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.
36 #include <sys/types.h>
38 #include <sys/param.h>
44 #ifdef HAVE_SYS_FILE_H
61 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
62 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
65 #if defined(__APPLE__) || defined (BSD)
66 # define MDB_USE_POSIX_SEM 1
67 # define MDB_FDATASYNC fsync
68 #elif defined(ANDROID)
69 # define MDB_FDATASYNC fsync
74 #ifdef MDB_USE_POSIX_SEM
75 #include <semaphore.h>
80 #include <valgrind/memcheck.h>
81 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
82 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
83 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
84 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
85 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
87 #define VGMEMP_CREATE(h,r,z)
88 #define VGMEMP_ALLOC(h,a,s)
89 #define VGMEMP_FREE(h,a)
90 #define VGMEMP_DESTROY(h)
91 #define VGMEMP_DEFINED(a,s)
95 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
96 /* Solaris just defines one or the other */
97 # define LITTLE_ENDIAN 1234
98 # define BIG_ENDIAN 4321
99 # ifdef _LITTLE_ENDIAN
100 # define BYTE_ORDER LITTLE_ENDIAN
102 # define BYTE_ORDER BIG_ENDIAN
105 # define BYTE_ORDER __BYTE_ORDER
109 #ifndef LITTLE_ENDIAN
110 #define LITTLE_ENDIAN __LITTLE_ENDIAN
113 #define BIG_ENDIAN __BIG_ENDIAN
116 #if defined(__i386) || defined(__x86_64)
117 #define MISALIGNED_OK 1
123 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
124 # error "Unknown or unsupported endianness (BYTE_ORDER)"
125 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
126 # error "Two's complement, reasonably sized integer types, please"
129 /** @defgroup internal MDB Internals
132 /** @defgroup compat Windows Compatibility Macros
133 * A bunch of macros to minimize the amount of platform-specific ifdefs
134 * needed throughout the rest of the code. When the features this library
135 * needs are similar enough to POSIX to be hidden in a one-or-two line
136 * replacement, this macro approach is used.
140 #define pthread_t DWORD
141 #define pthread_mutex_t HANDLE
142 #define pthread_key_t DWORD
143 #define pthread_self() GetCurrentThreadId()
144 #define pthread_key_create(x,y) \
145 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
146 #define pthread_key_delete(x) TlsFree(x)
147 #define pthread_getspecific(x) TlsGetValue(x)
148 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
149 #define pthread_mutex_unlock(x) ReleaseMutex(x)
150 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
151 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
152 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
153 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
154 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
155 #define getpid() GetCurrentProcessId()
156 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
157 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
158 #define ErrCode() GetLastError()
159 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
160 #define close(fd) CloseHandle(fd)
161 #define munmap(ptr,len) UnmapViewOfFile(ptr)
164 #ifdef MDB_USE_POSIX_SEM
166 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
167 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
168 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
169 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
172 mdb_sem_wait(sem_t *sem)
175 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
180 /** Lock the reader mutex.
182 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
183 /** Unlock the reader mutex.
185 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
187 /** Lock the writer mutex.
188 * Only a single write transaction is allowed at a time. Other writers
189 * will block waiting for this mutex.
191 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
192 /** Unlock the writer mutex.
194 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
195 #endif /* MDB_USE_POSIX_SEM */
197 /** Get the error code for the last failed system function.
199 #define ErrCode() errno
201 /** An abstraction for a file handle.
202 * On POSIX systems file handles are small integers. On Windows
203 * they're opaque pointers.
207 /** A value for an invalid file handle.
208 * Mainly used to initialize file variables and signify that they are
211 #define INVALID_HANDLE_VALUE (-1)
213 /** Get the size of a memory page for the system.
214 * This is the basic size that the platform's memory manager uses, and is
215 * fundamental to the use of memory-mapped files.
217 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
220 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
223 #define MNAME_LEN (sizeof(pthread_mutex_t))
229 /** A flag for opening a file and requesting synchronous data writes.
230 * This is only used when writing a meta page. It's not strictly needed;
231 * we could just do a normal write and then immediately perform a flush.
232 * But if this flag is available it saves us an extra system call.
234 * @note If O_DSYNC is undefined but exists in /usr/include,
235 * preferably set some compiler flag to get the definition.
236 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
239 # define MDB_DSYNC O_DSYNC
243 /** Function for flushing the data of a file. Define this to fsync
244 * if fdatasync() is not supported.
246 #ifndef MDB_FDATASYNC
247 # define MDB_FDATASYNC fdatasync
251 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
262 /** A page number in the database.
263 * Note that 64 bit page numbers are overkill, since pages themselves
264 * already represent 12-13 bits of addressable memory, and the OS will
265 * always limit applications to a maximum of 63 bits of address space.
267 * @note In the #MDB_node structure, we only store 48 bits of this value,
268 * which thus limits us to only 60 bits of addressable data.
270 typedef MDB_ID pgno_t;
272 /** A transaction ID.
273 * See struct MDB_txn.mt_txnid for details.
275 typedef MDB_ID txnid_t;
277 /** @defgroup debug Debug Macros
281 /** Enable debug output.
282 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
283 * read from and written to the database (used for free space management).
288 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
289 # define DPRINTF (void) /* Vararg macros may be unsupported */
291 static int mdb_debug;
292 static txnid_t mdb_debug_start;
294 /** Print a debug message with printf formatting. */
295 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
296 ((void) ((mdb_debug) && \
297 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
299 # define DPRINTF(fmt, ...) ((void) 0)
300 # define MDB_DEBUG_SKIP
302 /** Print a debug string.
303 * The string is printed literally, with no format processing.
305 #define DPUTS(arg) DPRINTF("%s", arg)
308 /** A default memory page size.
309 * The actual size is platform-dependent, but we use this for
310 * boot-strapping. We probably should not be using this any more.
311 * The #GET_PAGESIZE() macro is used to get the actual size.
313 * Note that we don't currently support Huge pages. On Linux,
314 * regular data files cannot use Huge pages, and in general
315 * Huge pages aren't actually pageable. We rely on the OS
316 * demand-pager to read our data and page it out when memory
317 * pressure from other processes is high. So until OSs have
318 * actual paging support for Huge pages, they're not viable.
320 #define MDB_PAGESIZE 4096
322 /** The minimum number of keys required in a database page.
323 * Setting this to a larger value will place a smaller bound on the
324 * maximum size of a data item. Data items larger than this size will
325 * be pushed into overflow pages instead of being stored directly in
326 * the B-tree node. This value used to default to 4. With a page size
327 * of 4096 bytes that meant that any item larger than 1024 bytes would
328 * go into an overflow page. That also meant that on average 2-3KB of
329 * each overflow page was wasted space. The value cannot be lower than
330 * 2 because then there would no longer be a tree structure. With this
331 * value, items larger than 2KB will go into overflow pages, and on
332 * average only 1KB will be wasted.
334 #define MDB_MINKEYS 2
336 /** A stamp that identifies a file as an MDB file.
337 * There's nothing special about this value other than that it is easily
338 * recognizable, and it will reflect any byte order mismatches.
340 #define MDB_MAGIC 0xBEEFC0DE
342 /** The version number for a database's file format. */
343 #define MDB_VERSION 1
345 /** @brief The maximum size of a key in the database.
347 * We require that keys all fit onto a regular page. This limit
348 * could be raised a bit further if needed; to something just
349 * under #MDB_PAGESIZE / #MDB_MINKEYS.
351 * Note that data items in an #MDB_DUPSORT database are actually keys
352 * of a subDB, so they're also limited to this size.
354 #ifndef MDB_MAXKEYSIZE
355 #define MDB_MAXKEYSIZE 511
358 /** @brief The maximum size of a data item.
360 * We only store a 32 bit value for node sizes.
362 #define MAXDATASIZE 0xffffffffUL
367 * This is used for printing a hex dump of a key's contents.
369 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
370 /** Display a key in hex.
372 * Invoke a function to display a key in hex.
374 #define DKEY(x) mdb_dkey(x, kbuf)
376 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
380 /** An invalid page number.
381 * Mainly used to denote an empty tree.
383 #define P_INVALID (~(pgno_t)0)
385 /** Test if a flag \b f is set in a flag word \b w. */
386 #define F_ISSET(w, f) (((w) & (f)) == (f))
388 /** Used for offsets within a single page.
389 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
392 typedef uint16_t indx_t;
394 /** Default size of memory map.
395 * This is certainly too small for any actual applications. Apps should always set
396 * the size explicitly using #mdb_env_set_mapsize().
398 #define DEFAULT_MAPSIZE 1048576
400 /** @defgroup readers Reader Lock Table
401 * Readers don't acquire any locks for their data access. Instead, they
402 * simply record their transaction ID in the reader table. The reader
403 * mutex is needed just to find an empty slot in the reader table. The
404 * slot's address is saved in thread-specific data so that subsequent read
405 * transactions started by the same thread need no further locking to proceed.
407 * Since the database uses multi-version concurrency control, readers don't
408 * actually need any locking. This table is used to keep track of which
409 * readers are using data from which old transactions, so that we'll know
410 * when a particular old transaction is no longer in use. Old transactions
411 * that have discarded any data pages can then have those pages reclaimed
412 * for use by a later write transaction.
414 * The lock table is constructed such that reader slots are aligned with the
415 * processor's cache line size. Any slot is only ever used by one thread.
416 * This alignment guarantees that there will be no contention or cache
417 * thrashing as threads update their own slot info, and also eliminates
418 * any need for locking when accessing a slot.
420 * A writer thread will scan every slot in the table to determine the oldest
421 * outstanding reader transaction. Any freed pages older than this will be
422 * reclaimed by the writer. The writer doesn't use any locks when scanning
423 * this table. This means that there's no guarantee that the writer will
424 * see the most up-to-date reader info, but that's not required for correct
425 * operation - all we need is to know the upper bound on the oldest reader,
426 * we don't care at all about the newest reader. So the only consequence of
427 * reading stale information here is that old pages might hang around a
428 * while longer before being reclaimed. That's actually good anyway, because
429 * the longer we delay reclaiming old pages, the more likely it is that a
430 * string of contiguous pages can be found after coalescing old pages from
431 * many old transactions together.
433 * @todo We don't actually do such coalescing yet, we grab pages from one
434 * old transaction at a time.
437 /** Number of slots in the reader table.
438 * This value was chosen somewhat arbitrarily. 126 readers plus a
439 * couple mutexes fit exactly into 8KB on my development machine.
440 * Applications should set the table size using #mdb_env_set_maxreaders().
442 #define DEFAULT_READERS 126
444 /** The size of a CPU cache line in bytes. We want our lock structures
445 * aligned to this size to avoid false cache line sharing in the
447 * This value works for most CPUs. For Itanium this should be 128.
453 /** The information we store in a single slot of the reader table.
454 * In addition to a transaction ID, we also record the process and
455 * thread ID that owns a slot, so that we can detect stale information,
456 * e.g. threads or processes that went away without cleaning up.
457 * @note We currently don't check for stale records. We simply re-init
458 * the table when we know that we're the only process opening the
461 typedef struct MDB_rxbody {
462 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
463 * Multiple readers that start at the same time will probably have the
464 * same ID here. Again, it's not important to exclude them from
465 * anything; all we need to know is which version of the DB they
466 * started from so we can avoid overwriting any data used in that
467 * particular version.
470 /** The process ID of the process owning this reader txn. */
472 /** The thread ID of the thread owning this txn. */
476 /** The actual reader record, with cacheline padding. */
477 typedef struct MDB_reader {
480 /** shorthand for mrb_txnid */
481 #define mr_txnid mru.mrx.mrb_txnid
482 #define mr_pid mru.mrx.mrb_pid
483 #define mr_tid mru.mrx.mrb_tid
484 /** cache line alignment */
485 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
489 /** The header for the reader table.
490 * The table resides in a memory-mapped file. (This is a different file
491 * than is used for the main database.)
493 * For POSIX the actual mutexes reside in the shared memory of this
494 * mapped file. On Windows, mutexes are named objects allocated by the
495 * kernel; we store the mutex names in this mapped file so that other
496 * processes can grab them. This same approach is also used on
497 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
498 * process-shared POSIX mutexes. For these cases where a named object
499 * is used, the object name is derived from a 64 bit FNV hash of the
500 * environment pathname. As such, naming collisions are extremely
501 * unlikely. If a collision occurs, the results are unpredictable.
503 typedef struct MDB_txbody {
504 /** Stamp identifying this as an MDB file. It must be set
507 /** Version number of this lock file. Must be set to #MDB_VERSION. */
508 uint32_t mtb_version;
509 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
510 char mtb_rmname[MNAME_LEN];
512 /** Mutex protecting access to this table.
513 * This is the reader lock that #LOCK_MUTEX_R acquires.
515 pthread_mutex_t mtb_mutex;
517 /** The ID of the last transaction committed to the database.
518 * This is recorded here only for convenience; the value can always
519 * be determined by reading the main database meta pages.
522 /** The number of slots that have been used in the reader table.
523 * This always records the maximum count, it is not decremented
524 * when readers release their slots.
526 unsigned mtb_numreaders;
529 /** The actual reader table definition. */
530 typedef struct MDB_txninfo {
533 #define mti_magic mt1.mtb.mtb_magic
534 #define mti_version mt1.mtb.mtb_version
535 #define mti_mutex mt1.mtb.mtb_mutex
536 #define mti_rmname mt1.mtb.mtb_rmname
537 #define mti_txnid mt1.mtb.mtb_txnid
538 #define mti_numreaders mt1.mtb.mtb_numreaders
539 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
542 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
543 char mt2_wmname[MNAME_LEN];
544 #define mti_wmname mt2.mt2_wmname
546 pthread_mutex_t mt2_wmutex;
547 #define mti_wmutex mt2.mt2_wmutex
549 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
551 MDB_reader mti_readers[1];
555 /** Common header for all page types.
556 * Overflow records occupy a number of contiguous pages with no
557 * headers on any page after the first.
559 typedef struct MDB_page {
560 #define mp_pgno mp_p.p_pgno
561 #define mp_next mp_p.p_next
563 pgno_t p_pgno; /**< page number */
564 void * p_next; /**< for in-memory list of freed structs */
567 /** @defgroup mdb_page Page Flags
569 * Flags for the page headers.
572 #define P_BRANCH 0x01 /**< branch page */
573 #define P_LEAF 0x02 /**< leaf page */
574 #define P_OVERFLOW 0x04 /**< overflow page */
575 #define P_META 0x08 /**< meta page */
576 #define P_DIRTY 0x10 /**< dirty page */
577 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
578 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
580 uint16_t mp_flags; /**< @ref mdb_page */
581 #define mp_lower mp_pb.pb.pb_lower
582 #define mp_upper mp_pb.pb.pb_upper
583 #define mp_pages mp_pb.pb_pages
586 indx_t pb_lower; /**< lower bound of free space */
587 indx_t pb_upper; /**< upper bound of free space */
589 uint32_t pb_pages; /**< number of overflow pages */
591 indx_t mp_ptrs[1]; /**< dynamic size */
594 /** Size of the page header, excluding dynamic data at the end */
595 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
597 /** Address of first usable data byte in a page, after the header */
598 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
600 /** Number of nodes on a page */
601 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
603 /** The amount of space remaining in the page */
604 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
606 /** The percentage of space used in the page, in tenths of a percent. */
607 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
608 ((env)->me_psize - PAGEHDRSZ))
609 /** The minimum page fill factor, in tenths of a percent.
610 * Pages emptier than this are candidates for merging.
612 #define FILL_THRESHOLD 250
614 /** Test if a page is a leaf page */
615 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
616 /** Test if a page is a LEAF2 page */
617 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
618 /** Test if a page is a branch page */
619 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
620 /** Test if a page is an overflow page */
621 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
622 /** Test if a page is a sub page */
623 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
625 /** The number of overflow pages needed to store the given size. */
626 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
628 /** Header for a single key/data pair within a page.
629 * We guarantee 2-byte alignment for nodes.
631 typedef struct MDB_node {
632 /** lo and hi are used for data size on leaf nodes and for
633 * child pgno on branch nodes. On 64 bit platforms, flags
634 * is also used for pgno. (Branch nodes have no flags).
635 * They are in host byte order in case that lets some
636 * accesses be optimized into a 32-bit word access.
638 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
639 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
640 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
641 /** @defgroup mdb_node Node Flags
643 * Flags for node headers.
646 #define F_BIGDATA 0x01 /**< data put on overflow page */
647 #define F_SUBDATA 0x02 /**< data is a sub-database */
648 #define F_DUPDATA 0x04 /**< data has duplicates */
650 /** valid flags for #mdb_node_add() */
651 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
654 unsigned short mn_flags; /**< @ref mdb_node */
655 unsigned short mn_ksize; /**< key size */
656 char mn_data[1]; /**< key and data are appended here */
659 /** Size of the node header, excluding dynamic data at the end */
660 #define NODESIZE offsetof(MDB_node, mn_data)
662 /** Bit position of top word in page number, for shifting mn_flags */
663 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
665 /** Size of a node in a branch page with a given key.
666 * This is just the node header plus the key, there is no data.
668 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
670 /** Size of a node in a leaf page with a given key and data.
671 * This is node header plus key plus data size.
673 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
675 /** Address of node \b i in page \b p */
676 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
678 /** Address of the key for the node */
679 #define NODEKEY(node) (void *)((node)->mn_data)
681 /** Address of the data for a node */
682 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
684 /** Get the page number pointed to by a branch node */
685 #define NODEPGNO(node) \
686 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
687 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
688 /** Set the page number in a branch node */
689 #define SETPGNO(node,pgno) do { \
690 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
691 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
693 /** Get the size of the data in a leaf node */
694 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
695 /** Set the size of the data for a leaf node */
696 #define SETDSZ(node,size) do { \
697 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
698 /** The size of a key in a node */
699 #define NODEKSZ(node) ((node)->mn_ksize)
701 /** Copy a page number from src to dst */
703 #define COPY_PGNO(dst,src) dst = src
705 #if SIZE_MAX > 4294967295UL
706 #define COPY_PGNO(dst,src) do { \
707 unsigned short *s, *d; \
708 s = (unsigned short *)&(src); \
709 d = (unsigned short *)&(dst); \
716 #define COPY_PGNO(dst,src) do { \
717 unsigned short *s, *d; \
718 s = (unsigned short *)&(src); \
719 d = (unsigned short *)&(dst); \
725 /** The address of a key in a LEAF2 page.
726 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
727 * There are no node headers, keys are stored contiguously.
729 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
731 /** Set the \b node's key into \b key, if requested. */
732 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
733 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
735 /** Information about a single database in the environment. */
736 typedef struct MDB_db {
737 uint32_t md_pad; /**< also ksize for LEAF2 pages */
738 uint16_t md_flags; /**< @ref mdb_dbi_open */
739 uint16_t md_depth; /**< depth of this tree */
740 pgno_t md_branch_pages; /**< number of internal pages */
741 pgno_t md_leaf_pages; /**< number of leaf pages */
742 pgno_t md_overflow_pages; /**< number of overflow pages */
743 size_t md_entries; /**< number of data items */
744 pgno_t md_root; /**< the root page of this tree */
747 /** Handle for the DB used to track free pages. */
749 /** Handle for the default DB. */
752 /** Meta page content. */
753 typedef struct MDB_meta {
754 /** Stamp identifying this as an MDB file. It must be set
757 /** Version number of this lock file. Must be set to #MDB_VERSION. */
759 void *mm_address; /**< address for fixed mapping */
760 size_t mm_mapsize; /**< size of mmap region */
761 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
762 /** The size of pages used in this DB */
763 #define mm_psize mm_dbs[0].md_pad
764 /** Any persistent environment flags. @ref mdb_env */
765 #define mm_flags mm_dbs[0].md_flags
766 pgno_t mm_last_pg; /**< last used page in file */
767 txnid_t mm_txnid; /**< txnid that committed this page */
770 /** Buffer for a stack-allocated dirty page.
