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__))
291 # define DPRINTF (void) /* Vararg macros may be unsupported */
293 static int mdb_debug;
294 static txnid_t mdb_debug_start;
296 /** Print a debug message with printf formatting. */
297 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
298 ((void) ((mdb_debug) && \
299 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
301 # define DPRINTF(fmt, ...) ((void) 0)
302 # define MDB_DEBUG_SKIP
304 /** Print a debug string.
305 * The string is printed literally, with no format processing.
307 #define DPUTS(arg) DPRINTF("%s", arg)
310 /** A default memory page size.
311 * The actual size is platform-dependent, but we use this for
312 * boot-strapping. We probably should not be using this any more.
313 * The #GET_PAGESIZE() macro is used to get the actual size.
315 * Note that we don't currently support Huge pages. On Linux,
316 * regular data files cannot use Huge pages, and in general
317 * Huge pages aren't actually pageable. We rely on the OS
318 * demand-pager to read our data and page it out when memory
319 * pressure from other processes is high. So until OSs have
320 * actual paging support for Huge pages, they're not viable.
322 #define MDB_PAGESIZE 4096
324 /** The minimum number of keys required in a database page.
325 * Setting this to a larger value will place a smaller bound on the
326 * maximum size of a data item. Data items larger than this size will
327 * be pushed into overflow pages instead of being stored directly in
328 * the B-tree node. This value used to default to 4. With a page size
329 * of 4096 bytes that meant that any item larger than 1024 bytes would
330 * go into an overflow page. That also meant that on average 2-3KB of
331 * each overflow page was wasted space. The value cannot be lower than
332 * 2 because then there would no longer be a tree structure. With this
333 * value, items larger than 2KB will go into overflow pages, and on
334 * average only 1KB will be wasted.
336 #define MDB_MINKEYS 2
338 /** A stamp that identifies a file as an MDB file.
339 * There's nothing special about this value other than that it is easily
340 * recognizable, and it will reflect any byte order mismatches.
342 #define MDB_MAGIC 0xBEEFC0DE
344 /** The version number for a database's file format. */
345 #define MDB_VERSION 1
347 /** @brief The maximum size of a key in the database.
349 * We require that keys all fit onto a regular page. This limit
350 * could be raised a bit further if needed; to something just
351 * under #MDB_PAGESIZE / #MDB_MINKEYS.
353 * Note that data items in an #MDB_DUPSORT database are actually keys
354 * of a subDB, so they're also limited to this size.
356 #ifndef MDB_MAXKEYSIZE
357 #define MDB_MAXKEYSIZE 511
360 /** @brief The maximum size of a data item.
362 * We only store a 32 bit value for node sizes.
364 #define MAXDATASIZE 0xffffffffUL
369 * This is used for printing a hex dump of a key's contents.
371 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
372 /** Display a key in hex.
374 * Invoke a function to display a key in hex.
376 #define DKEY(x) mdb_dkey(x, kbuf)
378 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
382 /** An invalid page number.
383 * Mainly used to denote an empty tree.
385 #define P_INVALID (~(pgno_t)0)
387 /** Test if the flags \b f are set in a flag word \b w. */
388 #define F_ISSET(w, f) (((w) & (f)) == (f))
390 /** Used for offsets within a single page.
391 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
394 typedef uint16_t indx_t;
396 /** Default size of memory map.
397 * This is certainly too small for any actual applications. Apps should always set
398 * the size explicitly using #mdb_env_set_mapsize().
400 #define DEFAULT_MAPSIZE 1048576
402 /** @defgroup readers Reader Lock Table
403 * Readers don't acquire any locks for their data access. Instead, they
404 * simply record their transaction ID in the reader table. The reader
405 * mutex is needed just to find an empty slot in the reader table. The
406 * slot's address is saved in thread-specific data so that subsequent read
407 * transactions started by the same thread need no further locking to proceed.
409 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
411 * No reader table is used if the database is on a read-only filesystem.
413 * Since the database uses multi-version concurrency control, readers don't
414 * actually need any locking. This table is used to keep track of which
415 * readers are using data from which old transactions, so that we'll know
416 * when a particular old transaction is no longer in use. Old transactions
417 * that have discarded any data pages can then have those pages reclaimed
418 * for use by a later write transaction.
420 * The lock table is constructed such that reader slots are aligned with the
421 * processor's cache line size. Any slot is only ever used by one thread.
422 * This alignment guarantees that there will be no contention or cache
423 * thrashing as threads update their own slot info, and also eliminates
424 * any need for locking when accessing a slot.
