc_snum; i++) {
8024 mc->mc_pg[i] = mx.mc_pg[i];
8029 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8030 } else if (rc == MDB_NOTFOUND) {
8036 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8038 MDB_cursor *mc, *m2;
8041 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8044 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8047 rc = mdb_cursor_open(txn, dbi, &mc);
8051 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8052 /* Invalidate the dropped DB's cursors */
8053 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8054 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8058 /* Can't delete the main DB */
8059 if (del && dbi > MAIN_DBI) {
8060 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8062 txn->mt_dbflags[dbi] = DB_STALE;
8063 mdb_dbi_close(txn->mt_env, dbi);
8066 /* reset the DB record, mark it dirty */
8067 txn->mt_dbflags[dbi] |= DB_DIRTY;
8068 txn->mt_dbs[dbi].md_depth = 0;
8069 txn->mt_dbs[dbi].md_branch_pages = 0;
8070 txn->mt_dbs[dbi].md_leaf_pages = 0;
8071 txn->mt_dbs[dbi].md_overflow_pages = 0;
8072 txn->mt_dbs[dbi].md_entries = 0;
8073 txn->mt_dbs[dbi].md_root = P_INVALID;
8075 txn->mt_flags |= MDB_TXN_DIRTY;
8078 mdb_cursor_close(mc);
8082 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8084 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8087 txn->mt_dbxs[dbi].md_cmp = cmp;
8091 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8093 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8096 txn->mt_dbxs[dbi].md_dcmp = cmp;
8100 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8102 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8105 txn->mt_dbxs[dbi].md_rel = rel;
8109 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8111 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8114 txn->mt_dbxs[dbi].md_relctx = ctx;
8118 int mdb_env_get_maxkeysize(MDB_env *env)
8120 return MDB_MAXKEYSIZE;
8123 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8125 unsigned int i, rdrs;
8132 if (!env->me_txns) {
8133 return func("(no reader locks)\n", ctx);
8135 rdrs = env->me_txns->mti_numreaders;
8136 mr = env->me_txns->mti_readers;
8137 for (i=0; i<rdrs; i++) {
8142 if (mr[i].mr_txnid == (txnid_t)-1) {
8143 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8145 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8149 func(" pid thread txnid\n", ctx);
8151 rc = func(buf, ctx);
8157 func("(no active readers)\n", ctx);
8162 /* insert pid into list if not already present.
8163 * return -1 if already present.
8165 static int mdb_pid_insert(pid_t *ids, pid_t pid)
8167 /* binary search of pid in list */
8169 unsigned cursor = 1;
8171 unsigned n = ids[0];
8174 unsigned pivot = n >> 1;
8175 cursor = base + pivot + 1;
8176 val = pid - ids[cursor];
8181 } else if ( val > 0 ) {
8186 /* found, so it's a duplicate */
8195 for (n = ids[0]; n > cursor; n--)
8201 int mdb_reader_check(MDB_env *env, int *dead)
8203 unsigned int i, j, rdrs;
8214 rdrs = env->me_txns->mti_numreaders;
8215 pids = malloc((rdrs+1) * sizeof(pid_t));
8219 mr = env->me_txns->mti_readers;
8221 for (i=0; i<rdrs; i++) {
8222 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8224 if (mdb_pid_insert(pids, pid) == 0) {
8225 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8227 /* Recheck, a new process may have reused pid */
8228 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8229 for (j=i; j<rdrs; j++)
8230 if (mr[j].mr_pid == pid) {
8235 UNLOCK_MUTEX_R(env);