771 * The members define size and alignment, and silence type
772 * aliasing warnings. They are not used directly; that could
773 * mean incorrectly using several union members in parallel.
775 typedef union MDB_pagebuf {
776 char mb_raw[MDB_PAGESIZE];
779 char mm_pad[PAGEHDRSZ];
784 /** Auxiliary DB info.
785 * The information here is mostly static/read-only. There is
786 * only a single copy of this record in the environment.
788 typedef struct MDB_dbx {
789 MDB_val md_name; /**< name of the database */
790 MDB_cmp_func *md_cmp; /**< function for comparing keys */
791 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
792 MDB_rel_func *md_rel; /**< user relocate function */
793 void *md_relctx; /**< user-provided context for md_rel */
796 /** A database transaction.
797 * Every operation requires a transaction handle.
800 MDB_txn *mt_parent; /**< parent of a nested txn */
801 MDB_txn *mt_child; /**< nested txn under this txn */
802 pgno_t mt_next_pgno; /**< next unallocated page */
803 /** The ID of this transaction. IDs are integers incrementing from 1.
804 * Only committed write transactions increment the ID. If a transaction
805 * aborts, the ID may be re-used by the next writer.
808 MDB_env *mt_env; /**< the DB environment */
809 /** The list of pages that became unused during this transaction.
813 MDB_ID2L dirty_list; /**< modified pages */
814 MDB_reader *reader; /**< this thread's slot in the reader table */
816 /** Array of records for each DB known in the environment. */
818 /** Array of MDB_db records for each known DB */
820 /** @defgroup mt_dbflag Transaction DB Flags
824 #define DB_DIRTY 0x01 /**< DB was written in this txn */
825 #define DB_STALE 0x02 /**< DB record is older than txnID */
827 /** Array of cursors for each DB */
828 MDB_cursor **mt_cursors;
829 /** Array of flags for each DB */
830 unsigned char *mt_dbflags;
831 /** Number of DB records in use. This number only ever increments;
832 * we don't decrement it when individual DB handles are closed.
836 /** @defgroup mdb_txn Transaction Flags
840 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
841 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
842 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
844 unsigned int mt_flags; /**< @ref mdb_txn */
845 /** Tracks which of the two meta pages was used at the start
846 * of this transaction.
848 unsigned int mt_toggle;
851 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
852 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
853 * raise this on a 64 bit machine.
855 #define CURSOR_STACK 32
859 /** Cursors are used for all DB operations */
861 /** Next cursor on this DB in this txn */
863 /** Original cursor if this is a shadow */
865 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
866 struct MDB_xcursor *mc_xcursor;
867 /** The transaction that owns this cursor */
869 /** The database handle this cursor operates on */
871 /** The database record for this cursor */
873 /** The database auxiliary record for this cursor */
875 /** The @ref mt_dbflag for this database */
876 unsigned char *mc_dbflag;
877 unsigned short mc_snum; /**< number of pushed pages */
878 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
879 /** @defgroup mdb_cursor Cursor Flags
881 * Cursor state flags.
884 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
885 #define C_EOF 0x02 /**< No more data */
886 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
887 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
888 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
889 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
891 unsigned int mc_flags; /**< @ref mdb_cursor */
892 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
893 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
896 /** Context for sorted-dup records.
897 * We could have gone to a fully recursive design, with arbitrarily
898 * deep nesting of sub-databases. But for now we only handle these
899 * levels - main DB, optional sub-DB, sorted-duplicate DB.
901 typedef struct MDB_xcursor {
902 /** A sub-cursor for traversing the Dup DB */
903 MDB_cursor mx_cursor;
904 /** The database record for this Dup DB */
906 /** The auxiliary DB record for this Dup DB */
908 /** The @ref mt_dbflag for this Dup DB */
909 unsigned char mx_dbflag;
912 /** A set of pages freed by an earlier transaction. */
913 typedef struct MDB_oldpages {
914 /** Usually we only read one record from the FREEDB at a time, but
915 * in case we read more, this will chain them together.
917 struct MDB_oldpages *mo_next;
918 /** The ID of the transaction in which these pages were freed. */
920 /** An #MDB_IDL of the pages */
921 pgno_t mo_pages[1]; /* dynamic */
924 /** The database environment. */
926 HANDLE me_fd; /**< The main data file */
927 HANDLE me_lfd; /**< The lock file */
928 HANDLE me_mfd; /**< just for writing the meta pages */
929 /** Failed to update the meta page. Probably an I/O error. */
930 #define MDB_FATAL_ERROR 0x80000000U
931 /** Read-only Filesystem. Allow read access, no locking. */
932 #define MDB_ROFS 0x40000000U
933 /** Some fields are initialized. */
934 #define MDB_ENV_ACTIVE 0x20000000U
935 uint32_t me_flags; /**< @ref mdb_env */
936 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
937 unsigned int me_maxreaders; /**< size of the reader table */
938 unsigned int me_numreaders; /**< max numreaders set by this env */
939 MDB_dbi me_numdbs; /**< number of DBs opened */
940 MDB_dbi me_maxdbs; /**< size of the DB table */
941 pid_t me_pid; /**< process ID of this env */
942 char *me_path; /**< path to the DB files */
943 char *me_map; /**< the memory map of the data file */
944 MDB_txninfo *me_txns; /**< the memory map of the lock file */
945 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
946 MDB_txn *me_txn; /**< current write transaction */
947 size_t me_mapsize; /**< size of the data memory map */
948 off_t me_size; /**< current file size */
949 pgno_t me_maxpg; /**< me_mapsize / me_psize */
950 txnid_t me_pgfirst; /**< ID of first old page record we used */
951 txnid_t me_pglast; /**< ID of last old page record we used */
952 MDB_dbx *me_dbxs; /**< array of static DB info */
953 uint16_t *me_dbflags; /**< array of DB flags */
954 MDB_oldpages *me_pghead; /**< list of old page records */
955 MDB_oldpages *me_pgfree; /**< list of page records to free */
956 pthread_key_t me_txkey; /**< thread-key for readers */
957 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
958 /** IDL of pages that became unused in a write txn */
960 /** ID2L of pages that were written during a write txn */
961 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
963 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
965 #elif defined(MDB_USE_POSIX_SEM)
966 sem_t *me_rmutex; /* Shared mutexes are not supported */
970 /** max number of pages to commit in one writev() call */
971 #define MDB_COMMIT_PAGES 64
972 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
973 #undef MDB_COMMIT_PAGES
974 #define MDB_COMMIT_PAGES IOV_MAX
977 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
978 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
979 static int mdb_page_touch(MDB_cursor *mc);
981 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
982 static int mdb_page_search_root(MDB_cursor *mc,
983 MDB_val *key, int modify);
984 #define MDB_PS_MODIFY 1
985 #define MDB_PS_ROOTONLY 2
986 static int mdb_page_search(MDB_cursor *mc,
987 MDB_val *key, int flags);
988 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
990 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
991 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
992 pgno_t newpgno, unsigned int nflags);
994 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
995 static int mdb_env_pick_meta(const MDB_env *env);
996 static int mdb_env_write_meta(MDB_txn *txn);
997 static void mdb_env_close0(MDB_env *env, int excl);
999 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1000 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1001 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1002 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1003 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1004 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1005 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1006 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1007 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1009 static int mdb_rebalance(MDB_cursor *mc);
1010 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
1012 static void mdb_cursor_pop(MDB_cursor *mc);
1013 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1015 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1016 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1017 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1018 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1019 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1021 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1022 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1024 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1025 static void mdb_xcursor_init0(MDB_cursor *mc);
1026 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1028 static int mdb_drop0(MDB_cursor *mc, int subs);
1029 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1032 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1036 static SECURITY_DESCRIPTOR mdb_null_sd;
1037 static SECURITY_ATTRIBUTES mdb_all_sa;
1038 static int mdb_sec_inited;
1041 /** Return the library version info. */
1043 mdb_version(int *major, int *minor, int *patch)
1045 if (major) *major = MDB_VERSION_MAJOR;
1046 if (minor) *minor = MDB_VERSION_MINOR;
1047 if (patch) *patch = MDB_VERSION_PATCH;
1048 return MDB_VERSION_STRING;
1051 /** Table of descriptions for MDB @ref errors */
1052 static char *const mdb_errstr[] = {
1053 "MDB_KEYEXIST: Key/data pair already exists",
1054 "MDB_NOTFOUND: No matching key/data pair found",
1055 "MDB_PAGE_NOTFOUND: Requested page not found",
1056 "MDB_CORRUPTED: Located page was wrong type",
1057 "MDB_PANIC: Update of meta page failed",
1058 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1059 "MDB_INVALID: File is not an MDB file",
1060 "MDB_MAP_FULL: Environment mapsize limit reached",
1061 "MDB_DBS_FULL: Environment maxdbs limit reached",
1062 "MDB_READERS_FULL: Environment maxreaders limit reached",
1063 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1064 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1065 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1066 "MDB_PAGE_FULL: Internal error - page has no more space"
1070 mdb_strerror(int err)
1074 return ("Successful return: 0");
1076 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1077 i = err - MDB_KEYEXIST;
1078 return mdb_errstr[i];
1081 return strerror(err);
1085 /** Display a key in hexadecimal and return the address of the result.
1086 * @param[in] key the key to display
1087 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1088 * @return The key in hexadecimal form.
1091 mdb_dkey(MDB_val *key, char *buf)
1094 unsigned char *c = key->mv_data;
1096 if (key->mv_size > MDB_MAXKEYSIZE)
1097 return "MDB_MAXKEYSIZE";
1098 /* may want to make this a dynamic check: if the key is mostly
1099 * printable characters, print it as-is instead of converting to hex.
1103 for (i=0; i<key->mv_size; i++)
1104 ptr += sprintf(ptr, "%02x", *c++);
1106 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1111 /** Display all the keys in the page. */
1113 mdb_page_list(MDB_page *mp)
1116 unsigned int i, nkeys, nsize;
1120 nkeys = NUMKEYS(mp);
1121 fprintf(stderr, "numkeys %d\n", nkeys);
1122 for (i=0; i<nkeys; i++) {
1123 node = NODEPTR(mp, i);
1124 key.mv_size = node->mn_ksize;
1125 key.mv_data = node->mn_data;
1126 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1127 if (F_ISSET(node->mn_flags, F_BIGDATA))
1128 nsize += sizeof(pgno_t);
1130 nsize += NODEDSZ(node);
1131 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1136 mdb_cursor_chk(MDB_cursor *mc)
1142 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1143 for (i=0; i<mc->mc_top; i++) {
1145 node = NODEPTR(mp, mc->mc_ki[i]);
1146 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1149 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1155 /** Count all the pages in each DB and in the freelist
1156 * and make sure it matches the actual number of pages
1159 static void mdb_audit(MDB_txn *txn)
1163 MDB_ID freecount, count;
1168 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1169 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1170 freecount += *(MDB_ID *)data.mv_data;
1173 for (i = 0; i<txn->mt_numdbs; i++) {
1174 MDB_xcursor mx, *mxp;
1175 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1176 mdb_cursor_init(&mc, txn, i, mxp);
1177 if (txn->mt_dbs[i].md_root == P_INVALID)
1179 count += txn->mt_dbs[i].md_branch_pages +
1180 txn->mt_dbs[i].md_leaf_pages +
1181 txn->mt_dbs[i].md_overflow_pages;
1182 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1183 mdb_page_search(&mc, NULL, 0);
1187 mp = mc.mc_pg[mc.mc_top];
1188 for (j=0; j<NUMKEYS(mp); j++) {
1189 MDB_node *leaf = NODEPTR(mp, j);
1190 if (leaf->mn_flags & F_SUBDATA) {
1192 memcpy(&db, NODEDATA(leaf), sizeof(db));
1193 count += db.md_branch_pages + db.md_leaf_pages +
1194 db.md_overflow_pages;
1198 while (mdb_cursor_sibling(&mc, 1) == 0);
1201 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1202 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1203 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1209 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1211 return txn->mt_dbxs[dbi].md_cmp(a, b);
1215 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1217 if (txn->mt_dbxs[dbi].md_dcmp)
1218 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1220 return EINVAL; /* too bad you can't distinguish this from a valid result */
1223 /** Allocate a single page.
1224 * Re-use old malloc'd pages first, otherwise just malloc.
1227 mdb_page_malloc(MDB_cursor *mc) {
1229 size_t sz = mc->mc_txn->mt_env->me_psize;
1230 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1231 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1232 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1233 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1234 } else if ((ret = malloc(sz)) != NULL) {
1235 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1241 mdb_page_free(MDB_env *env, MDB_page *mp)
1243 mp->mp_next = env->me_dpages;
1244 VGMEMP_FREE(env, mp);
1245 env->me_dpages = mp;
1248 /** Allocate pages for writing.
1249 * If there are free pages available from older transactions, they
1250 * will be re-used first. Otherwise a new page will be allocated.
1251 * @param[in] mc cursor A cursor handle identifying the transaction and
1252 * database for which we are allocating.
1253 * @param[in] num the number of pages to allocate.
1254 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1255 * will always be satisfied by a single contiguous chunk of memory.
1256 * @return 0 on success, non-zero on failure.
1259 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1261 MDB_txn *txn = mc->mc_txn;
1263 pgno_t pgno = P_INVALID;
1265 txnid_t oldest = 0, last;
1270 /* If our dirty list is already full, we can't do anything */
1271 if (txn->mt_u.dirty_list[0].mid >= MDB_IDL_UM_MAX)
1272 return MDB_TXN_FULL;
1274 /* The free list won't have any content at all until txn 2 has
1275 * committed. The pages freed by txn 2 will be unreferenced
1276 * after txn 3 commits, and so will be safe to re-use in txn 4.
1278 if (txn->mt_txnid > 3) {
1280 if (!txn->mt_env->me_pghead &&
1281 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1282 /* See if there's anything in the free DB */
1289 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1290 if (!txn->mt_env->me_pgfirst) {
1291 mdb_page_search(&m2, NULL, 0);
1292 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1293 kptr = (txnid_t *)NODEKEY(leaf);
1300 last = txn->mt_env->me_pglast + 1;
1302 key.mv_data = &last;
1303 key.mv_size = sizeof(last);
1304 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1307 last = *(txnid_t *)key.mv_data;
1313 oldest = txn->mt_txnid - 1;
1314 nr = txn->mt_env->me_txns->mti_numreaders;
1315 r = txn->mt_env->me_txns->mti_readers;
1316 for (i=0; i<nr; i++) {
1317 if (!r[i].mr_pid) continue;
1324 if (oldest > last) {
1325 /* It's usable, grab it.
1330 if (!txn->mt_env->me_pgfirst) {
1331 mdb_node_read(txn, leaf, &data);
1333 txn->mt_env->me_pglast = last;
1334 if (!txn->mt_env->me_pgfirst)
1335 txn->mt_env->me_pgfirst = last;
1336 idl = (MDB_ID *) data.mv_data;
1337 /* We might have a zero-length IDL due to freelist growth
1338 * during a prior commit
1340 if (!idl[0]) goto again;
1341 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1344 mop->mo_next = txn->mt_env->me_pghead;
1345 mop->mo_txnid = last;
1346 txn->mt_env->me_pghead = mop;
1347 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1352 DPRINTF("IDL read txn %zu root %zu num %zu",
1353 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1354 for (i=0; i<idl[0]; i++) {
1355 DPRINTF("IDL %zu", idl[i+1]);
1362 if (txn->mt_env->me_pghead) {
1363 MDB_oldpages *mop = txn->mt_env->me_pghead;
1366 int retry = 500, readit = 0, n2 = num-1;
1367 unsigned int i, j, k;
1369 /* If current list is too short, must fetch more and coalesce */
1370 if (mop->mo_pages[0] < (unsigned)num)
1373 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1375 /* If on freelist, don't try to read more. If what we have
1376 * right now isn't enough just use new pages.
1377 * TODO: get all of this working. Many circular dependencies...
1379 if (mc->mc_dbi == FREE_DBI) {
1389 last = mop->mo_txnid + 1;
1391 /* We haven't hit the readers list yet? */
1397 oldest = txn->mt_txnid - 1;
1398 nr = txn->mt_env->me_txns->mti_numreaders;
1399 r = txn->mt_env->me_txns->mti_readers;
1400 for (i=0; i<nr; i++) {
1401 if (!r[i].mr_pid) continue;
1408 /* There's nothing we can use on the freelist */
1409 if (oldest - last < 1)
1413 key.mv_data = &last;
1414 key.mv_size = sizeof(last);
1415 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1418 idl = (MDB_ID *) data.mv_data;
1419 mop2 = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - 2*sizeof(pgno_t) + MDB_IDL_SIZEOF(mop->mo_pages));
1422 /* merge in sorted order */
1423 i = idl[0]; j = mop->mo_pages[0]; mop2->mo_pages[0] = k = i+j;
1424 mop->mo_pages[0] = P_INVALID;
1425 while (i>0 || j>0) {
1426 if (i && idl[i] < mop->mo_pages[j])
1427 mop2->mo_pages[k--] = idl[i--];
1429 mop2->mo_pages[k--] = mop->mo_pages[j--];
1431 txn->mt_env->me_pglast = last;
1432 mop2->mo_txnid = last;
1433 mop2->mo_next = mop->mo_next;
1434 txn->mt_env->me_pghead = mop2;
1437 /* Keep trying to read until we have enough */
1438 if (mop->mo_pages[0] < (unsigned)num) {
1443 /* current list has enough pages, but are they contiguous? */
1444 for (i=mop->mo_pages[0]; i>=(unsigned)num; i--) {
1445 if (mop->mo_pages[i-n2] == mop->mo_pages[i] + n2) {
1446 pgno = mop->mo_pages[i];
1448 /* move any stragglers down */
1449 for (j=i+num; j<=mop->mo_pages[0]; j++)
1450 mop->mo_pages[i++] = mop->mo_pages[j];
1451 mop->mo_pages[0] -= num;
1456 /* Stop if we succeeded, or no more retries */
1457 if (!retry || pgno != P_INVALID)
1464 /* peel pages off tail, so we only have to truncate the list */
1465 pgno = MDB_IDL_LAST(mop->mo_pages);
1468 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1469 txn->mt_env->me_pghead = mop->mo_next;
1470 if (mc->mc_dbi == FREE_DBI) {
1471 mop->mo_next = txn->mt_env->me_pgfree;
1472 txn->mt_env->me_pgfree = mop;
1480 if (pgno == P_INVALID) {
1481 /* DB size is maxed out */
1482 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1483 DPUTS("DB size maxed out");
1484 return MDB_MAP_FULL;
1487 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1488 if (pgno == P_INVALID) {
1489 pgno = txn->mt_next_pgno;
1490 txn->mt_next_pgno += num;
1492 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1495 if (txn->mt_env->me_dpages && num == 1) {
1496 np = txn->mt_env->me_dpages;
1497 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1498 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1499 txn->mt_env->me_dpages = np->mp_next;
1501 size_t sz = txn->mt_env->me_psize * num;
1502 if ((np = malloc(sz)) == NULL)
1504 VGMEMP_ALLOC(txn->mt_env, np, sz);
1506 if (pgno == P_INVALID) {
1507 np->mp_pgno = txn->mt_next_pgno;
1508 txn->mt_next_pgno += num;
1513 mid.mid = np->mp_pgno;
1515 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1516 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1518 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1525 /** Copy a page: avoid copying unused portions of the page.