426 * A writer thread will scan every slot in the table to determine the oldest
427 * outstanding reader transaction. Any freed pages older than this will be
428 * reclaimed by the writer. The writer doesn't use any locks when scanning
429 * this table. This means that there's no guarantee that the writer will
430 * see the most up-to-date reader info, but that's not required for correct
431 * operation - all we need is to know the upper bound on the oldest reader,
432 * we don't care at all about the newest reader. So the only consequence of
433 * reading stale information here is that old pages might hang around a
434 * while longer before being reclaimed. That's actually good anyway, because
435 * the longer we delay reclaiming old pages, the more likely it is that a
436 * string of contiguous pages can be found after coalescing old pages from
437 * many old transactions together.
440 /** Number of slots in the reader table.
441 * This value was chosen somewhat arbitrarily. 126 readers plus a
442 * couple mutexes fit exactly into 8KB on my development machine.
443 * Applications should set the table size using #mdb_env_set_maxreaders().
445 #define DEFAULT_READERS 126
447 /** The size of a CPU cache line in bytes. We want our lock structures
448 * aligned to this size to avoid false cache line sharing in the
450 * This value works for most CPUs. For Itanium this should be 128.
456 /** The information we store in a single slot of the reader table.
457 * In addition to a transaction ID, we also record the process and
458 * thread ID that owns a slot, so that we can detect stale information,
459 * e.g. threads or processes that went away without cleaning up.
460 * @note We currently don't check for stale records. We simply re-init
461 * the table when we know that we're the only process opening the
464 typedef struct MDB_rxbody {
465 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
466 * Multiple readers that start at the same time will probably have the
467 * same ID here. Again, it's not important to exclude them from
468 * anything; all we need to know is which version of the DB they
469 * started from so we can avoid overwriting any data used in that
470 * particular version.
473 /** The process ID of the process owning this reader txn. */
475 /** The thread ID of the thread owning this txn. */
479 /** The actual reader record, with cacheline padding. */
480 typedef struct MDB_reader {
483 /** shorthand for mrb_txnid */
484 #define mr_txnid mru.mrx.mrb_txnid
485 #define mr_pid mru.mrx.mrb_pid
486 #define mr_tid mru.mrx.mrb_tid
487 /** cache line alignment */
488 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
492 /** The header for the reader table.
493 * The table resides in a memory-mapped file. (This is a different file
494 * than is used for the main database.)
496 * For POSIX the actual mutexes reside in the shared memory of this
497 * mapped file. On Windows, mutexes are named objects allocated by the
498 * kernel; we store the mutex names in this mapped file so that other
499 * processes can grab them. This same approach is also used on
500 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
501 * process-shared POSIX mutexes. For these cases where a named object
502 * is used, the object name is derived from a 64 bit FNV hash of the
503 * environment pathname. As such, naming collisions are extremely
504 * unlikely. If a collision occurs, the results are unpredictable.
506 typedef struct MDB_txbody {
507 /** Stamp identifying this as an MDB file. It must be set
510 /** Version number of this lock file. Must be set to #MDB_VERSION. */
511 uint32_t mtb_version;
512 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
513 char mtb_rmname[MNAME_LEN];
515 /** Mutex protecting access to this table.
516 * This is the reader lock that #LOCK_MUTEX_R acquires.
518 pthread_mutex_t mtb_mutex;
520 /** The ID of the last transaction committed to the database.
521 * This is recorded here only for convenience; the value can always
522 * be determined by reading the main database meta pages.
525 /** The number of slots that have been used in the reader table.
526 * This always records the maximum count, it is not decremented
527 * when readers release their slots.
529 unsigned mtb_numreaders;
532 /** The actual reader table definition. */
533 typedef struct MDB_txninfo {
536 #define mti_magic mt1.mtb.mtb_magic
537 #define mti_version mt1.mtb.mtb_version
538 #define mti_mutex mt1.mtb.mtb_mutex
539 #define mti_rmname mt1.mtb.mtb_rmname
540 #define mti_txnid mt1.mtb.mtb_txnid
541 #define mti_numreaders mt1.mtb.mtb_numreaders
542 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
545 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
546 char mt2_wmname[MNAME_LEN];
547 #define mti_wmname mt2.mt2_wmname
549 pthread_mutex_t mt2_wmutex;
550 #define mti_wmutex mt2.mt2_wmutex
552 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
554 MDB_reader mti_readers[1];
558 /** Common header for all page types.
559 * Overflow records occupy a number of contiguous pages with no
560 * headers on any page after the first.
562 typedef struct MDB_page {
563 #define mp_pgno mp_p.p_pgno
564 #define mp_next mp_p.p_next
566 pgno_t p_pgno; /**< page number */
567 void * p_next; /**< for in-memory list of freed structs */
570 /** @defgroup mdb_page Page Flags
572 * Flags for the page headers.