1526 * @param[in] dst page to copy into
1527 * @param[in] src page to copy from
1530 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1532 dst->mp_flags = src->mp_flags | P_DIRTY;
1533 dst->mp_pages = src->mp_pages;
1535 if (IS_LEAF2(src)) {
1536 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1538 unsigned int i, nkeys = NUMKEYS(src);
1539 for (i=0; i<nkeys; i++)
1540 dst->mp_ptrs[i] = src->mp_ptrs[i];
1541 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1542 psize - src->mp_upper);
1546 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1547 * @param[in] mc cursor pointing to the page to be touched
1548 * @return 0 on success, non-zero on failure.
1551 mdb_page_touch(MDB_cursor *mc)
1553 MDB_page *mp = mc->mc_pg[mc->mc_top];
1557 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1559 if ((rc = mdb_page_alloc(mc, 1, &np)))
1561 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1562 assert(mp->mp_pgno != np->mp_pgno);
1563 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1565 /* If page isn't full, just copy the used portion */
1566 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1569 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1571 np->mp_flags |= P_DIRTY;
1576 /* Adjust other cursors pointing to mp */
1577 if (mc->mc_flags & C_SUB) {
1578 MDB_cursor *m2, *m3;
1579 MDB_dbi dbi = mc->mc_dbi-1;
1581 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1582 if (m2 == mc) continue;
1583 m3 = &m2->mc_xcursor->mx_cursor;
1584 if (m3->mc_snum < mc->mc_snum) continue;
1585 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1586 m3->mc_pg[mc->mc_top] = mp;
1592 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1593 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1594 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1595 m2->mc_pg[mc->mc_top] = mp;
1599 mc->mc_pg[mc->mc_top] = mp;
1600 /** If this page has a parent, update the parent to point to
1604 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1606 mc->mc_db->md_root = mp->mp_pgno;
1607 } else if (mc->mc_txn->mt_parent) {
1610 /* If txn has a parent, make sure the page is in our
1613 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1614 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1615 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1616 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1617 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1618 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1619 mc->mc_pg[mc->mc_top] = mp;
1624 if (mc->mc_txn->mt_u.dirty_list[0].mid >= MDB_IDL_UM_MAX)
1625 return MDB_TXN_FULL;
1627 np = mdb_page_malloc(mc);
1630 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1631 mid.mid = np->mp_pgno;
1633 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1641 mdb_env_sync(MDB_env *env, int force)
1644 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1645 if (env->me_flags & MDB_WRITEMAP) {
1646 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1647 ? MS_ASYNC : MS_SYNC;
1648 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1651 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1655 if (MDB_FDATASYNC(env->me_fd))
1662 /** Make shadow copies of all of parent txn's cursors */
1664 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1666 MDB_cursor *mc, *m2;
1667 unsigned int i, j, size;
1669 for (i=0;i<src->mt_numdbs; i++) {
1670 if (src->mt_cursors[i]) {
1671 size = sizeof(MDB_cursor);
1672 if (src->mt_cursors[i]->mc_xcursor)
1673 size += sizeof(MDB_xcursor);
1674 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1681 mc->mc_db = &dst->mt_dbs[i];
1682 mc->mc_dbx = m2->mc_dbx;
1683 mc->mc_dbflag = &dst->mt_dbflags[i];
1684 mc->mc_snum = m2->mc_snum;
1685 mc->mc_top = m2->mc_top;
1686 mc->mc_flags = m2->mc_flags | C_SHADOW;
1687 for (j=0; j<mc->mc_snum; j++) {
1688 mc->mc_pg[j] = m2->mc_pg[j];
1689 mc->mc_ki[j] = m2->mc_ki[j];
1691 if (m2->mc_xcursor) {
1692 MDB_xcursor *mx, *mx2;
1693 mx = (MDB_xcursor *)(mc+1);
1694 mc->mc_xcursor = mx;
1695 mx2 = m2->mc_xcursor;
1696 mx->mx_db = mx2->mx_db;
1697 mx->mx_dbx = mx2->mx_dbx;
1698 mx->mx_dbflag = mx2->mx_dbflag;
1699 mx->mx_cursor.mc_txn = dst;
1700 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1701 mx->mx_cursor.mc_db = &mx->mx_db;
1702 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1703 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1704 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1705 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1706 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1707 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1708 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1709 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1712 mc->mc_xcursor = NULL;
1714 mc->mc_next = dst->mt_cursors[i];
1715 dst->mt_cursors[i] = mc;
1722 /** Merge shadow cursors back into parent's */
1724 mdb_cursor_merge(MDB_txn *txn)
1727 for (i=0; i<txn->mt_numdbs; i++) {
1728 if (txn->mt_cursors[i]) {
1730 while ((mc = txn->mt_cursors[i])) {
1731 txn->mt_cursors[i] = mc->mc_next;
1732 if (mc->mc_flags & C_SHADOW) {
1733 MDB_cursor *m2 = mc->mc_orig;
1735 m2->mc_snum = mc->mc_snum;
1736 m2->mc_top = mc->mc_top;
1737 for (j=0; j<mc->mc_snum; j++) {
1738 m2->mc_pg[j] = mc->mc_pg[j];
1739 m2->mc_ki[j] = mc->mc_ki[j];
1742 if (mc->mc_flags & C_ALLOCD)
1750 mdb_txn_reset0(MDB_txn *txn);
1752 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1753 * @param[in] txn the transaction handle to initialize
1754 * @return 0 on success, non-zero on failure. This can only
1755 * fail for read-only transactions, and then only if the
1756 * reader table is full.
1759 mdb_txn_renew0(MDB_txn *txn)
1761 MDB_env *env = txn->mt_env;
1766 txn->mt_numdbs = env->me_numdbs;
1767 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1769 if (txn->mt_flags & MDB_TXN_RDONLY) {
1770 if (env->me_flags & MDB_ROFS) {
1771 i = mdb_env_pick_meta(env);
1772 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1773 txn->mt_u.reader = NULL;
1775 MDB_reader *r = pthread_getspecific(env->me_txkey);
1777 pid_t pid = env->me_pid;
1778 pthread_t tid = pthread_self();
1781 for (i=0; i<env->me_txns->mti_numreaders; i++)
1782 if (env->me_txns->mti_readers[i].mr_pid == 0)
1784 if (i == env->me_maxreaders) {
1785 UNLOCK_MUTEX_R(env);
1786 return MDB_READERS_FULL;
1788 env->me_txns->mti_readers[i].mr_pid = pid;
1789 env->me_txns->mti_readers[i].mr_tid = tid;
1790 if (i >= env->me_txns->mti_numreaders)
1791 env->me_txns->mti_numreaders = i+1;
1792 /* Save numreaders for un-mutexed mdb_env_close() */
1793 env->me_numreaders = env->me_txns->mti_numreaders;
1794 UNLOCK_MUTEX_R(env);
1795 r = &env->me_txns->mti_readers[i];
1796 if ((rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1797 env->me_txns->mti_readers[i].mr_pid = 0;
1801 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1802 txn->mt_u.reader = r;
1804 txn->mt_toggle = txn->mt_txnid & 1;
1805 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1809 txn->mt_txnid = env->me_txns->mti_txnid;
1810 txn->mt_toggle = txn->mt_txnid & 1;
1811 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1814 if (txn->mt_txnid == mdb_debug_start)
1817 txn->mt_u.dirty_list = env->me_dirty_list;
1818 txn->mt_u.dirty_list[0].mid = 0;
1819 txn->mt_free_pgs = env->me_free_pgs;
1820 txn->mt_free_pgs[0] = 0;
1824 /* Copy the DB info and flags */
1825 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1826 for (i=2; i<txn->mt_numdbs; i++)
1827 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1828 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1829 if (txn->mt_numdbs > 2)
1830 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1836 mdb_txn_renew(MDB_txn *txn)
1840 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1843 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1844 DPUTS("environment had fatal error, must shutdown!");
1848 rc = mdb_txn_renew0(txn);
1849 if (rc == MDB_SUCCESS) {
1850 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1851 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1852 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1858 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1863 if (env->me_flags & MDB_FATAL_ERROR) {
1864 DPUTS("environment had fatal error, must shutdown!");
1867 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1870 /* Nested transactions: Max 1 child, write txns only, no writemap */
1871 if (parent->mt_child ||
1872 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1873 (env->me_flags & MDB_WRITEMAP))
1878 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1879 if (!(flags & MDB_RDONLY))
1880 size += env->me_maxdbs * sizeof(MDB_cursor *);
1882 if ((txn = calloc(1, size)) == NULL) {
1883 DPRINTF("calloc: %s", strerror(ErrCode()));
1886 txn->mt_dbs = (MDB_db *)(txn+1);
1887 if (flags & MDB_RDONLY) {
1888 txn->mt_flags |= MDB_TXN_RDONLY;
1889 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1891 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1892 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1897 txn->mt_free_pgs = mdb_midl_alloc();
1898 if (!txn->mt_free_pgs) {
1902 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1903 if (!txn->mt_u.dirty_list) {
1904 free(txn->mt_free_pgs);
1908 txn->mt_txnid = parent->mt_txnid;
1909 txn->mt_toggle = parent->mt_toggle;
1910 txn->mt_u.dirty_list[0].mid = 0;
1911 txn->mt_free_pgs[0] = 0;
1912 txn->mt_next_pgno = parent->mt_next_pgno;
1913 parent->mt_child = txn;
1914 txn->mt_parent = parent;
1915 txn->mt_numdbs = parent->mt_numdbs;
1916 txn->mt_dbxs = parent->mt_dbxs;
1917 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1918 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1919 mdb_cursor_shadow(parent, txn);
1922 rc = mdb_txn_renew0(txn);
1928 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1929 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1930 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1936 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1937 * @param[in] txn the transaction handle to reset
1940 mdb_txn_reset0(MDB_txn *txn)
1942 MDB_env *env = txn->mt_env;
1944 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1945 if (!(env->me_flags & MDB_ROFS))
1946 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1952 /* close(free) all cursors */
1953 for (i=0; i<txn->mt_numdbs; i++) {
1954 if (txn->mt_cursors[i]) {
1956 while ((mc = txn->mt_cursors[i])) {
1957 txn->mt_cursors[i] = mc->mc_next;
1958 if (mc->mc_flags & C_ALLOCD)
1964 if (!(env->me_flags & MDB_WRITEMAP)) {
1965 /* return all dirty pages to dpage list */
1966 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1967 dp = txn->mt_u.dirty_list[i].mptr;
1968 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1969 mdb_page_free(txn->mt_env, dp);
1971 /* large pages just get freed directly */
1972 VGMEMP_FREE(txn->mt_env, dp);
1978 if (txn->mt_parent) {
1979 txn->mt_parent->mt_child = NULL;
1980 mdb_midl_free(txn->mt_free_pgs);
1981 free(txn->mt_u.dirty_list);
1984 if (mdb_midl_shrink(&txn->mt_free_pgs))
1985 env->me_free_pgs = txn->mt_free_pgs;
1988 while ((mop = txn->mt_env->me_pghead)) {
1989 txn->mt_env->me_pghead = mop->mo_next;
1992 txn->mt_env->me_pgfirst = 0;
1993 txn->mt_env->me_pglast = 0;
1996 /* The writer mutex was locked in mdb_txn_begin. */
1997 UNLOCK_MUTEX_W(env);
2002 mdb_txn_reset(MDB_txn *txn)
2007 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2008 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2009 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2011 mdb_txn_reset0(txn);
2015 mdb_txn_abort(MDB_txn *txn)
2020 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2021 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2022 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2025 mdb_txn_abort(txn->mt_child);
2027 mdb_txn_reset0(txn);
2032 mdb_txn_commit(MDB_txn *txn)
2040 pgno_t next, freecnt;
2043 assert(txn != NULL);
2044 assert(txn->mt_env != NULL);
2046 if (txn->mt_child) {
2047 mdb_txn_commit(txn->mt_child);
2048 txn->mt_child = NULL;
2053 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2054 if (txn->mt_numdbs > env->me_numdbs) {
2055 /* update the DB flags */
2057 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2058 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2065 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2066 DPUTS("error flag is set, can't commit");
2068 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2073 if (txn->mt_parent) {
2079 /* Merge (and close) our cursors with parent's */
2080 mdb_cursor_merge(txn);
2082 /* Update parent's DB table */
2083 ip = &txn->mt_parent->mt_dbs[2];
2084 jp = &txn->mt_dbs[2];
2085 for (i = 2; i < txn->mt_numdbs; i++) {
2086 if (ip->md_root != jp->md_root)
2090 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2092 /* Append our free list to parent's */
2093 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
2095 mdb_midl_free(txn->mt_free_pgs);
2097 /* Merge our dirty list with parent's */
2098 dst = txn->mt_parent->mt_u.dirty_list;
2099 src = txn->mt_u.dirty_list;
2100 x = mdb_mid2l_search(dst, src[1].mid);
2101 for (y=1; y<=src[0].mid; y++) {
2102 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
2106 dst[x].mptr = src[y].mptr;
2109 for (; y<=src[0].mid; y++) {
2110 if (++x >= MDB_IDL_UM_MAX) {
2112 return MDB_TXN_FULL;
2117 free(txn->mt_u.dirty_list);
2118 txn->mt_parent->mt_child = NULL;
2123 if (txn != env->me_txn) {
2124 DPUTS("attempt to commit unknown transaction");
2129 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2132 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2133 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2135 /* Update DB root pointers. Their pages have already been
2136 * touched so this is all in-place and cannot fail.
2138 if (txn->mt_numdbs > 2) {
2141 data.mv_size = sizeof(MDB_db);
2143 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2144 for (i = 2; i < txn->mt_numdbs; i++) {
2145 if (txn->mt_dbflags[i] & DB_DIRTY) {
2146 data.mv_data = &txn->mt_dbs[i];
2147 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2152 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2154 /* should only be one record now */
2155 if (env->me_pghead || env->me_pgfirst) {
2156 /* make sure first page of freeDB is touched and on freelist */
2157 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2158 if (rc && rc != MDB_NOTFOUND) {
2165 /* Delete IDLs we used from the free list */
2166 if (env->me_pgfirst) {
2171 key.mv_size = sizeof(cur);
2172 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
2175 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
2176 rc = mdb_cursor_del(&mc, 0);
2180 env->me_pgfirst = 0;
2184 /* save to free list */
2186 freecnt = txn->mt_free_pgs[0];
2187 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2190 /* make sure last page of freeDB is touched and on freelist */
2191 key.mv_size = MDB_MAXKEYSIZE+1;
2193 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2194 if (rc && rc != MDB_NOTFOUND)
2200 MDB_IDL idl = txn->mt_free_pgs;
2201 mdb_midl_sort(txn->mt_free_pgs);
2202 DPRINTF("IDL write txn %zu root %zu num %zu",
2203 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2204 for (i=1; i<=idl[0]; i++) {
2205 DPRINTF("IDL %zu", idl[i]);
2209 /* write to last page of freeDB */
2210 key.mv_size = sizeof(pgno_t);
2211 key.mv_data = &txn->mt_txnid;
2212 data.mv_data = txn->mt_free_pgs;
2213 /* The free list can still grow during this call,
2214 * despite the pre-emptive touches above. So check
2215 * and make sure the entire thing got written.
2218 freecnt = txn->mt_free_pgs[0];
2219 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2220 mdb_midl_sort(txn->mt_free_pgs);
2221 rc = mdb_cursor_put(&mc, &key, &data, 0);
2224 } while (freecnt != txn->mt_free_pgs[0]);
2226 /* should only be one record now */
2228 if (env->me_pghead) {
2234 mop = env->me_pghead;
2236 key.mv_size = sizeof(id);
2238 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2239 data.mv_data = mop->mo_pages;
2240 orig = mop->mo_pages[0];
2241 /* These steps may grow the freelist again
2242 * due to freed overflow pages...
2244 rc = mdb_cursor_put(&mc, &key, &data, 0);
2247 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2248 /* could have been used again here */
2249 if (mop->mo_pages[0] != orig) {
2250 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2251 data.mv_data = mop->mo_pages;
2253 rc = mdb_cursor_put(&mc, &key, &data, 0);
2258 /* was completely used up */
2259 rc = mdb_cursor_del(&mc, 0);
2265 env->me_pgfirst = 0;
2269 /* Check for growth of freelist again */
2270 if (freecnt != txn->mt_free_pgs[0])
2273 if (env->me_pghead) {
2274 free(env->me_pghead);
2275 env->me_pghead = NULL;
2278 while (env->me_pgfree) {
2279 MDB_oldpages *mop = env->me_pgfree;
2280 env->me_pgfree = mop->mo_next;
2284 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2285 if (mdb_midl_shrink(&txn->mt_free_pgs))
2286 env->me_free_pgs = txn->mt_free_pgs;
2293 if (env->me_flags & MDB_WRITEMAP) {
2294 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2295 dp = txn->mt_u.dirty_list[i].mptr;
2296 /* clear dirty flag */
2297 dp->mp_flags &= ~P_DIRTY;
2298 txn->mt_u.dirty_list[i].mid = 0;
2300 txn->mt_u.dirty_list[0].mid = 0;
2304 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2310 /* Windows actually supports scatter/gather I/O, but only on
2311 * unbuffered file handles. Since we're relying on the OS page
2312 * cache for all our data, that's self-defeating. So we just
2313 * write pages one at a time. We use the ov structure to set
2314 * the write offset, to at least save the overhead of a Seek
2318 memset(&ov, 0, sizeof(ov));
2319 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2321 dp = txn->mt_u.dirty_list[i].mptr;
2322 DPRINTF("committing page %zu", dp->mp_pgno);
2323 size = dp->mp_pgno * env->me_psize;
2324 ov.Offset = size & 0xffffffff;
2325 ov.OffsetHigh = size >> 16;
2326 ov.OffsetHigh >>= 16;
2327 /* clear dirty flag */
2328 dp->mp_flags &= ~P_DIRTY;
2329 wsize = env->me_psize;
2330 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2331 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2334 DPRINTF("WriteFile: %d", n);
2341 struct iovec iov[MDB_COMMIT_PAGES];
2345 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2346 dp = txn->mt_u.dirty_list[i].mptr;
2347 if (dp->mp_pgno != next) {
2349 rc = writev(env->me_fd, iov, n);
2353 DPUTS("short write, filesystem full?");
2355 DPRINTF("writev: %s", strerror(n));
2362 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2365 DPRINTF("committing page %zu", dp->mp_pgno);
2366 iov[n].iov_len = env->me_psize;
2367 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2368 iov[n].iov_base = (char *)dp;
2369 size += iov[n].iov_len;
2370 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2371 /* clear dirty flag */
2372 dp->mp_flags &= ~P_DIRTY;
2373 if (++n >= MDB_COMMIT_PAGES) {
2383 rc = writev(env->me_fd, iov, n);
2387 DPUTS("short write, filesystem full?");
2389 DPRINTF("writev: %s", strerror(n));
2396 /* Drop the dirty pages.