575 #define P_BRANCH 0x01 /**< branch page */
576 #define P_LEAF 0x02 /**< leaf page */
577 #define P_OVERFLOW 0x04 /**< overflow page */
578 #define P_META 0x08 /**< meta page */
579 #define P_DIRTY 0x10 /**< dirty page */
580 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
581 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
583 uint16_t mp_flags; /**< @ref mdb_page */
584 #define mp_lower mp_pb.pb.pb_lower
585 #define mp_upper mp_pb.pb.pb_upper
586 #define mp_pages mp_pb.pb_pages
589 indx_t pb_lower; /**< lower bound of free space */
590 indx_t pb_upper; /**< upper bound of free space */
592 uint32_t pb_pages; /**< number of overflow pages */
594 indx_t mp_ptrs[1]; /**< dynamic size */
597 /** Size of the page header, excluding dynamic data at the end */
598 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
600 /** Address of first usable data byte in a page, after the header */
601 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
603 /** Number of nodes on a page */
604 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
606 /** The amount of space remaining in the page */
607 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
609 /** The percentage of space used in the page, in tenths of a percent. */
610 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
611 ((env)->me_psize - PAGEHDRSZ))
612 /** The minimum page fill factor, in tenths of a percent.
613 * Pages emptier than this are candidates for merging.
615 #define FILL_THRESHOLD 250
617 /** Test if a page is a leaf page */
618 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
619 /** Test if a page is a LEAF2 page */
620 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
621 /** Test if a page is a branch page */
622 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
623 /** Test if a page is an overflow page */
624 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
625 /** Test if a page is a sub page */
626 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
628 /** The number of overflow pages needed to store the given size. */
629 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
631 /** Header for a single key/data pair within a page.
632 * We guarantee 2-byte alignment for nodes.
634 typedef struct MDB_node {
635 /** lo and hi are used for data size on leaf nodes and for
636 * child pgno on branch nodes. On 64 bit platforms, flags
637 * is also used for pgno. (Branch nodes have no flags).
638 * They are in host byte order in case that lets some
639 * accesses be optimized into a 32-bit word access.
641 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
642 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
643 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
644 /** @defgroup mdb_node Node Flags
646 * Flags for node headers.
649 #define F_BIGDATA 0x01 /**< data put on overflow page */
650 #define F_SUBDATA 0x02 /**< data is a sub-database */
651 #define F_DUPDATA 0x04 /**< data has duplicates */
653 /** valid flags for #mdb_node_add() */
654 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
657 unsigned short mn_flags; /**< @ref mdb_node */
658 unsigned short mn_ksize; /**< key size */
659 char mn_data[1]; /**< key and data are appended here */
662 /** Size of the node header, excluding dynamic data at the end */
663 #define NODESIZE offsetof(MDB_node, mn_data)
665 /** Bit position of top word in page number, for shifting mn_flags */
666 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
668 /** Size of a node in a branch page with a given key.
669 * This is just the node header plus the key, there is no data.
671 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
673 /** Size of a node in a leaf page with a given key and data.
674 * This is node header plus key plus data size.
676 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
678 /** Address of node \b i in page \b p */
679 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
681 /** Address of the key for the node */
682 #define NODEKEY(node) (void *)((node)->mn_data)
684 /** Address of the data for a node */
685 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
687 /** Get the page number pointed to by a branch node */
688 #define NODEPGNO(node) \
689 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
690 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
691 /** Set the page number in a branch node */
692 #define SETPGNO(node,pgno) do { \
693 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
694 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
696 /** Get the size of the data in a leaf node */
697 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
698 /** Set the size of the data for a leaf node */
699 #define SETDSZ(node,size) do { \
700 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
701 /** The size of a key in a node */
702 #define NODEKSZ(node) ((node)->mn_ksize)
704 /** Copy a page number from src to dst */
706 #define COPY_PGNO(dst,src) dst = src
708 #if SIZE_MAX > 4294967295UL
709 #define COPY_PGNO(dst,src) do { \
710 unsigned short *s, *d; \
711 s = (unsigned short *)&(src); \
712 d = (unsigned short *)&(dst); \
719 #define COPY_PGNO(dst,src) do { \
720 unsigned short *s, *d; \
721 s = (unsigned short *)&(src); \
722 d = (unsigned short *)&(dst); \
728 /** The address of a key in a LEAF2 page.
729 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
730 * There are no node headers, keys are stored contiguously.