2398 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2399 dp = txn->mt_u.dirty_list[i].mptr;
2400 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2401 mdb_page_free(txn->mt_env, dp);
2403 VGMEMP_FREE(txn->mt_env, dp);
2406 txn->mt_u.dirty_list[i].mid = 0;
2408 txn->mt_u.dirty_list[0].mid = 0;
2411 if ((n = mdb_env_sync(env, 0)) != 0 ||
2412 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2419 if (txn->mt_numdbs > env->me_numdbs) {
2420 /* update the DB flags */
2422 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2423 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2427 UNLOCK_MUTEX_W(env);
2433 /** Read the environment parameters of a DB environment before
2434 * mapping it into memory.
2435 * @param[in] env the environment handle
2436 * @param[out] meta address of where to store the meta information
2437 * @return 0 on success, non-zero on failure.
2440 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2447 /* We don't know the page size yet, so use a minimum value.
2448 * Read both meta pages so we can use the latest one.
2451 for (i=0; i<2; i++) {
2453 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2455 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2460 else if (rc != MDB_PAGESIZE) {
2464 DPRINTF("read: %s", strerror(err));
2468 p = (MDB_page *)&pbuf;
2470 if (!F_ISSET(p->mp_flags, P_META)) {
2471 DPRINTF("page %zu not a meta page", p->mp_pgno);
2476 if (m->mm_magic != MDB_MAGIC) {
2477 DPUTS("meta has invalid magic");
2481 if (m->mm_version != MDB_VERSION) {
2482 DPRINTF("database is version %u, expected version %u",
2483 m->mm_version, MDB_VERSION);
2484 return MDB_VERSION_MISMATCH;
2488 if (m->mm_txnid > meta->mm_txnid)
2489 memcpy(meta, m, sizeof(*m));
2491 memcpy(meta, m, sizeof(*m));
2493 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2495 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2503 /** Write the environment parameters of a freshly created DB environment.
2504 * @param[in] env the environment handle
2505 * @param[out] meta address of where to store the meta information
2506 * @return 0 on success, non-zero on failure.
2509 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2516 DPUTS("writing new meta page");
2518 GET_PAGESIZE(psize);
2520 meta->mm_magic = MDB_MAGIC;
2521 meta->mm_version = MDB_VERSION;
2522 meta->mm_mapsize = env->me_mapsize;
2523 meta->mm_psize = psize;
2524 meta->mm_last_pg = 1;
2525 meta->mm_flags = env->me_flags & 0xffff;
2526 meta->mm_flags |= MDB_INTEGERKEY;
2527 meta->mm_dbs[0].md_root = P_INVALID;
2528 meta->mm_dbs[1].md_root = P_INVALID;
2530 p = calloc(2, psize);
2532 p->mp_flags = P_META;
2535 memcpy(m, meta, sizeof(*meta));
2537 q = (MDB_page *)((char *)p + psize);
2540 q->mp_flags = P_META;
2543 memcpy(m, meta, sizeof(*meta));
2548 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2549 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2550 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2553 lseek(env->me_fd, 0, SEEK_SET);
2554 rc = write(env->me_fd, p, psize * 2);
2555 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2561 /** Update the environment info to commit a transaction.
2562 * @param[in] txn the transaction that's being committed
2563 * @return 0 on success, non-zero on failure.
2566 mdb_env_write_meta(MDB_txn *txn)
2569 MDB_meta meta, metab, *mp;
2571 int rc, len, toggle;
2578 assert(txn != NULL);
2579 assert(txn->mt_env != NULL);
2581 toggle = !txn->mt_toggle;
2582 DPRINTF("writing meta page %d for root page %zu",
2583 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2586 mp = env->me_metas[toggle];
2588 if (env->me_flags & MDB_WRITEMAP) {
2589 /* Persist any increases of mapsize config */
2590 if (env->me_mapsize > mp->mm_mapsize)
2591 mp->mm_mapsize = env->me_mapsize;
2592 mp->mm_dbs[0] = txn->mt_dbs[0];
2593 mp->mm_dbs[1] = txn->mt_dbs[1];
2594 mp->mm_last_pg = txn->mt_next_pgno - 1;
2595 mp->mm_txnid = txn->mt_txnid;
2596 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2597 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2600 ptr += env->me_psize;
2601 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2608 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2609 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2611 ptr = (char *)&meta;
2612 if (env->me_mapsize > mp->mm_mapsize) {
2613 /* Persist any increases of mapsize config */
2614 meta.mm_mapsize = env->me_mapsize;
2615 off = offsetof(MDB_meta, mm_mapsize);
2617 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2619 len = sizeof(MDB_meta) - off;
2622 meta.mm_dbs[0] = txn->mt_dbs[0];
2623 meta.mm_dbs[1] = txn->mt_dbs[1];
2624 meta.mm_last_pg = txn->mt_next_pgno - 1;
2625 meta.mm_txnid = txn->mt_txnid;
2628 off += env->me_psize;
2631 /* Write to the SYNC fd */
2632 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2633 env->me_fd : env->me_mfd;
2636 memset(&ov, 0, sizeof(ov));
2638 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2641 rc = pwrite(mfd, ptr, len, off);
2646 DPUTS("write failed, disk error?");
2647 /* On a failure, the pagecache still contains the new data.
2648 * Write some old data back, to prevent it from being used.
2649 * Use the non-SYNC fd; we know it will fail anyway.
2651 meta.mm_last_pg = metab.mm_last_pg;
2652 meta.mm_txnid = metab.mm_txnid;
2654 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2656 r2 = pwrite(env->me_fd, ptr, len, off);
2659 env->me_flags |= MDB_FATAL_ERROR;
2663 /* Memory ordering issues are irrelevant; since the entire writer
2664 * is wrapped by wmutex, all of these changes will become visible
2665 * after the wmutex is unlocked. Since the DB is multi-version,
2666 * readers will get consistent data regardless of how fresh or
2667 * how stale their view of these values is.
2669 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2674 /** Check both meta pages to see which one is newer.
2675 * @param[in] env the environment handle
2676 * @return meta toggle (0 or 1).
2679 mdb_env_pick_meta(const MDB_env *env)
2681 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2685 mdb_env_create(MDB_env **env)
2689 e = calloc(1, sizeof(MDB_env));
2693 e->me_free_pgs = mdb_midl_alloc();
2694 if (!e->me_free_pgs) {
2698 e->me_maxreaders = DEFAULT_READERS;
2700 e->me_fd = INVALID_HANDLE_VALUE;
2701 e->me_lfd = INVALID_HANDLE_VALUE;
2702 e->me_mfd = INVALID_HANDLE_VALUE;
2703 #ifdef MDB_USE_POSIX_SEM
2704 e->me_rmutex = SEM_FAILED;
2705 e->me_wmutex = SEM_FAILED;
2707 e->me_pid = getpid();
2708 VGMEMP_CREATE(e,0,0);
2714 mdb_env_set_mapsize(MDB_env *env, size_t size)
2718 env->me_mapsize = size;
2720 env->me_maxpg = env->me_mapsize / env->me_psize;
2725 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2729 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2734 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2736 if (env->me_map || readers < 1)
2738 env->me_maxreaders = readers;
2743 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2745 if (!env || !readers)
2747 *readers = env->me_maxreaders;
2751 /** Further setup required for opening an MDB environment
2754 mdb_env_open2(MDB_env *env)
2756 unsigned int flags = env->me_flags;
2757 int i, newenv = 0, prot;
2761 memset(&meta, 0, sizeof(meta));
2763 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2766 DPUTS("new mdbenv");
2770 /* Was a mapsize configured? */
2771 if (!env->me_mapsize) {
2772 /* If this is a new environment, take the default,
2773 * else use the size recorded in the existing env.
2775 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2776 } else if (env->me_mapsize < meta.mm_mapsize) {
2777 /* If the configured size is smaller, make sure it's
2778 * still big enough. Silently round up to minimum if not.
2780 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2781 if (env->me_mapsize < minsize)
2782 env->me_mapsize = minsize;
2788 LONG sizelo, sizehi;
2789 sizelo = env->me_mapsize & 0xffffffff;
2790 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2792 /* Windows won't create mappings for zero length files.
2793 * Just allocate the maxsize right now.
2796 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2797 if (!SetEndOfFile(env->me_fd))
2799 SetFilePointer(env->me_fd, 0, NULL, 0);
2801 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2802 PAGE_READWRITE : PAGE_READONLY,
2803 sizehi, sizelo, NULL);
2806 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2807 FILE_MAP_WRITE : FILE_MAP_READ,
2808 0, 0, env->me_mapsize, meta.mm_address);
2816 if (flags & MDB_WRITEMAP) {
2818 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2821 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2823 if (env->me_map == MAP_FAILED) {
2830 if (flags & MDB_FIXEDMAP)
2831 meta.mm_address = env->me_map;
2832 i = mdb_env_init_meta(env, &meta);
2833 if (i != MDB_SUCCESS) {
2836 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2837 /* Can happen because the address argument to mmap() is just a
2838 * hint. mmap() can pick another, e.g. if the range is in use.
2839 * The MAP_FIXED flag would prevent that, but then mmap could
2840 * instead unmap existing pages to make room for the new map.
2842 return EBUSY; /* TODO: Make a new MDB_* error code? */
2844 env->me_psize = meta.mm_psize;
2846 env->me_maxpg = env->me_mapsize / env->me_psize;
2848 p = (MDB_page *)env->me_map;
2849 env->me_metas[0] = METADATA(p);
2850 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2854 int toggle = mdb_env_pick_meta(env);
2855 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2857 DPRINTF("opened database version %u, pagesize %u",
2858 env->me_metas[0]->mm_version, env->me_psize);
2859 DPRINTF("using meta page %d", toggle);
2860 DPRINTF("depth: %u", db->md_depth);
2861 DPRINTF("entries: %zu", db->md_entries);
2862 DPRINTF("branch pages: %zu", db->md_branch_pages);
2863 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2864 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2865 DPRINTF("root: %zu", db->md_root);
2873 /** Release a reader thread's slot in the reader lock table.
2874 * This function is called automatically when a thread exits.
2875 * @param[in] ptr This points to the slot in the reader lock table.
2878 mdb_env_reader_dest(void *ptr)
2880 MDB_reader *reader = ptr;
2886 /** Junk for arranging thread-specific callbacks on Windows. This is
2887 * necessarily platform and compiler-specific. Windows supports up
2888 * to 1088 keys. Let's assume nobody opens more than 64 environments
2889 * in a single process, for now. They can override this if needed.
2891 #ifndef MAX_TLS_KEYS
2892 #define MAX_TLS_KEYS 64
2894 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2895 static int mdb_tls_nkeys;
2897 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2901 case DLL_PROCESS_ATTACH: break;
2902 case DLL_THREAD_ATTACH: break;
2903 case DLL_THREAD_DETACH:
2904 for (i=0; i<mdb_tls_nkeys; i++) {
2905 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2906 mdb_env_reader_dest(r);
2909 case DLL_PROCESS_DETACH: break;
2914 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2916 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2920 /* Force some symbol references.
2921 * _tls_used forces the linker to create the TLS directory if not already done
2922 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2924 #pragma comment(linker, "/INCLUDE:_tls_used")
2925 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2926 #pragma const_seg(".CRT$XLB")
2927 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2928 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2931 #pragma comment(linker, "/INCLUDE:__tls_used")
2932 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2933 #pragma data_seg(".CRT$XLB")
2934 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2936 #endif /* WIN 32/64 */
2937 #endif /* !__GNUC__ */
2940 /** Downgrade the exclusive lock on the region back to shared */
2942 mdb_env_share_locks(MDB_env *env, int *excl)
2944 int rc = 0, toggle = mdb_env_pick_meta(env);
2946 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2951 /* First acquire a shared lock. The Unlock will
2952 * then release the existing exclusive lock.
2954 memset(&ov, 0, sizeof(ov));
2955 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2958 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2964 struct flock lock_info;
2965 /* The shared lock replaces the existing lock */
2966 memset((void *)&lock_info, 0, sizeof(lock_info));
2967 lock_info.l_type = F_RDLCK;
2968 lock_info.l_whence = SEEK_SET;
2969 lock_info.l_start = 0;
2970 lock_info.l_len = 1;
2971 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2972 (rc = ErrCode()) == EINTR) ;
2973 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
2980 /** Try to get exlusive lock, otherwise shared.
2981 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
2984 mdb_env_excl_lock(MDB_env *env, int *excl)
2988 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2992 memset(&ov, 0, sizeof(ov));
2993 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3000 struct flock lock_info;
3001 memset((void *)&lock_info, 0, sizeof(lock_info));
3002 lock_info.l_type = F_WRLCK;
3003 lock_info.l_whence = SEEK_SET;
3004 lock_info.l_start = 0;
3005 lock_info.l_len = 1;
3006 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3007 (rc = ErrCode()) == EINTR) ;
3011 # ifdef MDB_USE_POSIX_SEM
3012 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3015 lock_info.l_type = F_RDLCK;
3016 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3017 (rc = ErrCode()) == EINTR) ;
3025 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3027 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3029 * @(#) $Revision: 5.1 $
3030 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3031 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3033 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3037 * Please do not copyright this code. This code is in the public domain.
3039 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3040 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3041 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3042 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3043 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3044 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3045 * PERFORMANCE OF THIS SOFTWARE.
3048 * chongo <Landon Curt Noll> /\oo/\
3049 * http://www.isthe.com/chongo/
3051 * Share and Enjoy! :-)
3054 typedef unsigned long long mdb_hash_t;
3055 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3057 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3058 * @param[in] str string to hash
3059 * @param[in] hval initial value for hash
3060 * @return 64 bit hash
3062 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3063 * hval arg on the first call.
3066 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3068 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3069 unsigned char *end = s + val->mv_size;
3071 * FNV-1a hash each octet of the string
3074 /* xor the bottom with the current octet */
3075 hval ^= (mdb_hash_t)*s++;
3077 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3078 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3079 (hval << 7) + (hval << 8) + (hval << 40);
3081 /* return our new hash value */
3085 /** Hash the string and output the hash in hex.
3086 * @param[in] str string to hash
3087 * @param[out] hexbuf an array of 17 chars to hold the hash
3090 mdb_hash_hex(MDB_val *val, char *hexbuf)
3093 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3094 for (i=0; i<8; i++) {
3095 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3101 /** Open and/or initialize the lock region for the environment.
3102 * @param[in] env The MDB environment.
3103 * @param[in] lpath The pathname of the file used for the lock region.
3104 * @param[in] mode The Unix permissions for the file, if we create it.
3105 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3106 * @return 0 on success, non-zero on failure.
3109 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3117 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3118 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3119 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
3121 if (rc == ERROR_WRITE_PROTECT && (env->me_flags & MDB_RDONLY)) {
3122 env->me_flags |= MDB_ROFS;
3127 /* Try to get exclusive lock. If we succeed, then
3128 * nobody is using the lock region and we should initialize it.
3130 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3131 size = GetFileSize(env->me_lfd, NULL);
3137 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
3139 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3140 env->me_flags |= MDB_ROFS;
3145 /* Lose record locks when exec*() */
3146 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3147 fcntl(env->me_lfd, F_SETFD, fdflags);
3149 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
3150 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1) {
3152 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3153 env->me_flags |= MDB_ROFS;
3160 /* Try to get exclusive lock. If we succeed, then
3161 * nobody is using the lock region and we should initialize it.
3163 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3165 size = lseek(env->me_lfd, 0, SEEK_END);
3167 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3168 if (size < rsize && *excl > 0) {
3170 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3171 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3173 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3177 size = rsize - sizeof(MDB_txninfo);
3178 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3183 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3185 if (!mh) goto fail_errno;
3186 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3188 if (!env->me_txns) goto fail_errno;
3190 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3192 if (m == MAP_FAILED) goto fail_errno;
3198 BY_HANDLE_FILE_INFORMATION stbuf;
3207 if (!mdb_sec_inited) {
3208 InitializeSecurityDescriptor(&mdb_null_sd,
3209 SECURITY_DESCRIPTOR_REVISION);
3210 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3211 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3212 mdb_all_sa.bInheritHandle = FALSE;
3213 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3216 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3217 idbuf.volume = stbuf.dwVolumeSerialNumber;
3218 idbuf.nhigh = stbuf.nFileIndexHigh;
3219 idbuf.nlow = stbuf.nFileIndexLow;
3220 val.mv_data = &idbuf;
3221 val.mv_size = sizeof(idbuf);
3222 mdb_hash_hex(&val, hexbuf);
3223 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3224 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3225 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3226 if (!env->me_rmutex) goto fail_errno;
3227 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3228 if (!env->me_wmutex) goto fail_errno;
3229 #elif defined(MDB_USE_POSIX_SEM)
3238 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3239 idbuf.dev = stbuf.st_dev;
3240 idbuf.ino = stbuf.st_ino;
3241 val.mv_data = &idbuf;
3242 val.mv_size = sizeof(idbuf);
3243 mdb_hash_hex(&val, hexbuf);
3244 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3245 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3246 /* Clean up after a previous run, if needed: Try to
3247 * remove both semaphores before doing anything else.
3249 sem_unlink(env->me_txns->mti_rmname);
3250 sem_unlink(env->me_txns->mti_wmname);
3251 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3252 O_CREAT|O_EXCL, mode, 1);
3253 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3254 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3255 O_CREAT|O_EXCL, mode, 1);
3256 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3257 #else /* MDB_USE_POSIX_SEM */
3258 pthread_mutexattr_t mattr;
3260 if ((rc = pthread_mutexattr_init(&mattr))
3261 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3262 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3263 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3265 pthread_mutexattr_destroy(&mattr);
3266 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3268 env->me_txns->mti_version = MDB_VERSION;
3269 env->me_txns->mti_magic = MDB_MAGIC;
3270 env->me_txns->mti_txnid = 0;
3271 env->me_txns->mti_numreaders = 0;
3274 if (env->me_txns->mti_magic != MDB_MAGIC) {
3275 DPUTS("lock region has invalid magic");
3279 if (env->me_txns->mti_version != MDB_VERSION) {
3280 DPRINTF("lock region is version %u, expected version %u",
3281 env->me_txns->mti_version, MDB_VERSION);
3282 rc = MDB_VERSION_MISMATCH;
3286 if (rc != EACCES && rc != EAGAIN) {
3290 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3291 if (!env->me_rmutex) goto fail_errno;
3292 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3293 if (!env->me_wmutex) goto fail_errno;
3294 #elif defined(MDB_USE_POSIX_SEM)
3295 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3296 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3297 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3298 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3309 /** The name of the lock file in the DB environment */
3310 #define LOCKNAME "/lock.mdb"
3311 /** The name of the data file in the DB environment */
3312 #define DATANAME "/data.mdb"
3313 /** The suffix of the lock file when no subdir is used */
3314 #define LOCKSUFF "-lock"
3315 /** Only a subset of the @ref mdb_env flags can be changed
3316 * at runtime. Changing other flags requires closing the
3317 * environment and re-opening it with the new flags.
3319 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3320 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP)
3323 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3325 int oflags, rc, len, excl;
3326 char *lpath, *dpath;
3328 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3332 if (flags & MDB_NOSUBDIR) {
3333 rc = len + sizeof(LOCKSUFF) + len + 1;
3335 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3340 if (flags & MDB_NOSUBDIR) {
3341 dpath = lpath + len + sizeof(LOCKSUFF);
3342 sprintf(lpath, "%s" LOCKSUFF, path);
3343 strcpy(dpath, path);
3345 dpath = lpath + len + sizeof(LOCKNAME);
3346 sprintf(lpath, "%s" LOCKNAME, path);
3347 sprintf(dpath, "%s" DATANAME, path);
3350 flags |= env->me_flags;
3351 /* silently ignore WRITEMAP if we're only getting read access */
3352 if (F_ISSET(flags, MDB_RDONLY|MDB_WRITEMAP))
3353 flags ^= MDB_WRITEMAP;
3354 env->me_flags = flags |= MDB_ENV_ACTIVE;
3356 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3361 if (F_ISSET(flags, MDB_RDONLY)) {
3362 oflags = GENERIC_READ;
3363 len = OPEN_EXISTING;
3365 oflags = GENERIC_READ|GENERIC_WRITE;
3368 mode = FILE_ATTRIBUTE_NORMAL;
3369 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3370 NULL, len, mode, NULL);
3372 if (F_ISSET(flags, MDB_RDONLY))
3375 oflags = O_RDWR | O_CREAT;
3377 env->me_fd = open(dpath, oflags, mode);
3379 if (env->me_fd == INVALID_HANDLE_VALUE) {
3384 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3385 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3386 env->me_mfd = env->me_fd;
3388 /* Synchronous fd for meta writes. Needed even with
3389 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3392 env->me_mfd = CreateFile(dpath, oflags,
3393 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3394 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3396 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3398 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3403 DPRINTF("opened dbenv %p", (void *) env);
3404 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3407 env->me_numdbs = 2; /* this notes that me_txkey was set */
3409 /* Windows TLS callbacks need help finding their TLS info. */
3410 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3411 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3418 rc = mdb_env_share_locks(env, &excl);
3422 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3423 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3424 env->me_path = strdup(path);
3425 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3431 mdb_env_close0(env, excl);
3437 /** Destroy resources from mdb_env_open() and clear our readers */
3439 mdb_env_close0(MDB_env *env, int excl)
3443 if (!(env->me_flags & MDB_ENV_ACTIVE))
3446 free(env->me_dbflags);
3450 if (env->me_numdbs) {
3451 pthread_key_delete(env->me_txkey);
3453 /* Delete our key from the global list */
3454 for (i=0; i<mdb_tls_nkeys; i++)
3455 if (mdb_tls_keys[i] == env->me_txkey) {
3456 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3464 munmap(env->me_map, env->me_mapsize);
3466 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3468 if (env->me_fd != INVALID_HANDLE_VALUE)
3471 pid_t pid = env->me_pid;
3472 /* Clearing readers is done in this function because
3473 * me_txkey with its destructor must be disabled first.
3475 for (i = env->me_numreaders; --i >= 0; )
3476 if (env->me_txns->mti_readers[i].mr_pid == pid)
3477 env->me_txns->mti_readers[i].mr_pid = 0;
3479 if (env->me_rmutex) {
3480 CloseHandle(env->me_rmutex);
3481 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3483 /* Windows automatically destroys the mutexes when
3484 * the last handle closes.
3486 #elif defined(MDB_USE_POSIX_SEM)
3487 if (env->me_rmutex != SEM_FAILED) {
3488 sem_close(env->me_rmutex);
3489 if (env->me_wmutex != SEM_FAILED)
3490 sem_close(env->me_wmutex);
3491 /* If we have the filelock: If we are the
3492 * only remaining user, clean up semaphores.
3495 mdb_env_excl_lock(env, &excl);
3497 sem_unlink(env->me_txns->mti_rmname);
3498 sem_unlink(env->me_txns->mti_wmname);
3502 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3504 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3507 /* Unlock the lockfile. Windows would have unlocked it
3508 * after closing anyway, but not necessarily at once.
3510 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3516 env->me_flags &= ~MDB_ENV_ACTIVE;
3520 mdb_env_copy(MDB_env *env, const char *path)
3522 MDB_txn *txn = NULL;
3526 HANDLE newfd = INVALID_HANDLE_VALUE;
3528 if (env->me_flags & MDB_NOSUBDIR) {
3529 lpath = (char *)path;
3532 len += sizeof(DATANAME);
3533 lpath = malloc(len);
3536 sprintf(lpath, "%s" DATANAME, path);
3539 /* The destination path must exist, but the destination file must not.
3540 * We don't want the OS to cache the writes, since the source data is
3541 * already in the OS cache.
3544 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3545 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3547 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3553 if (!(env->me_flags & MDB_NOSUBDIR))
3555 if (newfd == INVALID_HANDLE_VALUE) {
3560 #ifdef F_NOCACHE /* __APPLE__ */
3561 rc = fcntl(newfd, F_NOCACHE, 1);
3568 /* Do the lock/unlock of the reader mutex before starting the
3569 * write txn. Otherwise other read txns could block writers.
3571 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3575 if (!(env->me_flags & MDB_ROFS)) {
3576 /* We must start the actual read txn after blocking writers */
3577 mdb_txn_reset0(txn);
3579 /* Temporarily block writers until we snapshot the meta pages */
3582 rc = mdb_txn_renew0(txn);
3584 UNLOCK_MUTEX_W(env);
3589 wsize = env->me_psize * 2;
3593 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3594 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3597 rc = write(newfd, env->me_map, wsize);
3598 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3600 if (! (env->me_flags & MDB_ROFS))
3601 UNLOCK_MUTEX_W(env);
3606 ptr = env->me_map + wsize;
3607 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3608 #define MAX_WRITE 2147483648U
3612 if (wsize > MAX_WRITE)
3616 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3617 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3626 if (wsize > MAX_WRITE)
3630 wres = write(newfd, ptr, w2);
3631 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3640 if (newfd != INVALID_HANDLE_VALUE)
3647 mdb_env_close(MDB_env *env)
3654 VGMEMP_DESTROY(env);
3655 while ((dp = env->me_dpages) != NULL) {
3656 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3657 env->me_dpages = dp->mp_next;
3661 mdb_env_close0(env, 0);
3662 mdb_midl_free(env->me_free_pgs);
3666 /** Compare two items pointing at aligned size_t's */
3668 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3670 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3671 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3674 /** Compare two items pointing at aligned int's */
3676 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3678 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3679 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3682 /** Compare two items pointing at ints of unknown alignment.
3683 * Nodes and keys are guaranteed to be 2-byte aligned.
3686 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3688 #if BYTE_ORDER == LITTLE_ENDIAN
3689 unsigned short *u, *c;
3692 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3693 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3696 } while(!x && u > (unsigned short *)a->mv_data);
3699 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3703 /** Compare two items lexically */
3705 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3712 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3718 diff = memcmp(a->mv_data, b->mv_data, len);
3719 return diff ? diff : len_diff<0 ? -1 : len_diff;
3722 /** Compare two items in reverse byte order */
3724 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3726 const unsigned char *p1, *p2, *p1_lim;
3730 p1_lim = (const unsigned char *)a->mv_data;
3731 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3732 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3734 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3740 while (p1 > p1_lim) {
3741 diff = *--p1 - *--p2;
3745 return len_diff<0 ? -1 : len_diff;
3748 /** Search for key within a page, using binary search.
3749 * Returns the smallest entry larger or equal to the key.
3750 * If exactp is non-null, stores whether the found entry was an exact match
3751 * in *exactp (1 or 0).
3752 * Updates the cursor index with the index of the found entry.
3753 * If no entry larger or equal to the key is found, returns NULL.
3756 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3758 unsigned int i = 0, nkeys;
3761 MDB_page *mp = mc->mc_pg[mc->mc_top];
3762 MDB_node *node = NULL;
3767 nkeys = NUMKEYS(mp);
3772 COPY_PGNO(pgno, mp->mp_pgno);
3773 DPRINTF("searching %u keys in %s %spage %zu",
3774 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3781 low = IS_LEAF(mp) ? 0 : 1;
3783 cmp = mc->mc_dbx->md_cmp;
3785 /* Branch pages have no data, so if using integer keys,
3786 * alignment is guaranteed. Use faster mdb_cmp_int.
3788 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3789 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3796 nodekey.mv_size = mc->mc_db->md_pad;
3797 node = NODEPTR(mp, 0); /* fake */
3798 while (low <= high) {
3799 i = (low + high) >> 1;
3800 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3801 rc = cmp(key, &nodekey);
3802 DPRINTF("found leaf index %u [%s], rc = %i",
3803 i, DKEY(&nodekey), rc);
3812 while (low <= high) {
3813 i = (low + high) >> 1;
3815 node = NODEPTR(mp, i);
3816 nodekey.mv_size = NODEKSZ(node);
3817 nodekey.mv_data = NODEKEY(node);
3819 rc = cmp(key, &nodekey);
3822 DPRINTF("found leaf index %u [%s], rc = %i",
3823 i, DKEY(&nodekey), rc);
3825 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3826 i, DKEY(&nodekey), NODEPGNO(node), rc);
3837 if (rc > 0) { /* Found entry is less than the key. */
3838 i++; /* Skip to get the smallest entry larger than key. */
3840 node = NODEPTR(mp, i);
3843 *exactp = (rc == 0);
3844 /* store the key index */
3845 mc->mc_ki[mc->mc_top] = i;
3847 /* There is no entry larger or equal to the key. */
3850 /* nodeptr is fake for LEAF2 */
3856 mdb_cursor_adjust(MDB_cursor *mc, func)
3860 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3861 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3868 /** Pop a page off the top of the cursor's stack. */
3870 mdb_cursor_pop(MDB_cursor *mc)
3873 #ifndef MDB_DEBUG_SKIP
3874 MDB_page *top = mc->mc_pg[mc->mc_top];
3880 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3881 mc->mc_dbi, (void *) mc);
3885 /** Push a page onto the top of the cursor's stack. */
3887 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3889 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3890 mc->mc_dbi, (void *) mc);
3892 if (mc->mc_snum >= CURSOR_STACK) {
3893 assert(mc->mc_snum < CURSOR_STACK);
3894 return MDB_CURSOR_FULL;
3897 mc->mc_top = mc->mc_snum++;
3898 mc->mc_pg[mc->mc_top] = mp;
3899 mc->mc_ki[mc->mc_top] = 0;
3904 /** Find the address of the page corresponding to a given page number.
3905 * @param[in] txn the transaction for this access.
3906 * @param[in] pgno the page number for the page to retrieve.
3907 * @param[out] ret address of a pointer where the page's address will be stored.
3908 * @return 0 on success, non-zero on failure.
3911 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3915 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
3916 if (pgno < txn->mt_next_pgno)
3917 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3920 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3922 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3923 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3924 p = txn->mt_u.dirty_list[x].mptr;
3928 if (pgno < txn->mt_next_pgno)
3929 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3934 DPRINTF("page %zu not found", pgno);
3937 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3940 /** Search for the page a given key should be in.
3941 * Pushes parent pages on the cursor stack. This function continues a
3942 * search on a cursor that has already been initialized. (Usually by
3943 * #mdb_page_search() but also by #mdb_node_move().)
3944 * @param[in,out] mc the cursor for this operation.
3945 * @param[in] key the key to search for. If NULL, search for the lowest
3946 * page. (This is used by #mdb_cursor_first().)
3947 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3948 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3949 * @return 0 on success, non-zero on failure.
3952 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3954 MDB_page *mp = mc->mc_pg[mc->mc_top];
3959 while (IS_BRANCH(mp)) {
3963 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3964 assert(NUMKEYS(mp) > 1);
3965 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3967 if (key == NULL) /* Initialize cursor to first page. */
3969 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
3970 /* cursor to last page */
3974 node = mdb_node_search(mc, key, &exact);
3976 i = NUMKEYS(mp) - 1;
3978 i = mc->mc_ki[mc->mc_top];
3987 DPRINTF("following index %u for key [%s]",
3989 assert(i < NUMKEYS(mp));
3990 node = NODEPTR(mp, i);
3992 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3995 mc->mc_ki[mc->mc_top] = i;
3996 if ((rc = mdb_cursor_push(mc, mp)))
4000 if ((rc = mdb_page_touch(mc)) != 0)
4002 mp = mc->mc_pg[mc->mc_top];
4007 DPRINTF("internal error, index points to a %02X page!?",
4009 return MDB_CORRUPTED;
4012 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4013 key ? DKEY(key) : NULL);
4018 /** Search for the page a given key should be in.
4019 * Pushes parent pages on the cursor stack. This function just sets up
4020 * the search; it finds the root page for \b mc's database and sets this
4021 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4022 * called to complete the search.
4023 * @param[in,out] mc the cursor for this operation.
4024 * @param[in] key the key to search for. If NULL, search for the lowest
4025 * page. (This is used by #mdb_cursor_first().)
4026 * @param[in] modify If true, visited pages are updated with new page numbers.
4027 * @return 0 on success, non-zero on failure.
4030 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4035 /* Make sure the txn is still viable, then find the root from
4036 * the txn's db table.
4038 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4039 DPUTS("transaction has failed, must abort");
4042 /* Make sure we're using an up-to-date root */
4043 if (mc->mc_dbi > MAIN_DBI) {
4044 if ((*mc->mc_dbflag & DB_STALE) ||
4045 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4047 unsigned char dbflag = 0;
4048 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4049 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4052 if (*mc->mc_dbflag & DB_STALE) {
4055 MDB_node *leaf = mdb_node_search(&mc2,
4056 &mc->mc_dbx->md_name, &exact);
4058 return MDB_NOTFOUND;
4059 mdb_node_read(mc->mc_txn, leaf, &data);
4060 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4062 if (flags & MDB_PS_MODIFY)
4064 *mc->mc_dbflag = dbflag;
4067 root = mc->mc_db->md_root;
4069 if (root == P_INVALID) { /* Tree is empty. */
4070 DPUTS("tree is empty");
4071 return MDB_NOTFOUND;
4076 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4077 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
4083 DPRINTF("db %u root page %zu has flags 0x%X",
4084 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4086 if (flags & MDB_PS_MODIFY) {
4087 if ((rc = mdb_page_touch(mc)))
4091 if (flags & MDB_PS_ROOTONLY)
4094 return mdb_page_search_root(mc, key, flags);
4097 /** Return the data associated with a given node.
4098 * @param[in] txn The transaction for this operation.
4099 * @param[in] leaf The node being read.
4100 * @param[out] data Updated to point to the node's data.
4101 * @return 0 on success, non-zero on failure.
4104 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4106 MDB_page *omp; /* overflow page */
4110 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4111 data->mv_size = NODEDSZ(leaf);
4112 data->mv_data = NODEDATA(leaf);
4116 /* Read overflow data.
4118 data->mv_size = NODEDSZ(leaf);
4119 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4120 if ((rc = mdb_page_get(txn, pgno, &omp))) {
4121 DPRINTF("read overflow page %zu failed", pgno);
4124 data->mv_data = METADATA(omp);
4130 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4131 MDB_val *key, MDB_val *data)
4140 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4142 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4145 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4149 mdb_cursor_init(&mc, txn, dbi, &mx);
4150 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4153 /** Find a sibling for a page.
4154 * Replaces the page at the top of the cursor's stack with the
4155 * specified sibling, if one exists.
4156 * @param[in] mc The cursor for this operation.
4157 * @param[in] move_right Non-zero if the right sibling is requested,
4158 * otherwise the left sibling.
4159 * @return 0 on success, non-zero on failure.
4162 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4168 if (mc->mc_snum < 2) {
4169 return MDB_NOTFOUND; /* root has no siblings */
4173 DPRINTF("parent page is page %zu, index %u",
4174 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4176 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4177 : (mc->mc_ki[mc->mc_top] == 0)) {
4178 DPRINTF("no more keys left, moving to %s sibling",
4179 move_right ? "right" : "left");
4180 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4181 /* undo cursor_pop before returning */
4188 mc->mc_ki[mc->mc_top]++;
4190 mc->mc_ki[mc->mc_top]--;
4191 DPRINTF("just moving to %s index key %u",
4192 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4194 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4196 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4197 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4200 mdb_cursor_push(mc, mp);
4205 /** Move the cursor to the next data item. */
4207 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4213 if (mc->mc_flags & C_EOF) {
4214 return MDB_NOTFOUND;
4217 assert(mc->mc_flags & C_INITIALIZED);
4219 mp = mc->mc_pg[mc->mc_top];
4221 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4222 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4223 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4224 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4225 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4226 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4230 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4231 if (op == MDB_NEXT_DUP)
4232 return MDB_NOTFOUND;
4236 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4238 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4239 DPUTS("=====> move to next sibling page");
4240 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4241 mc->mc_flags |= C_EOF;
4242 mc->mc_flags &= ~C_INITIALIZED;
4243 return MDB_NOTFOUND;
4245 mp = mc->mc_pg[mc->mc_top];
4246 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4248 mc->mc_ki[mc->mc_top]++;
4250 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4251 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4254 key->mv_size = mc->mc_db->md_pad;
4255 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4259 assert(IS_LEAF(mp));
4260 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4262 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4263 mdb_xcursor_init1(mc, leaf);
4266 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4269 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4270 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4271 if (rc != MDB_SUCCESS)
4276 MDB_GET_KEY(leaf, key);
4280 /** Move the cursor to the previous data item. */
4282 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4288 assert(mc->mc_flags & C_INITIALIZED);
4290 mp = mc->mc_pg[mc->mc_top];
4292 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4293 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4294 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4295 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4296 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4297 if (op != MDB_PREV || rc == MDB_SUCCESS)
4300 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4301 if (op == MDB_PREV_DUP)
4302 return MDB_NOTFOUND;
4307 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4309 if (mc->mc_ki[mc->mc_top] == 0) {
4310 DPUTS("=====> move to prev sibling page");
4311 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4312 mc->mc_flags &= ~C_INITIALIZED;
4313 return MDB_NOTFOUND;
4315 mp = mc->mc_pg[mc->mc_top];
4316 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4317 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4319 mc->mc_ki[mc->mc_top]--;
4321 mc->mc_flags &= ~C_EOF;
4323 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4324 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4327 key->mv_size = mc->mc_db->md_pad;
4328 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4332 assert(IS_LEAF(mp));
4333 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4335 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4336 mdb_xcursor_init1(mc, leaf);
4339 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4342 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4343 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4344 if (rc != MDB_SUCCESS)
4349 MDB_GET_KEY(leaf, key);
4353 /** Set the cursor on a specific data item. */
4355 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4356 MDB_cursor_op op, int *exactp)
4360 MDB_node *leaf = NULL;
4365 assert(key->mv_size > 0);
4367 /* See if we're already on the right page */
4368 if (mc->mc_flags & C_INITIALIZED) {
4371 mp = mc->mc_pg[mc->mc_top];
4373 mc->mc_ki[mc->mc_top] = 0;
4374 return MDB_NOTFOUND;
4376 if (mp->mp_flags & P_LEAF2) {
4377 nodekey.mv_size = mc->mc_db->md_pad;
4378 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4380 leaf = NODEPTR(mp, 0);
4381 MDB_GET_KEY(leaf, &nodekey);
4383 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4385 /* Probably happens rarely, but first node on the page
4386 * was the one we wanted.