732 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
734 /** Set the \b node's key into \b key, if requested. */
735 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
736 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
738 /** Information about a single database in the environment. */
739 typedef struct MDB_db {
740 uint32_t md_pad; /**< also ksize for LEAF2 pages */
741 uint16_t md_flags; /**< @ref mdb_dbi_open */
742 uint16_t md_depth; /**< depth of this tree */
743 pgno_t md_branch_pages; /**< number of internal pages */
744 pgno_t md_leaf_pages; /**< number of leaf pages */
745 pgno_t md_overflow_pages; /**< number of overflow pages */
746 size_t md_entries; /**< number of data items */
747 pgno_t md_root; /**< the root page of this tree */
750 /** mdb_dbi_open flags */
751 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
752 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
753 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
754 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
756 /** Handle for the DB used to track free pages. */
758 /** Handle for the default DB. */
761 /** Meta page content. */
762 typedef struct MDB_meta {
763 /** Stamp identifying this as an MDB file. It must be set
766 /** Version number of this lock file. Must be set to #MDB_VERSION. */
768 void *mm_address; /**< address for fixed mapping */
769 size_t mm_mapsize; /**< size of mmap region */
770 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
771 /** The size of pages used in this DB */
772 #define mm_psize mm_dbs[0].md_pad
773 /** Any persistent environment flags. @ref mdb_env */
774 #define mm_flags mm_dbs[0].md_flags
775 pgno_t mm_last_pg; /**< last used page in file */
776 txnid_t mm_txnid; /**< txnid that committed this page */
779 /** Buffer for a stack-allocated dirty page.
780 * The members define size and alignment, and silence type
781 * aliasing warnings. They are not used directly; that could
782 * mean incorrectly using several union members in parallel.
784 typedef union MDB_pagebuf {
785 char mb_raw[MDB_PAGESIZE];
788 char mm_pad[PAGEHDRSZ];
793 /** Auxiliary DB info.
794 * The information here is mostly static/read-only. There is
795 * only a single copy of this record in the environment.
797 typedef struct MDB_dbx {
798 MDB_val md_name; /**< name of the database */
799 MDB_cmp_func *md_cmp; /**< function for comparing keys */
800 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
801 MDB_rel_func *md_rel; /**< user relocate function */
802 void *md_relctx; /**< user-provided context for md_rel */
805 /** A database transaction.
806 * Every operation requires a transaction handle.
809 MDB_txn *mt_parent; /**< parent of a nested txn */
810 MDB_txn *mt_child; /**< nested txn under this txn */
811 pgno_t mt_next_pgno; /**< next unallocated page */
812 /** The ID of this transaction. IDs are integers incrementing from 1.
813 * Only committed write transactions increment the ID. If a transaction
814 * aborts, the ID may be re-used by the next writer.
817 MDB_env *mt_env; /**< the DB environment */
818 /** The list of pages that became unused during this transaction.
822 MDB_ID2L dirty_list; /**< for write txns: modified pages */
823 MDB_reader *reader; /**< this thread's reader table slot or NULL */
825 /** Array of records for each DB known in the environment. */
827 /** Array of MDB_db records for each known DB */
829 /** @defgroup mt_dbflag Transaction DB Flags
833 #define DB_DIRTY 0x01 /**< DB was written in this txn */
834 #define DB_STALE 0x02 /**< DB record is older than txnID */
835 #define DB_NEW 0x04 /**< DB handle opened in this txn */
836 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
838 /** In write txns, array of cursors for each DB */
839 MDB_cursor **mt_cursors;
840 /** Array of flags for each DB */
841 unsigned char *mt_dbflags;
842 /** Number of DB records in use. This number only ever increments;
843 * we don't decrement it when individual DB handles are closed.
847 /** @defgroup mdb_txn Transaction Flags
851 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
852 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
853 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
855 unsigned int mt_flags; /**< @ref mdb_txn */
856 /** dirty_list maxsize - #allocated pages including in parent txns */
857 unsigned int mt_dirty_room;
858 /** Tracks which of the two meta pages was used at the start
859 * of this transaction.
861 unsigned int mt_toggle;
864 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
865 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
866 * raise this on a 64 bit machine.
868 #define CURSOR_STACK 32
872 /** Cursors are used for all DB operations */
874 /** Next cursor on this DB in this txn */
876 /** Original cursor if this is a shadow */
878 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
879 struct MDB_xcursor *mc_xcursor;
880 /** The transaction that owns this cursor */
882 /** The database handle this cursor operates on */
884 /** The database record for this cursor */
886 /** The database auxiliary record for this cursor */
888 /** The @ref mt_dbflag for this database */
889 unsigned char *mc_dbflag;
890 unsigned short mc_snum; /**< number of pushed pages */
891 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
892 /** @defgroup mdb_cursor Cursor Flags
894 * Cursor state flags.
897 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
898 #define C_EOF 0x02 /**< No more data */
899 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
900 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
901 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
902 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
904 unsigned int mc_flags; /**< @ref mdb_cursor */
905 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
906 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
909 /** Context for sorted-dup records.
910 * We could have gone to a fully recursive design, with arbitrarily
911 * deep nesting of sub-databases. But for now we only handle these
912 * levels - main DB, optional sub-DB, sorted-duplicate DB.
914 typedef struct MDB_xcursor {
915 /** A sub-cursor for traversing the Dup DB */
916 MDB_cursor mx_cursor;
917 /** The database record for this Dup DB */
919 /** The auxiliary DB record for this Dup DB */
921 /** The @ref mt_dbflag for this Dup DB */
922 unsigned char mx_dbflag;
925 /** State of FreeDB old pages, stored in the MDB_env */
926 typedef struct MDB_pgstate {
927 txnid_t mf_pglast; /**< ID of last old page record we used */
928 pgno_t *mf_pghead; /**< old pages reclaimed from freelist */
929 pgno_t *mf_pgfree; /**< memory to free when dropping me_pghead */
932 /** The database environment. */
934 HANDLE me_fd; /**< The main data file */
935 HANDLE me_lfd; /**< The lock file */
936 HANDLE me_mfd; /**< just for writing the meta pages */
937 /** Failed to update the meta page. Probably an I/O error. */
938 #define MDB_FATAL_ERROR 0x80000000U
939 /** Some fields are initialized. */
940 #define MDB_ENV_ACTIVE 0x20000000U
941 /** me_txkey is set */
942 #define MDB_ENV_TXKEY 0x10000000U
943 uint32_t me_flags; /**< @ref mdb_env */
944 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
945 unsigned int me_maxreaders; /**< size of the reader table */
946 unsigned int me_numreaders; /**< max numreaders set by this env */
947 MDB_dbi me_numdbs; /**< number of DBs opened */
948 MDB_dbi me_maxdbs; /**< size of the DB table */
949 pid_t me_pid; /**< process ID of this env */
950 char *me_path; /**< path to the DB files */
951 char *me_map; /**< the memory map of the data file */
952 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
953 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
954 MDB_txn *me_txn; /**< current write transaction */
955 size_t me_mapsize; /**< size of the data memory map */
956 off_t me_size; /**< current file size */
957 pgno_t me_maxpg; /**< me_mapsize / me_psize */
958 MDB_dbx *me_dbxs; /**< array of static DB info */
959 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
960 pthread_key_t me_txkey; /**< thread-key for readers */
961 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
962 # define me_pglast me_pgstate.mf_pglast
963 # define me_pghead me_pgstate.mf_pghead
964 # define me_pgfree me_pgstate.mf_pgfree
965 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
966 /** IDL of pages that became unused in a write txn */
968 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
969 MDB_ID2L me_dirty_list;
970 /** Max number of freelist items that can fit in a single overflow page */
971 unsigned int me_maxfree_1pg;
972 /** Max size of a node on a page */
973 unsigned int me_nodemax;
975 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
977 #elif defined(MDB_USE_POSIX_SEM)
978 sem_t *me_rmutex; /* Shared mutexes are not supported */
983 /** Nested transaction */
984 typedef struct MDB_ntxn {
985 MDB_txn mnt_txn; /* the transaction */
986 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
989 /** max number of pages to commit in one writev() call */
990 #define MDB_COMMIT_PAGES 64
991 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
992 #undef MDB_COMMIT_PAGES
993 #define MDB_COMMIT_PAGES IOV_MAX
996 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
997 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
998 static int mdb_page_touch(MDB_cursor *mc);
1000 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
1001 static int mdb_page_search_root(MDB_cursor *mc,
1002 MDB_val *key, int modify);
1003 #define MDB_PS_MODIFY 1
1004 #define MDB_PS_ROOTONLY 2
1005 static int mdb_page_search(MDB_cursor *mc,
1006 MDB_val *key, int flags);
1007 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1009 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1010 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1011 pgno_t newpgno, unsigned int nflags);
1013 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1014 static int mdb_env_pick_meta(const MDB_env *env);
1015 static int mdb_env_write_meta(MDB_txn *txn);
1016 static void mdb_env_close0(MDB_env *env, int excl);
1018 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1019 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1020 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1021 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1022 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1023 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1024 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1025 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1026 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1028 static int mdb_rebalance(MDB_cursor *mc);
1029 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1031 static void mdb_cursor_pop(MDB_cursor *mc);