4388 mc->mc_ki[mc->mc_top] = 0;
4395 unsigned int nkeys = NUMKEYS(mp);
4397 if (mp->mp_flags & P_LEAF2) {
4398 nodekey.mv_data = LEAF2KEY(mp,
4399 nkeys-1, nodekey.mv_size);
4401 leaf = NODEPTR(mp, nkeys-1);
4402 MDB_GET_KEY(leaf, &nodekey);
4404 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4406 /* last node was the one we wanted */
4407 mc->mc_ki[mc->mc_top] = nkeys-1;
4413 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4414 /* This is definitely the right page, skip search_page */
4415 if (mp->mp_flags & P_LEAF2) {
4416 nodekey.mv_data = LEAF2KEY(mp,
4417 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4419 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4420 MDB_GET_KEY(leaf, &nodekey);
4422 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4424 /* current node was the one we wanted */
4434 /* If any parents have right-sibs, search.
4435 * Otherwise, there's nothing further.
4437 for (i=0; i<mc->mc_top; i++)
4439 NUMKEYS(mc->mc_pg[i])-1)
4441 if (i == mc->mc_top) {
4442 /* There are no other pages */
4443 mc->mc_ki[mc->mc_top] = nkeys;
4444 return MDB_NOTFOUND;
4448 /* There are no other pages */
4449 mc->mc_ki[mc->mc_top] = 0;
4450 return MDB_NOTFOUND;
4454 rc = mdb_page_search(mc, key, 0);
4455 if (rc != MDB_SUCCESS)
4458 mp = mc->mc_pg[mc->mc_top];
4459 assert(IS_LEAF(mp));
4462 leaf = mdb_node_search(mc, key, exactp);
4463 if (exactp != NULL && !*exactp) {
4464 /* MDB_SET specified and not an exact match. */
4465 return MDB_NOTFOUND;
4469 DPUTS("===> inexact leaf not found, goto sibling");
4470 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4471 return rc; /* no entries matched */
4472 mp = mc->mc_pg[mc->mc_top];
4473 assert(IS_LEAF(mp));
4474 leaf = NODEPTR(mp, 0);
4478 mc->mc_flags |= C_INITIALIZED;
4479 mc->mc_flags &= ~C_EOF;
4482 key->mv_size = mc->mc_db->md_pad;
4483 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4487 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4488 mdb_xcursor_init1(mc, leaf);
4491 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4492 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4493 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4496 if (op == MDB_GET_BOTH) {
4502 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4503 if (rc != MDB_SUCCESS)
4506 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4508 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4510 rc = mc->mc_dbx->md_dcmp(data, &d2);
4512 if (op == MDB_GET_BOTH || rc > 0)
4513 return MDB_NOTFOUND;
4518 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4519 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4524 /* The key already matches in all other cases */
4525 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4526 MDB_GET_KEY(leaf, key);
4527 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4532 /** Move the cursor to the first item in the database. */
4534 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4539 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4540 rc = mdb_page_search(mc, NULL, 0);
4541 if (rc != MDB_SUCCESS)
4544 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4546 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4547 mc->mc_flags |= C_INITIALIZED;
4548 mc->mc_flags &= ~C_EOF;
4550 mc->mc_ki[mc->mc_top] = 0;
4552 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4553 key->mv_size = mc->mc_db->md_pad;
4554 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4559 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4560 mdb_xcursor_init1(mc, leaf);
4561 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4566 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4567 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4571 MDB_GET_KEY(leaf, key);
4575 /** Move the cursor to the last item in the database. */
4577 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4582 if (!(mc->mc_flags & C_EOF)) {
4584 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4587 lkey.mv_size = MDB_MAXKEYSIZE+1;
4588 lkey.mv_data = NULL;
4589 rc = mdb_page_search(mc, &lkey, 0);
4590 if (rc != MDB_SUCCESS)
4593 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4595 mc->mc_flags |= C_INITIALIZED|C_EOF;
4596 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4598 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4600 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4601 key->mv_size = mc->mc_db->md_pad;
4602 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4607 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4608 mdb_xcursor_init1(mc, leaf);
4609 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4614 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4615 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4620 MDB_GET_KEY(leaf, key);
4625 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4634 case MDB_GET_CURRENT:
4635 if (!mc->mc_flags & C_INITIALIZED) {
4638 MDB_page *mp = mc->mc_pg[mc->mc_top];
4640 mc->mc_ki[mc->mc_top] = 0;
4646 key->mv_size = mc->mc_db->md_pad;
4647 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4649 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4650 MDB_GET_KEY(leaf, key);
4652 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4653 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4655 rc = mdb_node_read(mc->mc_txn, leaf, data);
4662 case MDB_GET_BOTH_RANGE:
4663 if (data == NULL || mc->mc_xcursor == NULL) {
4671 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4673 } else if (op == MDB_SET_RANGE)
4674 rc = mdb_cursor_set(mc, key, data, op, NULL);
4676 rc = mdb_cursor_set(mc, key, data, op, &exact);
4678 case MDB_GET_MULTIPLE:
4680 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4681 !(mc->mc_flags & C_INITIALIZED)) {
4686 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4687 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4690 case MDB_NEXT_MULTIPLE:
4692 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4696 if (!(mc->mc_flags & C_INITIALIZED))
4697 rc = mdb_cursor_first(mc, key, data);
4699 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4700 if (rc == MDB_SUCCESS) {
4701 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4704 mx = &mc->mc_xcursor->mx_cursor;
4705 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4707 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4708 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4716 case MDB_NEXT_NODUP:
4717 if (!(mc->mc_flags & C_INITIALIZED))
4718 rc = mdb_cursor_first(mc, key, data);
4720 rc = mdb_cursor_next(mc, key, data, op);
4724 case MDB_PREV_NODUP:
4725 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4726 rc = mdb_cursor_last(mc, key, data);
4727 mc->mc_flags |= C_INITIALIZED;
4728 mc->mc_ki[mc->mc_top]++;
4730 rc = mdb_cursor_prev(mc, key, data, op);
4733 rc = mdb_cursor_first(mc, key, data);
4737 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4738 !(mc->mc_flags & C_INITIALIZED) ||
4739 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4743 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4746 rc = mdb_cursor_last(mc, key, data);
4750 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4751 !(mc->mc_flags & C_INITIALIZED) ||
4752 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4756 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4759 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4767 /** Touch all the pages in the cursor stack.
4768 * Makes sure all the pages are writable, before attempting a write operation.
4769 * @param[in] mc The cursor to operate on.
4772 mdb_cursor_touch(MDB_cursor *mc)
4776 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4779 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4780 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4781 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4784 *mc->mc_dbflag = DB_DIRTY;
4786 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4787 rc = mdb_page_touch(mc);
4791 mc->mc_top = mc->mc_snum-1;
4796 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4799 MDB_node *leaf = NULL;
4800 MDB_val xdata, *rdata, dkey;
4803 int do_sub = 0, insert = 0;
4804 unsigned int mcount = 0;
4808 char dbuf[MDB_MAXKEYSIZE+1];
4809 unsigned int nflags;
4812 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4815 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE)
4818 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4821 #if SIZE_MAX > MAXDATASIZE
4822 if (data->mv_size > MAXDATASIZE)
4826 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4827 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4831 if (flags == MDB_CURRENT) {
4832 if (!(mc->mc_flags & C_INITIALIZED))
4835 } else if (mc->mc_db->md_root == P_INVALID) {
4837 /* new database, write a root leaf page */
4838 DPUTS("allocating new root leaf page");
4839 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4843 mdb_cursor_push(mc, np);
4844 mc->mc_db->md_root = np->mp_pgno;
4845 mc->mc_db->md_depth++;
4846 *mc->mc_dbflag = DB_DIRTY;
4847 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4849 np->mp_flags |= P_LEAF2;
4850 mc->mc_flags |= C_INITIALIZED;
4856 if (flags & MDB_APPEND) {
4858 rc = mdb_cursor_last(mc, &k2, &d2);
4860 rc = mc->mc_dbx->md_cmp(key, &k2);
4863 mc->mc_ki[mc->mc_top]++;
4869 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4871 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4872 DPRINTF("duplicate key [%s]", DKEY(key));
4874 return MDB_KEYEXIST;
4876 if (rc && rc != MDB_NOTFOUND)
4880 /* Cursor is positioned, now make sure all pages are writable */
4881 rc2 = mdb_cursor_touch(mc);
4886 /* The key already exists */
4887 if (rc == MDB_SUCCESS) {
4888 /* there's only a key anyway, so this is a no-op */
4889 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4890 unsigned int ksize = mc->mc_db->md_pad;
4891 if (key->mv_size != ksize)
4893 if (flags == MDB_CURRENT) {
4894 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4895 memcpy(ptr, key->mv_data, ksize);
4900 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4903 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4904 /* Was a single item before, must convert now */
4906 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4907 /* Just overwrite the current item */
4908 if (flags == MDB_CURRENT)
4911 dkey.mv_size = NODEDSZ(leaf);
4912 dkey.mv_data = NODEDATA(leaf);
4913 #if UINT_MAX < SIZE_MAX
4914 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4915 #ifdef MISALIGNED_OK
4916 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4918 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4921 /* if data matches, ignore it */
4922 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4923 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4925 /* create a fake page for the dup items */
4926 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4927 dkey.mv_data = dbuf;
4928 fp = (MDB_page *)&pbuf;
4929 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4930 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4931 fp->mp_lower = PAGEHDRSZ;
4932 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4933 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4934 fp->mp_flags |= P_LEAF2;
4935 fp->mp_pad = data->mv_size;
4936 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4938 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4939 (dkey.mv_size & 1) + (data->mv_size & 1);
4941 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4944 xdata.mv_size = fp->mp_upper;
4949 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4950 /* See if we need to convert from fake page to subDB */
4952 unsigned int offset;
4955 fp = NODEDATA(leaf);
4956 if (flags == MDB_CURRENT) {
4958 fp->mp_flags |= P_DIRTY;
4959 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4960 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4964 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4965 offset = fp->mp_pad;
4966 if (SIZELEFT(fp) >= offset)
4968 offset *= 4; /* space for 4 more */
4970 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4972 offset += offset & 1;
4973 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4974 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4976 /* yes, convert it */
4978 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4979 dummy.md_pad = fp->mp_pad;
4980 dummy.md_flags = MDB_DUPFIXED;
4981 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4982 dummy.md_flags |= MDB_INTEGERKEY;
4985 dummy.md_branch_pages = 0;
4986 dummy.md_leaf_pages = 1;
4987 dummy.md_overflow_pages = 0;
4988 dummy.md_entries = NUMKEYS(fp);
4990 xdata.mv_size = sizeof(MDB_db);
4991 xdata.mv_data = &dummy;
4992 if ((rc = mdb_page_alloc(mc, 1, &mp)))
4994 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4995 flags |= F_DUPDATA|F_SUBDATA;
4996 dummy.md_root = mp->mp_pgno;
4998 /* no, just grow it */
5000 xdata.mv_size = NODEDSZ(leaf) + offset;
5001 xdata.mv_data = &pbuf;
5002 mp = (MDB_page *)&pbuf;
5003 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5006 mp->mp_flags = fp->mp_flags | P_DIRTY;
5007 mp->mp_pad = fp->mp_pad;
5008 mp->mp_lower = fp->mp_lower;
5009 mp->mp_upper = fp->mp_upper + offset;
5011 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5013 nsize = NODEDSZ(leaf) - fp->mp_upper;
5014 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5015 for (i=0; i<NUMKEYS(fp); i++)
5016 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5018 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5022 /* data is on sub-DB, just store it */
5023 flags |= F_DUPDATA|F_SUBDATA;
5027 /* overflow page overwrites need special handling */
5028 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5031 int ovpages, dpages;
5033 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5034 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5035 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5036 mdb_page_get(mc->mc_txn, pg, &omp);
5037 /* Is the ov page writable and large enough? */
5038 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5039 /* yes, overwrite it. Note in this case we don't
5040 * bother to try shrinking the node if the new data
5041 * is smaller than the overflow threshold.
5043 if (F_ISSET(flags, MDB_RESERVE))
5044 data->mv_data = METADATA(omp);
5046 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5049 /* no, free ovpages */
5051 mc->mc_db->md_overflow_pages -= ovpages;
5052 for (i=0; i<ovpages; i++) {
5053 DPRINTF("freed ov page %zu", pg);
5054 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5058 } else if (NODEDSZ(leaf) == data->mv_size) {
5059 /* same size, just replace it. Note that we could
5060 * also reuse this node if the new data is smaller,
5061 * but instead we opt to shrink the node in that case.
5063 if (F_ISSET(flags, MDB_RESERVE))
5064 data->mv_data = NODEDATA(leaf);
5066 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5069 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5070 mc->mc_db->md_entries--;
5072 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5079 nflags = flags & NODE_ADD_FLAGS;
5080 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5081 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5082 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5083 nflags &= ~MDB_APPEND;
5085 nflags |= MDB_SPLIT_REPLACE;
5086 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5088 /* There is room already in this leaf page. */
5089 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5090 if (rc == 0 && !do_sub && insert) {
5091 /* Adjust other cursors pointing to mp */
5092 MDB_cursor *m2, *m3;
5093 MDB_dbi dbi = mc->mc_dbi;
5094 unsigned i = mc->mc_top;
5095 MDB_page *mp = mc->mc_pg[i];
5097 if (mc->mc_flags & C_SUB)
5100 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5101 if (mc->mc_flags & C_SUB)
5102 m3 = &m2->mc_xcursor->mx_cursor;
5105 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5106 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5113 if (rc != MDB_SUCCESS)
5114 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5116 /* Now store the actual data in the child DB. Note that we're
5117 * storing the user data in the keys field, so there are strict
5118 * size limits on dupdata. The actual data fields of the child
5119 * DB are all zero size.
5126 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5127 if (flags & MDB_CURRENT) {
5128 xflags = MDB_CURRENT;
5130 mdb_xcursor_init1(mc, leaf);
5131 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5133 /* converted, write the original data first */
5135 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5139 /* Adjust other cursors pointing to mp */
5141 unsigned i = mc->mc_top;
5142 MDB_page *mp = mc->mc_pg[i];
5144 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5145 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5146 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5147 mdb_xcursor_init1(m2, leaf);
5152 if (flags & MDB_APPENDDUP)
5153 xflags |= MDB_APPEND;
5154 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5155 if (flags & F_SUBDATA) {
5156 void *db = NODEDATA(leaf);
5157 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5160 /* sub-writes might have failed so check rc again.
5161 * Don't increment count if we just replaced an existing item.
5163 if (!rc && !(flags & MDB_CURRENT))
5164 mc->mc_db->md_entries++;
5165 if (flags & MDB_MULTIPLE) {
5167 if (mcount < data[1].mv_size) {
5168 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5169 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5179 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5184 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5187 if (!mc->mc_flags & C_INITIALIZED)
5190 rc = mdb_cursor_touch(mc);
5194 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5196 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5197 if (flags != MDB_NODUPDATA) {
5198 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5199 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5201 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5202 /* If sub-DB still has entries, we're done */
5203 if (mc->mc_xcursor->mx_db.md_entries) {
5204 if (leaf->mn_flags & F_SUBDATA) {
5205 /* update subDB info */
5206 void *db = NODEDATA(leaf);
5207 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5209 /* shrink fake page */
5210 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5212 mc->mc_db->md_entries--;
5215 /* otherwise fall thru and delete the sub-DB */
5218 if (leaf->mn_flags & F_SUBDATA) {
5219 /* add all the child DB's pages to the free list */
5220 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5221 if (rc == MDB_SUCCESS) {
5222 mc->mc_db->md_entries -=
5223 mc->mc_xcursor->mx_db.md_entries;
5228 return mdb_cursor_del0(mc, leaf);
5231 /** Allocate and initialize new pages for a database.
5232 * @param[in] mc a cursor on the database being added to.
5233 * @param[in] flags flags defining what type of page is being allocated.
5234 * @param[in] num the number of pages to allocate. This is usually 1,
5235 * unless allocating overflow pages for a large record.
5236 * @param[out] mp Address of a page, or NULL on failure.
5237 * @return 0 on success, non-zero on failure.
5240 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5245 if ((rc = mdb_page_alloc(mc, num, &np)))
5247 DPRINTF("allocated new mpage %zu, page size %u",
5248 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5249 np->mp_flags = flags | P_DIRTY;
5250 np->mp_lower = PAGEHDRSZ;
5251 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5254 mc->mc_db->md_branch_pages++;
5255 else if (IS_LEAF(np))
5256 mc->mc_db->md_leaf_pages++;
5257 else if (IS_OVERFLOW(np)) {
5258 mc->mc_db->md_overflow_pages += num;
5266 /** Calculate the size of a leaf node.
5267 * The size depends on the environment's page size; if a data item
5268 * is too large it will be put onto an overflow page and the node
5269 * size will only include the key and not the data. Sizes are always
5270 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5271 * of the #MDB_node headers.
5272 * @param[in] env The environment handle.
5273 * @param[in] key The key for the node.
5274 * @param[in] data The data for the node.
5275 * @return The number of bytes needed to store the node.
5278 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5282 sz = LEAFSIZE(key, data);
5283 if (sz >= env->me_psize / MDB_MINKEYS) {
5284 /* put on overflow page */
5285 sz -= data->mv_size - sizeof(pgno_t);
5289 return sz + sizeof(indx_t);
5292 /** Calculate the size of a branch node.
5293 * The size should depend on the environment's page size but since
5294 * we currently don't support spilling large keys onto overflow
5295 * pages, it's simply the size of the #MDB_node header plus the
5296 * size of the key. Sizes are always rounded up to an even number
5297 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5298 * @param[in] env The environment handle.
5299 * @param[in] key The key for the node.
5300 * @return The number of bytes needed to store the node.
5303 mdb_branch_size(MDB_env *env, MDB_val *key)
5308 if (sz >= env->me_psize / MDB_MINKEYS) {
5309 /* put on overflow page */
5310 /* not implemented */
5311 /* sz -= key->size - sizeof(pgno_t); */
5314 return sz + sizeof(indx_t);
5317 /** Add a node to the page pointed to by the cursor.
5318 * @param[in] mc The cursor for this operation.
5319 * @param[in] indx The index on the page where the new node should be added.
5320 * @param[in] key The key for the new node.
5321 * @param[in] data The data for the new node, if any.
5322 * @param[in] pgno The page number, if adding a branch node.
5323 * @param[in] flags Flags for the node.
5324 * @return 0 on success, non-zero on failure. Possible errors are:
5326 * <li>ENOMEM - failed to allocate overflow pages for the node.
5327 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5328 * should never happen since all callers already calculate the
5329 * page's free space before calling this function.