1032 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1034 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1035 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1036 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1037 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1038 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1040 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1041 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1043 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1044 static void mdb_xcursor_init0(MDB_cursor *mc);
1045 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1047 static int mdb_drop0(MDB_cursor *mc, int subs);
1048 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1051 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1055 static SECURITY_DESCRIPTOR mdb_null_sd;
1056 static SECURITY_ATTRIBUTES mdb_all_sa;
1057 static int mdb_sec_inited;
1060 /** Return the library version info. */
1062 mdb_version(int *major, int *minor, int *patch)
1064 if (major) *major = MDB_VERSION_MAJOR;
1065 if (minor) *minor = MDB_VERSION_MINOR;
1066 if (patch) *patch = MDB_VERSION_PATCH;
1067 return MDB_VERSION_STRING;
1070 /** Table of descriptions for MDB @ref errors */
1071 static char *const mdb_errstr[] = {
1072 "MDB_KEYEXIST: Key/data pair already exists",
1073 "MDB_NOTFOUND: No matching key/data pair found",
1074 "MDB_PAGE_NOTFOUND: Requested page not found",
1075 "MDB_CORRUPTED: Located page was wrong type",
1076 "MDB_PANIC: Update of meta page failed",
1077 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1078 "MDB_INVALID: File is not an MDB file",
1079 "MDB_MAP_FULL: Environment mapsize limit reached",
1080 "MDB_DBS_FULL: Environment maxdbs limit reached",
1081 "MDB_READERS_FULL: Environment maxreaders limit reached",
1082 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1083 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1084 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1085 "MDB_PAGE_FULL: Internal error - page has no more space",
1086 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1087 "MDB_INCOMPATIBLE: Database flags changed or would change",
1088 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1092 mdb_strerror(int err)
1096 return ("Successful return: 0");
1098 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1099 i = err - MDB_KEYEXIST;
1100 return mdb_errstr[i];
1103 return strerror(err);
1107 /** Display a key in hexadecimal and return the address of the result.
1108 * @param[in] key the key to display
1109 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1110 * @return The key in hexadecimal form.
1113 mdb_dkey(MDB_val *key, char *buf)
1116 unsigned char *c = key->mv_data;
1122 if (key->mv_size > MDB_MAXKEYSIZE)
1123 return "MDB_MAXKEYSIZE";
1124 /* may want to make this a dynamic check: if the key is mostly
1125 * printable characters, print it as-is instead of converting to hex.
1129 for (i=0; i<key->mv_size; i++)
1130 ptr += sprintf(ptr, "%02x", *c++);
1132 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1137 /** Display all the keys in the page. */
1139 mdb_page_list(MDB_page *mp)
1142 unsigned int i, nkeys, nsize;
1146 nkeys = NUMKEYS(mp);
1147 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1148 for (i=0; i<nkeys; i++) {
1149 node = NODEPTR(mp, i);
1150 key.mv_size = node->mn_ksize;
1151 key.mv_data = node->mn_data;
1152 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1153 if (IS_BRANCH(mp)) {
1154 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1157 if (F_ISSET(node->mn_flags, F_BIGDATA))
1158 nsize += sizeof(pgno_t);
1160 nsize += NODEDSZ(node);
1161 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1167 mdb_cursor_chk(MDB_cursor *mc)
1173 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1174 for (i=0; i<mc->mc_top; i++) {
1176 node = NODEPTR(mp, mc->mc_ki[i]);
1177 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1180 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1186 /** Count all the pages in each DB and in the freelist
1187 * and make sure it matches the actual number of pages
1190 static void mdb_audit(MDB_txn *txn)
1194 MDB_ID freecount, count;
1199 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1200 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1201 freecount += *(MDB_ID *)data.mv_data;
1204 for (i = 0; i<txn->mt_numdbs; i++) {
1205 MDB_xcursor mx, *mxp;
1206 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1207 mdb_cursor_init(&mc, txn, i, mxp);
1208 if (txn->mt_dbs[i].md_root == P_INVALID)
1210 count += txn->mt_dbs[i].md_branch_pages +
1211 txn->mt_dbs[i].md_leaf_pages +
1212 txn->mt_dbs[i].md_overflow_pages;
1213 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1214 mdb_page_search(&mc, NULL, 0);
1218 mp = mc.mc_pg[mc.mc_top];
1219 for (j=0; j<NUMKEYS(mp); j++) {
1220 MDB_node *leaf = NODEPTR(mp, j);
1221 if (leaf->mn_flags & F_SUBDATA) {
1223 memcpy(&db, NODEDATA(leaf), sizeof(db));
1224 count += db.md_branch_pages + db.md_leaf_pages +
1225 db.md_overflow_pages;
1229 while (mdb_cursor_sibling(&mc, 1) == 0);
1232 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1233 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1234 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1240 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1242 return txn->mt_dbxs[dbi].md_cmp(a, b);
1246 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1248 if (txn->mt_dbxs[dbi].md_dcmp)
1249 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1251 return EINVAL; /* too bad you can't distinguish this from a valid result */
1254 /** Allocate a single page.