5333 mdb_node_add(MDB_cursor *mc, indx_t indx,
5334 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5337 size_t node_size = NODESIZE;
5340 MDB_page *mp = mc->mc_pg[mc->mc_top];
5341 MDB_page *ofp = NULL; /* overflow page */
5344 assert(mp->mp_upper >= mp->mp_lower);
5346 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5347 IS_LEAF(mp) ? "leaf" : "branch",
5348 IS_SUBP(mp) ? "sub-" : "",
5349 mp->mp_pgno, indx, data ? data->mv_size : 0,
5350 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5353 /* Move higher keys up one slot. */
5354 int ksize = mc->mc_db->md_pad, dif;
5355 char *ptr = LEAF2KEY(mp, indx, ksize);
5356 dif = NUMKEYS(mp) - indx;
5358 memmove(ptr+ksize, ptr, dif*ksize);
5359 /* insert new key */
5360 memcpy(ptr, key->mv_data, ksize);
5362 /* Just using these for counting */
5363 mp->mp_lower += sizeof(indx_t);
5364 mp->mp_upper -= ksize - sizeof(indx_t);
5369 node_size += key->mv_size;
5373 if (F_ISSET(flags, F_BIGDATA)) {
5374 /* Data already on overflow page. */
5375 node_size += sizeof(pgno_t);
5376 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5377 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5379 /* Put data on overflow page. */
5380 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5381 data->mv_size, node_size+data->mv_size);
5382 node_size += sizeof(pgno_t);
5383 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5385 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5388 node_size += data->mv_size;
5391 node_size += node_size & 1;
5393 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5394 DPRINTF("not enough room in page %zu, got %u ptrs",
5395 mp->mp_pgno, NUMKEYS(mp));
5396 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5397 mp->mp_upper - mp->mp_lower);
5398 DPRINTF("node size = %zu", node_size);
5399 return MDB_PAGE_FULL;
5402 /* Move higher pointers up one slot. */
5403 for (i = NUMKEYS(mp); i > indx; i--)
5404 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5406 /* Adjust free space offsets. */
5407 ofs = mp->mp_upper - node_size;
5408 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5409 mp->mp_ptrs[indx] = ofs;
5411 mp->mp_lower += sizeof(indx_t);
5413 /* Write the node data. */
5414 node = NODEPTR(mp, indx);
5415 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5416 node->mn_flags = flags;
5418 SETDSZ(node,data->mv_size);
5423 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5428 if (F_ISSET(flags, F_BIGDATA))
5429 memcpy(node->mn_data + key->mv_size, data->mv_data,
5431 else if (F_ISSET(flags, MDB_RESERVE))
5432 data->mv_data = node->mn_data + key->mv_size;
5434 memcpy(node->mn_data + key->mv_size, data->mv_data,
5437 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5439 if (F_ISSET(flags, MDB_RESERVE))
5440 data->mv_data = METADATA(ofp);
5442 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5449 /** Delete the specified node from a page.
5450 * @param[in] mp The page to operate on.
5451 * @param[in] indx The index of the node to delete.
5452 * @param[in] ksize The size of a node. Only used if the page is
5453 * part of a #MDB_DUPFIXED database.
5456 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5459 indx_t i, j, numkeys, ptr;
5466 COPY_PGNO(pgno, mp->mp_pgno);
5467 DPRINTF("delete node %u on %s page %zu", indx,
5468 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5471 assert(indx < NUMKEYS(mp));
5474 int x = NUMKEYS(mp) - 1 - indx;
5475 base = LEAF2KEY(mp, indx, ksize);
5477 memmove(base, base + ksize, x * ksize);
5478 mp->mp_lower -= sizeof(indx_t);
5479 mp->mp_upper += ksize - sizeof(indx_t);
5483 node = NODEPTR(mp, indx);
5484 sz = NODESIZE + node->mn_ksize;
5486 if (F_ISSET(node->mn_flags, F_BIGDATA))
5487 sz += sizeof(pgno_t);
5489 sz += NODEDSZ(node);
5493 ptr = mp->mp_ptrs[indx];
5494 numkeys = NUMKEYS(mp);
5495 for (i = j = 0; i < numkeys; i++) {
5497 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5498 if (mp->mp_ptrs[i] < ptr)
5499 mp->mp_ptrs[j] += sz;
5504 base = (char *)mp + mp->mp_upper;
5505 memmove(base + sz, base, ptr - mp->mp_upper);
5507 mp->mp_lower -= sizeof(indx_t);
5511 /** Compact the main page after deleting a node on a subpage.
5512 * @param[in] mp The main page to operate on.
5513 * @param[in] indx The index of the subpage on the main page.
5516 mdb_node_shrink(MDB_page *mp, indx_t indx)
5523 indx_t i, numkeys, ptr;
5525 node = NODEPTR(mp, indx);
5526 sp = (MDB_page *)NODEDATA(node);
5527 osize = NODEDSZ(node);
5529 delta = sp->mp_upper - sp->mp_lower;
5530 SETDSZ(node, osize - delta);
5531 xp = (MDB_page *)((char *)sp + delta);
5533 /* shift subpage upward */
5535 nsize = NUMKEYS(sp) * sp->mp_pad;
5536 memmove(METADATA(xp), METADATA(sp), nsize);
5539 nsize = osize - sp->mp_upper;
5540 numkeys = NUMKEYS(sp);
5541 for (i=numkeys-1; i>=0; i--)
5542 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5544 xp->mp_upper = sp->mp_lower;
5545 xp->mp_lower = sp->mp_lower;
5546 xp->mp_flags = sp->mp_flags;
5547 xp->mp_pad = sp->mp_pad;
5548 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5550 /* shift lower nodes upward */
5551 ptr = mp->mp_ptrs[indx];
5552 numkeys = NUMKEYS(mp);
5553 for (i = 0; i < numkeys; i++) {
5554 if (mp->mp_ptrs[i] <= ptr)
5555 mp->mp_ptrs[i] += delta;
5558 base = (char *)mp + mp->mp_upper;
5559 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5560 mp->mp_upper += delta;
5563 /** Initial setup of a sorted-dups cursor.
5564 * Sorted duplicates are implemented as a sub-database for the given key.
5565 * The duplicate data items are actually keys of the sub-database.
5566 * Operations on the duplicate data items are performed using a sub-cursor
5567 * initialized when the sub-database is first accessed. This function does
5568 * the preliminary setup of the sub-cursor, filling in the fields that
5569 * depend only on the parent DB.
5570 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5573 mdb_xcursor_init0(MDB_cursor *mc)
5575 MDB_xcursor *mx = mc->mc_xcursor;
5577 mx->mx_cursor.mc_xcursor = NULL;
5578 mx->mx_cursor.mc_txn = mc->mc_txn;
5579 mx->mx_cursor.mc_db = &mx->mx_db;
5580 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5581 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5582 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5583 mx->mx_cursor.mc_snum = 0;
5584 mx->mx_cursor.mc_top = 0;
5585 mx->mx_cursor.mc_flags = C_SUB;
5586 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5587 mx->mx_dbx.md_dcmp = NULL;
5588 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5591 /** Final setup of a sorted-dups cursor.
5592 * Sets up the fields that depend on the data from the main cursor.
5593 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5594 * @param[in] node The data containing the #MDB_db record for the
5595 * sorted-dup database.
5598 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5600 MDB_xcursor *mx = mc->mc_xcursor;
5602 if (node->mn_flags & F_SUBDATA) {
5603 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5604 mx->mx_cursor.mc_pg[0] = 0;
5605 mx->mx_cursor.mc_snum = 0;
5606 mx->mx_cursor.mc_flags = C_SUB;
5608 MDB_page *fp = NODEDATA(node);
5609 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5610 mx->mx_db.md_flags = 0;
5611 mx->mx_db.md_depth = 1;
5612 mx->mx_db.md_branch_pages = 0;
5613 mx->mx_db.md_leaf_pages = 1;
5614 mx->mx_db.md_overflow_pages = 0;
5615 mx->mx_db.md_entries = NUMKEYS(fp);
5616 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5617 mx->mx_cursor.mc_snum = 1;
5618 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5619 mx->mx_cursor.mc_top = 0;
5620 mx->mx_cursor.mc_pg[0] = fp;
5621 mx->mx_cursor.mc_ki[0] = 0;
5622 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5623 mx->mx_db.md_flags = MDB_DUPFIXED;
5624 mx->mx_db.md_pad = fp->mp_pad;
5625 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5626 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5629 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5631 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5633 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5634 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5635 #if UINT_MAX < SIZE_MAX
5636 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5637 #ifdef MISALIGNED_OK
5638 mx->mx_dbx.md_cmp = mdb_cmp_long;
5640 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5645 /** Initialize a cursor for a given transaction and database. */
5647 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5652 mc->mc_db = &txn->mt_dbs[dbi];
5653 mc->mc_dbx = &txn->mt_dbxs[dbi];
5654 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5659 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5661 mc->mc_xcursor = mx;
5662 mdb_xcursor_init0(mc);
5664 mc->mc_xcursor = NULL;
5666 if (*mc->mc_dbflag & DB_STALE) {
5667 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5672 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5675 MDB_xcursor *mx = NULL;
5676 size_t size = sizeof(MDB_cursor);
5678 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5681 /* Allow read access to the freelist */
5682 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5685 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5686 size += sizeof(MDB_xcursor);
5688 if ((mc = malloc(size)) != NULL) {
5689 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5690 mx = (MDB_xcursor *)(mc + 1);
5692 mdb_cursor_init(mc, txn, dbi, mx);
5693 if (txn->mt_cursors) {
5694 mc->mc_next = txn->mt_cursors[dbi];
5695 txn->mt_cursors[dbi] = mc;
5697 mc->mc_flags |= C_ALLOCD;
5708 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5710 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5713 if (txn->mt_cursors)
5716 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5720 /* Return the count of duplicate data items for the current key */
5722 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5726 if (mc == NULL || countp == NULL)
5729 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5732 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5733 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5736 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5739 *countp = mc->mc_xcursor->mx_db.md_entries;
5745 mdb_cursor_close(MDB_cursor *mc)
5748 /* remove from txn, if tracked */
5749 if (mc->mc_txn->mt_cursors) {
5750 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5751 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5753 *prev = mc->mc_next;
5755 if (mc->mc_flags & C_ALLOCD)
5761 mdb_cursor_txn(MDB_cursor *mc)
5763 if (!mc) return NULL;
5768 mdb_cursor_dbi(MDB_cursor *mc)
5774 /** Replace the key for a node with a new key.
5775 * @param[in] mp The page containing the node to operate on.
5776 * @param[in] indx The index of the node to operate on.
5777 * @param[in] key The new key to use.
5778 * @return 0 on success, non-zero on failure.
5781 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5787 indx_t ptr, i, numkeys;
5790 node = NODEPTR(mp, indx);
5791 ptr = mp->mp_ptrs[indx];
5795 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5796 k2.mv_data = NODEKEY(node);
5797 k2.mv_size = node->mn_ksize;
5798 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5800 mdb_dkey(&k2, kbuf2),
5806 delta0 = delta = key->mv_size - node->mn_ksize;
5808 /* Must be 2-byte aligned. If new key is
5809 * shorter by 1, the shift will be skipped.
5811 delta += (delta & 1);
5813 if (delta > 0 && SIZELEFT(mp) < delta) {
5814 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5815 return MDB_PAGE_FULL;
5818 numkeys = NUMKEYS(mp);
5819 for (i = 0; i < numkeys; i++) {
5820 if (mp->mp_ptrs[i] <= ptr)
5821 mp->mp_ptrs[i] -= delta;
5824 base = (char *)mp + mp->mp_upper;
5825 len = ptr - mp->mp_upper + NODESIZE;
5826 memmove(base - delta, base, len);
5827 mp->mp_upper -= delta;
5829 node = NODEPTR(mp, indx);
5832 /* But even if no shift was needed, update ksize */
5834 node->mn_ksize = key->mv_size;
5837 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5842 /** Move a node from csrc to cdst.
5845 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5851 unsigned short flags;
5855 /* Mark src and dst as dirty. */
5856 if ((rc = mdb_page_touch(csrc)) ||
5857 (rc = mdb_page_touch(cdst)))
5860 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5861 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5862 key.mv_size = csrc->mc_db->md_pad;
5863 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5865 data.mv_data = NULL;
5869 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5870 assert(!((long)srcnode&1));
5871 srcpg = NODEPGNO(srcnode);
5872 flags = srcnode->mn_flags;
5873 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5874 unsigned int snum = csrc->mc_snum;
5876 /* must find the lowest key below src */
5877 mdb_page_search_root(csrc, NULL, 0);
5878 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5879 key.mv_size = csrc->mc_db->md_pad;
5880 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5882 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5883 key.mv_size = NODEKSZ(s2);
5884 key.mv_data = NODEKEY(s2);
5886 csrc->mc_snum = snum--;
5887 csrc->mc_top = snum;
5889 key.mv_size = NODEKSZ(srcnode);
5890 key.mv_data = NODEKEY(srcnode);
5892 data.mv_size = NODEDSZ(srcnode);
5893 data.mv_data = NODEDATA(srcnode);
5895 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5896 unsigned int snum = cdst->mc_snum;
5899 /* must find the lowest key below dst */
5900 mdb_page_search_root(cdst, NULL, 0);
5901 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5902 bkey.mv_size = cdst->mc_db->md_pad;
5903 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5905 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5906 bkey.mv_size = NODEKSZ(s2);
5907 bkey.mv_data = NODEKEY(s2);
5909 cdst->mc_snum = snum--;
5910 cdst->mc_top = snum;
5911 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5914 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5915 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5916 csrc->mc_ki[csrc->mc_top],
5918 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5919 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5921 /* Add the node to the destination page.
5923 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5924 if (rc != MDB_SUCCESS)
5927 /* Delete the node from the source page.
5929 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5932 /* Adjust other cursors pointing to mp */
5933 MDB_cursor *m2, *m3;
5934 MDB_dbi dbi = csrc->mc_dbi;
5935 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5937 if (csrc->mc_flags & C_SUB)
5940 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5941 if (m2 == csrc) continue;
5942 if (csrc->mc_flags & C_SUB)
5943 m3 = &m2->mc_xcursor->mx_cursor;
5946 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5947 csrc->mc_ki[csrc->mc_top]) {
5948 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5949 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5954 /* Update the parent separators.
5956 if (csrc->mc_ki[csrc->mc_top] == 0) {
5957 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5958 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5959 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5961 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5962 key.mv_size = NODEKSZ(srcnode);
5963 key.mv_data = NODEKEY(srcnode);
5965 DPRINTF("update separator for source page %zu to [%s]",
5966 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5967 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5968 &key)) != MDB_SUCCESS)
5971 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5973 nullkey.mv_size = 0;
5974 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5975 assert(rc == MDB_SUCCESS);
5979 if (cdst->mc_ki[cdst->mc_top] == 0) {
5980 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5981 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5982 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5984 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5985 key.mv_size = NODEKSZ(srcnode);
5986 key.mv_data = NODEKEY(srcnode);
5988 DPRINTF("update separator for destination page %zu to [%s]",
5989 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5990 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5991 &key)) != MDB_SUCCESS)
5994 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5996 nullkey.mv_size = 0;
5997 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5998 assert(rc == MDB_SUCCESS);
6005 /** Merge one page into another.
6006 * The nodes from the page pointed to by \b csrc will
6007 * be copied to the page pointed to by \b cdst and then
6008 * the \b csrc page will be freed.
6009 * @param[in] csrc Cursor pointing to the source page.
6010 * @param[in] cdst Cursor pointing to the destination page.
6013 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6021 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6022 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6024 assert(csrc->mc_snum > 1); /* can't merge root page */
6025 assert(cdst->mc_snum > 1);
6027 /* Mark dst as dirty. */
6028 if ((rc = mdb_page_touch(cdst)))
6031 /* Move all nodes from src to dst.
6033 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6034 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6035 key.mv_size = csrc->mc_db->md_pad;
6036 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6037 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6038 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6039 if (rc != MDB_SUCCESS)
6041 key.mv_data = (char *)key.mv_data + key.mv_size;
6044 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6045 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6046 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6047 unsigned int snum = csrc->mc_snum;
6049 /* must find the lowest key below src */
6050 mdb_page_search_root(csrc, NULL, 0);
6051 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6052 key.mv_size = csrc->mc_db->md_pad;
6053 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6055 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6056 key.mv_size = NODEKSZ(s2);
6057 key.mv_data = NODEKEY(s2);
6059 csrc->mc_snum = snum--;
6060 csrc->mc_top = snum;
6062 key.mv_size = srcnode->mn_ksize;
6063 key.mv_data = NODEKEY(srcnode);
6066 data.mv_size = NODEDSZ(srcnode);
6067 data.mv_data = NODEDATA(srcnode);
6068 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6069 if (rc != MDB_SUCCESS)
6074 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6075 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]), (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10);
6077 /* Unlink the src page from parent and add to free list.
6079 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6080 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6082 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
6086 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6087 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6088 csrc->mc_db->md_leaf_pages--;
6090 csrc->mc_db->md_branch_pages--;
6092 /* Adjust other cursors pointing to mp */
6093 MDB_cursor *m2, *m3;
6094 MDB_dbi dbi = csrc->mc_dbi;
6095 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6097 if (csrc->mc_flags & C_SUB)
6100 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6101 if (csrc->mc_flags & C_SUB)
6102 m3 = &m2->mc_xcursor->mx_cursor;
6105 if (m3 == csrc) continue;
6106 if (m3->mc_snum < csrc->mc_snum) continue;
6107 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6108 m3->mc_pg[csrc->mc_top] = mp;
6109 m3->mc_ki[csrc->mc_top] += nkeys;
6113 mdb_cursor_pop(csrc);
6115 return mdb_rebalance(csrc);
6118 /** Copy the contents of a cursor.
6119 * @param[in] csrc The cursor to copy from.
6120 * @param[out] cdst The cursor to copy to.
6123 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6127 cdst->mc_txn = csrc->mc_txn;
6128 cdst->mc_dbi = csrc->mc_dbi;
6129 cdst->mc_db = csrc->mc_db;
6130 cdst->mc_dbx = csrc->mc_dbx;
6131 cdst->mc_snum = csrc->mc_snum;
6132 cdst->mc_top = csrc->mc_top;
6133 cdst->mc_flags = csrc->mc_flags;
6135 for (i=0; i<csrc->mc_snum; i++) {
6136 cdst->mc_pg[i] = csrc->mc_pg[i];
6137 cdst->mc_ki[i] = csrc->mc_ki[i];
6141 /** Rebalance the tree after a delete operation.
6142 * @param[in] mc Cursor pointing to the page where rebalancing
6144 * @return 0 on success, non-zero on failure.
6147 mdb_rebalance(MDB_cursor *mc)
6157 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6158 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6159 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6160 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6164 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
6167 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6168 DPRINTF("no need to rebalance page %zu, above fill threshold",
6174 if (mc->mc_snum < 2) {
6175 MDB_page *mp = mc->mc_pg[0];
6176 if (NUMKEYS(mp) == 0) {
6177 DPUTS("tree is completely empty");
6178 mc->mc_db->md_root = P_INVALID;
6179 mc->mc_db->md_depth = 0;
6180 mc->mc_db->md_leaf_pages = 0;
6181 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6185 /* Adjust other cursors pointing to mp */
6186 MDB_cursor *m2, *m3;
6187 MDB_dbi dbi = mc->mc_dbi;
6189 if (mc->mc_flags & C_SUB)
6192 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6193 if (m2 == mc) continue;
6194 if (mc->mc_flags & C_SUB)
6195 m3 = &m2->mc_xcursor->mx_cursor;
6198 if (m3->mc_snum < mc->mc_snum) continue;
6199 if (m3->mc_pg[0] == mp) {
6205 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6206 DPUTS("collapsing root page!");
6207 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6208 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6209 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6212 mc->mc_db->md_depth--;
6213 mc->mc_db->md_branch_pages--;
6215 /* Adjust other cursors pointing to mp */
6216 MDB_cursor *m2, *m3;
6217 MDB_dbi dbi = mc->mc_dbi;
6219 if (mc->mc_flags & C_SUB)
6222 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6223 if (m2 == mc) continue;
6224 if (mc->mc_flags & C_SUB)
6225 m3 = &m2->mc_xcursor->mx_cursor;
6228 if (m3->mc_snum < mc->mc_snum) continue;
6229 if (m3->mc_pg[0] == mp) {
6230 m3->mc_pg[0] = mc->mc_pg[0];
6235 DPUTS("root page doesn't need rebalancing");
6239 /* The parent (branch page) must have at least 2 pointers,
6240 * otherwise the tree is invalid.