1255 * Re-use old malloc'd pages first, otherwise just malloc.
1258 mdb_page_malloc(MDB_cursor *mc) {
1260 size_t sz = mc->mc_txn->mt_env->me_psize;
1261 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1262 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1263 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1264 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1265 } else if ((ret = malloc(sz)) != NULL) {
1266 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1272 mdb_page_free(MDB_env *env, MDB_page *mp)
1274 mp->mp_next = env->me_dpages;
1275 VGMEMP_FREE(env, mp);
1276 env->me_dpages = mp;
1279 /** Allocate pages for writing.
1280 * If there are free pages available from older transactions, they
1281 * will be re-used first. Otherwise a new page will be allocated.
1282 * @param[in] mc cursor A cursor handle identifying the transaction and
1283 * database for which we are allocating.
1284 * @param[in] num the number of pages to allocate.
1285 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1286 * will always be satisfied by a single contiguous chunk of memory.
1287 * @return 0 on success, non-zero on failure.
1290 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1292 MDB_txn *txn = mc->mc_txn;
1294 pgno_t pgno = P_INVALID;
1296 txnid_t oldest = 0, last;
1301 /* If our dirty list is already full, we can't do anything */
1302 if (txn->mt_dirty_room == 0)
1303 return MDB_TXN_FULL;
1305 /* The free list won't have any content at all until txn 2 has
1306 * committed. The pages freed by txn 2 will be unreferenced
1307 * after txn 3 commits, and so will be safe to re-use in txn 4.
1309 if (txn->mt_txnid > 3) {
1310 if (!txn->mt_env->me_pghead &&
1311 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1312 /* See if there's anything in the free DB */
1319 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1320 if (!txn->mt_env->me_pglast) {
1321 mdb_page_search(&m2, NULL, 0);
1322 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1323 kptr = (txnid_t *)NODEKEY(leaf);
1328 last = txn->mt_env->me_pglast + 1;
1330 key.mv_data = &last;
1331 key.mv_size = sizeof(last);
1332 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET_RANGE, NULL);
1335 last = *(txnid_t *)key.mv_data;
1341 oldest = txn->mt_txnid - 1;
1342 nr = txn->mt_env->me_txns->mti_numreaders;
1343 r = txn->mt_env->me_txns->mti_readers;
1344 for (i=0; i<nr; i++) {
1345 if (!r[i].mr_pid) continue;
1352 if (oldest > last) {
1353 /* It's usable, grab it.
1357 if (!txn->mt_env->me_pglast) {
1358 mdb_node_read(txn, leaf, &data);
1360 idl = (MDB_ID *) data.mv_data;
1361 /* We might have a zero-length IDL due to freelist growth
1362 * during a prior commit
1365 txn->mt_env->me_pglast = last;
1368 mop = malloc(MDB_IDL_SIZEOF(idl));
1371 txn->mt_env->me_pglast = last;
1372 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop;
1373 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1378 DPRINTF("IDL read txn %zu root %zu num %zu",
1379 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1380 for (i=0; i<idl[0]; i++) {
1381 DPRINTF("IDL %zu", idl[i+1]);
1388 if (txn->mt_env->me_pghead) {
1389 pgno_t *mop = txn->mt_env->me_pghead;
1392 int retry = 1, readit = 0, n2 = num-1;
1393 unsigned int i, j, k;
1395 /* If current list is too short, must fetch more and coalesce */
1396 if (mop[0] < (unsigned)num)
1399 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1401 /* If on freelist, don't try to read more. If what we have
1402 * right now isn't enough just use new pages.
1403 * TODO: get all of this working. Many circular dependencies...