6242 ptop = mc->mc_top-1;
6243 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6245 /* Leaf page fill factor is below the threshold.
6246 * Try to move keys from left or right neighbor, or
6247 * merge with a neighbor page.
6252 mdb_cursor_copy(mc, &mn);
6253 mn.mc_xcursor = NULL;
6255 if (mc->mc_ki[ptop] == 0) {
6256 /* We're the leftmost leaf in our parent.
6258 DPUTS("reading right neighbor");
6260 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6261 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6263 mn.mc_ki[mn.mc_top] = 0;
6264 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6266 /* There is at least one neighbor to the left.
6268 DPUTS("reading left neighbor");
6270 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6271 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6273 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6274 mc->mc_ki[mc->mc_top] = 0;
6277 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6278 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10);
6280 /* If the neighbor page is above threshold and has at least two
6281 * keys, move one key from it.
6283 * Otherwise we should try to merge them.
6285 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6286 return mdb_node_move(&mn, mc);
6287 else { /* FIXME: if (has_enough_room()) */
6288 mc->mc_flags &= ~C_INITIALIZED;
6289 if (mc->mc_ki[ptop] == 0)
6290 return mdb_page_merge(&mn, mc);
6292 return mdb_page_merge(mc, &mn);
6296 /** Complete a delete operation started by #mdb_cursor_del(). */
6298 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6302 /* add overflow pages to free list */
6303 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6307 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6308 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6309 mc->mc_db->md_overflow_pages -= ovpages;
6310 for (i=0; i<ovpages; i++) {
6311 DPRINTF("freed ov page %zu", pg);
6312 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6316 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6317 mc->mc_db->md_entries--;
6318 rc = mdb_rebalance(mc);
6319 if (rc != MDB_SUCCESS)
6320 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6326 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6327 MDB_val *key, MDB_val *data)
6332 MDB_val rdata, *xdata;
6336 assert(key != NULL);
6338 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6340 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6343 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6347 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6351 mdb_cursor_init(&mc, txn, dbi, &mx);
6362 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6364 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6368 /** Split a page and insert a new node.
6369 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6370 * The cursor will be updated to point to the actual page and index where
6371 * the node got inserted after the split.
6372 * @param[in] newkey The key for the newly inserted node.
6373 * @param[in] newdata The data for the newly inserted node.
6374 * @param[in] newpgno The page number, if the new node is a branch node.
6375 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6376 * @return 0 on success, non-zero on failure.
6379 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6380 unsigned int nflags)
6383 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6386 unsigned int i, j, split_indx, nkeys, pmax;
6388 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6390 MDB_page *mp, *rp, *pp;
6395 mp = mc->mc_pg[mc->mc_top];
6396 newindx = mc->mc_ki[mc->mc_top];
6398 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6399 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6400 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6402 /* Create a right sibling. */
6403 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6405 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6407 if (mc->mc_snum < 2) {
6408 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6410 /* shift current top to make room for new parent */
6411 mc->mc_pg[1] = mc->mc_pg[0];
6412 mc->mc_ki[1] = mc->mc_ki[0];
6415 mc->mc_db->md_root = pp->mp_pgno;
6416 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6417 mc->mc_db->md_depth++;
6420 /* Add left (implicit) pointer. */
6421 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6422 /* undo the pre-push */
6423 mc->mc_pg[0] = mc->mc_pg[1];
6424 mc->mc_ki[0] = mc->mc_ki[1];
6425 mc->mc_db->md_root = mp->mp_pgno;
6426 mc->mc_db->md_depth--;
6433 ptop = mc->mc_top-1;
6434 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6437 mc->mc_flags |= C_SPLITTING;
6438 mdb_cursor_copy(mc, &mn);
6439 mn.mc_pg[mn.mc_top] = rp;
6440 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6442 if (nflags & MDB_APPEND) {
6443 mn.mc_ki[mn.mc_top] = 0;
6445 split_indx = newindx;
6450 nkeys = NUMKEYS(mp);
6451 split_indx = nkeys / 2;
6452 if (newindx < split_indx)
6458 unsigned int lsize, rsize, ksize;
6459 /* Move half of the keys to the right sibling */
6461 x = mc->mc_ki[mc->mc_top] - split_indx;
6462 ksize = mc->mc_db->md_pad;
6463 split = LEAF2KEY(mp, split_indx, ksize);
6464 rsize = (nkeys - split_indx) * ksize;
6465 lsize = (nkeys - split_indx) * sizeof(indx_t);
6466 mp->mp_lower -= lsize;
6467 rp->mp_lower += lsize;
6468 mp->mp_upper += rsize - lsize;
6469 rp->mp_upper -= rsize - lsize;
6470 sepkey.mv_size = ksize;
6471 if (newindx == split_indx) {
6472 sepkey.mv_data = newkey->mv_data;
6474 sepkey.mv_data = split;
6477 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6478 memcpy(rp->mp_ptrs, split, rsize);
6479 sepkey.mv_data = rp->mp_ptrs;
6480 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6481 memcpy(ins, newkey->mv_data, ksize);
6482 mp->mp_lower += sizeof(indx_t);
6483 mp->mp_upper -= ksize - sizeof(indx_t);
6486 memcpy(rp->mp_ptrs, split, x * ksize);
6487 ins = LEAF2KEY(rp, x, ksize);
6488 memcpy(ins, newkey->mv_data, ksize);
6489 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6490 rp->mp_lower += sizeof(indx_t);
6491 rp->mp_upper -= ksize - sizeof(indx_t);
6492 mc->mc_ki[mc->mc_top] = x;
6493 mc->mc_pg[mc->mc_top] = rp;
6498 /* For leaf pages, check the split point based on what
6499 * fits where, since otherwise mdb_node_add can fail.
6501 * This check is only needed when the data items are
6502 * relatively large, such that being off by one will
6503 * make the difference between success or failure.
6505 * It's also relevant if a page happens to be laid out
6506 * such that one half of its nodes are all "small" and
6507 * the other half of its nodes are "large." If the new
6508 * item is also "large" and falls on the half with
6509 * "large" nodes, it also may not fit.
6512 unsigned int psize, nsize;
6513 /* Maximum free space in an empty page */
6514 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6515 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6516 if ((nkeys < 20) || (nsize > pmax/16)) {
6517 if (newindx <= split_indx) {
6520 for (i=0; i<split_indx; i++) {
6521 node = NODEPTR(mp, i);
6522 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6523 if (F_ISSET(node->mn_flags, F_BIGDATA))
6524 psize += sizeof(pgno_t);
6526 psize += NODEDSZ(node);
6530 split_indx = newindx;
6541 for (i=nkeys-1; i>=split_indx; i--) {
6542 node = NODEPTR(mp, i);
6543 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6544 if (F_ISSET(node->mn_flags, F_BIGDATA))
6545 psize += sizeof(pgno_t);
6547 psize += NODEDSZ(node);
6551 split_indx = newindx;
6562 /* First find the separating key between the split pages.
6563 * The case where newindx == split_indx is ambiguous; the
6564 * new item could go to the new page or stay on the original
6565 * page. If newpos == 1 it goes to the new page.
6567 if (newindx == split_indx && newpos) {
6568 sepkey.mv_size = newkey->mv_size;
6569 sepkey.mv_data = newkey->mv_data;
6571 node = NODEPTR(mp, split_indx);
6572 sepkey.mv_size = node->mn_ksize;
6573 sepkey.mv_data = NODEKEY(node);
6577 DPRINTF("separator is [%s]", DKEY(&sepkey));
6579 /* Copy separator key to the parent.
6581 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6585 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6588 if (mn.mc_snum == mc->mc_snum) {
6589 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6590 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6591 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6592 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6597 /* Right page might now have changed parent.
6598 * Check if left page also changed parent.
6600 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6601 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6602 for (i=0; i<ptop; i++) {
6603 mc->mc_pg[i] = mn.mc_pg[i];
6604 mc->mc_ki[i] = mn.mc_ki[i];
6606 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6607 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6611 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6614 mc->mc_flags ^= C_SPLITTING;
6615 if (rc != MDB_SUCCESS) {
6618 if (nflags & MDB_APPEND) {
6619 mc->mc_pg[mc->mc_top] = rp;
6620 mc->mc_ki[mc->mc_top] = 0;
6621 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6624 for (i=0; i<mc->mc_top; i++)
6625 mc->mc_ki[i] = mn.mc_ki[i];
6632 /* Move half of the keys to the right sibling. */
6634 /* grab a page to hold a temporary copy */
6635 copy = mdb_page_malloc(mc);
6639 copy->mp_pgno = mp->mp_pgno;
6640 copy->mp_flags = mp->mp_flags;
6641 copy->mp_lower = PAGEHDRSZ;
6642 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6643 mc->mc_pg[mc->mc_top] = copy;
6644 for (i = j = 0; i <= nkeys; j++) {
6645 if (i == split_indx) {
6646 /* Insert in right sibling. */
6647 /* Reset insert index for right sibling. */
6648 if (i != newindx || (newpos ^ ins_new)) {
6650 mc->mc_pg[mc->mc_top] = rp;
6654 if (i == newindx && !ins_new) {
6655 /* Insert the original entry that caused the split. */
6656 rkey.mv_data = newkey->mv_data;
6657 rkey.mv_size = newkey->mv_size;
6666 /* Update index for the new key. */
6667 mc->mc_ki[mc->mc_top] = j;
6668 } else if (i == nkeys) {
6671 node = NODEPTR(mp, i);
6672 rkey.mv_data = NODEKEY(node);
6673 rkey.mv_size = node->mn_ksize;
6675 xdata.mv_data = NODEDATA(node);
6676 xdata.mv_size = NODEDSZ(node);
6679 pgno = NODEPGNO(node);
6680 flags = node->mn_flags;
6685 if (!IS_LEAF(mp) && j == 0) {
6686 /* First branch index doesn't need key data. */
6690 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6694 nkeys = NUMKEYS(copy);
6695 for (i=0; i<nkeys; i++)
6696 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6697 mp->mp_lower = copy->mp_lower;
6698 mp->mp_upper = copy->mp_upper;
6699 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6700 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6702 /* reset back to original page */
6703 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6704 mc->mc_pg[mc->mc_top] = mp;
6705 if (nflags & MDB_RESERVE) {
6706 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6707 if (!(node->mn_flags & F_BIGDATA))
6708 newdata->mv_data = NODEDATA(node);
6714 /* return tmp page to freelist */
6715 mdb_page_free(mc->mc_txn->mt_env, copy);
6718 /* Adjust other cursors pointing to mp */
6719 MDB_cursor *m2, *m3;
6720 MDB_dbi dbi = mc->mc_dbi;
6721 int fixup = NUMKEYS(mp);
6723 if (mc->mc_flags & C_SUB)
6726 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6727 if (m2 == mc) continue;
6728 if (mc->mc_flags & C_SUB)
6729 m3 = &m2->mc_xcursor->mx_cursor;
6732 if (!(m3->mc_flags & C_INITIALIZED))
6734 if (m3->mc_flags & C_SPLITTING)
6739 for (k=m3->mc_top; k>=0; k--) {
6740 m3->mc_ki[k+1] = m3->mc_ki[k];
6741 m3->mc_pg[k+1] = m3->mc_pg[k];
6743 if (m3->mc_ki[0] >= split_indx) {
6748 m3->mc_pg[0] = mc->mc_pg[0];
6752 if (m3->mc_pg[mc->mc_top] == mp) {
6753 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6754 m3->mc_ki[mc->mc_top]++;
6755 if (m3->mc_ki[mc->mc_top] >= fixup) {
6756 m3->mc_pg[mc->mc_top] = rp;
6757 m3->mc_ki[mc->mc_top] -= fixup;
6758 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6760 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6761 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6770 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6771 MDB_val *key, MDB_val *data, unsigned int flags)
6776 assert(key != NULL);
6777 assert(data != NULL);
6779 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6782 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6786 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6790 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6793 mdb_cursor_init(&mc, txn, dbi, &mx);
6794 return mdb_cursor_put(&mc, key, data, flags);
6798 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6800 if ((flag & CHANGEABLE) != flag)
6803 env->me_flags |= flag;
6805 env->me_flags &= ~flag;
6810 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6815 *arg = env->me_flags;
6820 mdb_env_get_path(MDB_env *env, const char **arg)
6825 *arg = env->me_path;
6829 /** Common code for #mdb_stat() and #mdb_env_stat().
6830 * @param[in] env the environment to operate in.
6831 * @param[in] db the #MDB_db record containing the stats to return.
6832 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6833 * @return 0, this function always succeeds.
6836 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6838 arg->ms_psize = env->me_psize;
6839 arg->ms_depth = db->md_depth;
6840 arg->ms_branch_pages = db->md_branch_pages;
6841 arg->ms_leaf_pages = db->md_leaf_pages;
6842 arg->ms_overflow_pages = db->md_overflow_pages;
6843 arg->ms_entries = db->md_entries;
6848 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6852 if (env == NULL || arg == NULL)
6855 toggle = mdb_env_pick_meta(env);
6857 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6861 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
6865 if (env == NULL || arg == NULL)
6868 toggle = mdb_env_pick_meta(env);
6869 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
6870 arg->me_mapsize = env->me_mapsize;
6871 arg->me_maxreaders = env->me_maxreaders;
6872 arg->me_numreaders = env->me_numreaders;
6873 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
6874 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
6878 /** Set the default comparison functions for a database.
6879 * Called immediately after a database is opened to set the defaults.
6880 * The user can then override them with #mdb_set_compare() or
6881 * #mdb_set_dupsort().
6882 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6883 * @param[in] dbi A database handle returned by #mdb_dbi_open()
6886 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6888 uint16_t f = txn->mt_dbs[dbi].md_flags;
6890 txn->mt_dbxs[dbi].md_cmp =
6891 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6892 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6894 txn->mt_dbxs[dbi].md_dcmp =
6895 !(f & MDB_DUPSORT) ? 0 :
6896 ((f & MDB_INTEGERDUP)
6897 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6898 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6901 #define PERSISTENT_FLAGS 0xffff
6902 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
6903 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
6904 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6909 int rc, dbflag, exact;
6910 unsigned int unused = 0;
6913 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6914 mdb_default_cmp(txn, FREE_DBI);
6917 if ((flags & VALID_FLAGS) != flags)
6923 if (flags & PERSISTENT_FLAGS) {
6924 uint16_t f2 = flags & PERSISTENT_FLAGS;
6925 /* make sure flag changes get committed */
6926 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
6927 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
6928 txn->mt_flags |= MDB_TXN_DIRTY;
6931 mdb_default_cmp(txn, MAIN_DBI);
6935 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6936 mdb_default_cmp(txn, MAIN_DBI);
6939 /* Is the DB already open? */
6941 for (i=2; i<txn->mt_numdbs; i++) {
6942 if (!txn->mt_dbxs[i].md_name.mv_size) {
6943 /* Remember this free slot */
6944 if (!unused) unused = i;
6947 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6948 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6954 /* If no free slot and max hit, fail */
6955 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
6956 return MDB_DBS_FULL;
6958 /* Find the DB info */
6962 key.mv_data = (void *)name;
6963 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6964 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6965 if (rc == MDB_SUCCESS) {
6966 /* make sure this is actually a DB */
6967 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6968 if (!(node->mn_flags & F_SUBDATA))
6970 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6971 /* Create if requested */
6973 data.mv_size = sizeof(MDB_db);
6974 data.mv_data = &dummy;
6975 memset(&dummy, 0, sizeof(dummy));
6976 dummy.md_root = P_INVALID;
6977 dummy.md_flags = flags & PERSISTENT_FLAGS;
6978 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6982 /* OK, got info, add to table */
6983 if (rc == MDB_SUCCESS) {
6984 unsigned int slot = unused ? unused : txn->mt_numdbs;
6985 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
6986 txn->mt_dbxs[slot].md_name.mv_size = len;
6987 txn->mt_dbxs[slot].md_rel = NULL;
6988 txn->mt_dbflags[slot] = dbflag;
6989 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
6991 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
6992 mdb_default_cmp(txn, slot);
6995 txn->mt_env->me_numdbs++;
7002 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7004 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7007 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7010 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7013 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
7015 ptr = env->me_dbxs[dbi].md_name.mv_data;
7016 env->me_dbxs[dbi].md_name.mv_data = NULL;
7017 env->me_dbxs[dbi].md_name.mv_size = 0;
7021 /** Add all the DB's pages to the free list.
7022 * @param[in] mc Cursor on the DB to free.
7023 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7024 * @return 0 on success, non-zero on failure.
7027 mdb_drop0(MDB_cursor *mc, int subs)
7031 rc = mdb_page_search(mc, NULL, 0);
7032 if (rc == MDB_SUCCESS) {
7037 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7038 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
7041 mdb_cursor_copy(mc, &mx);
7042 while (mc->mc_snum > 0) {
7043 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7044 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7045 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7046 if (ni->mn_flags & F_SUBDATA) {
7047 mdb_xcursor_init1(mc, ni);
7048 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7054 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7056 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7059 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7064 rc = mdb_cursor_sibling(mc, 1);
7066 /* no more siblings, go back to beginning
7067 * of previous level.
7070 for (i=1; i<mc->mc_top; i++)
7071 mc->mc_pg[i] = mx.mc_pg[i];
7075 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7076 mc->mc_db->md_root);
7081 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7086 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1)
7089 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7092 rc = mdb_cursor_open(txn, dbi, &mc);
7096 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7100 /* Can't delete the main DB */
7101 if (del && dbi > MAIN_DBI) {
7102 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7104 mdb_dbi_close(txn->mt_env, dbi);
7106 /* reset the DB record, mark it dirty */
7107 txn->mt_dbflags[dbi] |= DB_DIRTY;
7108 txn->mt_dbs[dbi].md_depth = 0;
7109 txn->mt_dbs[dbi].md_branch_pages = 0;
7110 txn->mt_dbs[dbi].md_leaf_pages = 0;
7111 txn->mt_dbs[dbi].md_overflow_pages = 0;
7112 txn->mt_dbs[dbi].md_entries = 0;
7113 txn->mt_dbs[dbi].md_root = P_INVALID;
7115 if (!txn->mt_u.dirty_list[0].mid) {
7118 /* make sure we have at least one dirty page in this txn
7119 * otherwise these changes will be ignored.
7121 key.mv_size = sizeof(txnid_t);
7122 key.mv_data = &txn->mt_txnid;
7123 data.mv_size = sizeof(MDB_ID);
7124 data.mv_data = txn->mt_free_pgs;
7125 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
7126 rc = mdb_cursor_put(&m2, &key, &data, 0);
7130 mdb_cursor_close(mc);
7134 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7136 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7139 txn->mt_dbxs[dbi].md_cmp = cmp;
7143 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7145 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7148 txn->mt_dbxs[dbi].md_dcmp = cmp;
7152 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7154 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7157 txn->mt_dbxs[dbi].md_rel = rel;
7161 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7163 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7166 txn->mt_dbxs[dbi].md_relctx = ctx;