1405 if (mc->mc_dbi == FREE_DBI) {
1413 last = txn->mt_env->me_pglast + 1;
1415 /* We haven't hit the readers list yet? */
1421 oldest = txn->mt_txnid - 1;
1422 nr = txn->mt_env->me_txns->mti_numreaders;
1423 r = txn->mt_env->me_txns->mti_readers;
1424 for (i=0; i<nr; i++) {
1425 if (!r[i].mr_pid) continue;
1432 /* There's nothing we can use on the freelist */
1433 if (oldest - last < 1)
1436 key.mv_data = &last;
1437 key.mv_size = sizeof(last);
1438 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1440 if (rc == MDB_NOTFOUND)
1444 last = *(txnid_t*)key.mv_data;
1447 idl = (MDB_ID *) data.mv_data;
1448 mop2 = malloc(MDB_IDL_SIZEOF(idl) + MDB_IDL_SIZEOF(mop));
1451 /* merge in sorted order */
1452 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1454 while (i>0 || j>0) {
1455 if (i && idl[i] < mop[j])
1456 mop2[k--] = idl[i--];
1458 mop2[k--] = mop[j--];
1460 txn->mt_env->me_pglast = last;
1461 free(txn->mt_env->me_pgfree);
1462 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop2;
1464 /* Keep trying to read until we have enough */
1465 if (mop[0] < (unsigned)num) {
1470 /* current list has enough pages, but are they contiguous? */
1471 for (i=mop[0]; i>=(unsigned)num; i--) {
1472 if (mop[i-n2] == mop[i] + n2) {
1475 /* move any stragglers down */
1476 for (j=i+num; j<=mop[0]; j++)
1483 /* Stop if we succeeded, or no retries */
1484 if (!retry || pgno != P_INVALID)
1490 /* peel pages off tail, so we only have to truncate the list */
1491 pgno = MDB_IDL_LAST(mop);
1494 if (MDB_IDL_IS_ZERO(mop)) {
1495 free(txn->mt_env->me_pgfree);
1496 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1501 if (pgno == P_INVALID) {
1502 /* DB size is maxed out */
1503 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1504 DPUTS("DB size maxed out");
1505 return MDB_MAP_FULL;
1508 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1509 if (pgno == P_INVALID) {
1510 pgno = txn->mt_next_pgno;
1511 txn->mt_next_pgno += num;
1513 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1516 if (txn->mt_env->me_dpages && num == 1) {
1517 np = txn->mt_env->me_dpages;
1518 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1519 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1520 txn->mt_env->me_dpages = np->mp_next;
1522 size_t sz = txn->mt_env->me_psize * num;
1523 if ((np = malloc(sz)) == NULL)
1525 VGMEMP_ALLOC(txn->mt_env, np, sz);
1527 if (pgno == P_INVALID) {
1528 np->mp_pgno = txn->mt_next_pgno;
1529 txn->mt_next_pgno += num;
1534 mid.mid = np->mp_pgno;
1536 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1537 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1539 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1541 txn->mt_dirty_room--;
1547 /** Copy a page: avoid copying unused portions of the page.
1548 * @param[in] dst page to copy into
1549 * @param[in] src page to copy from
1552 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1554 dst->mp_flags = src->mp_flags | P_DIRTY;
1555 dst->mp_pages = src->mp_pages;
1557 if (IS_LEAF2(src)) {
1558 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1560 unsigned int i, nkeys = NUMKEYS(src);
1561 for (i=0; i<nkeys; i++)
1562 dst->mp_ptrs[i] = src->mp_ptrs[i];
1563 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1564 psize - src->mp_upper);
1568 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1569 * @param[in] mc cursor pointing to the page to be touched
1570 * @return 0 on success, non-zero on failure.
1573 mdb_page_touch(MDB_cursor *mc)
1575 MDB_page *mp = mc->mc_pg[mc->mc_top];
1579 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1581 if ((rc = mdb_page_alloc(mc, 1, &np)))
1583 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1584 assert(mp->mp_pgno != np->mp_pgno);
1585 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1587 /* If page isn't full, just copy the used portion */
1588 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1591 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1593 np->mp_flags |= P_DIRTY;
1598 /* Adjust other cursors pointing to mp */
1599 if (mc->mc_flags & C_SUB) {
1600 MDB_cursor *m2, *m3;
1601 MDB_dbi dbi = mc->mc_dbi-1;
1603 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1604 if (m2 == mc) continue;
1605 m3 = &m2->mc_xcursor->mx_cursor;
1606 if (m3->mc_snum < mc->mc_snum) continue;
1607 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1608 m3->mc_pg[mc->mc_top] = mp;
1614 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1615 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1616 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1617 m2->mc_pg[mc->mc_top] = mp;
1621 mc->mc_pg[mc->mc_top] = mp;
1622 /** If this page has a parent, update the parent to point to
1626 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1628 mc->mc_db->md_root = mp->mp_pgno;
1629 } else if (mc->mc_txn->mt_parent) {
1632 /* If txn has a parent, make sure the page is in our
1635 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1636 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1637 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1638 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1639 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1640 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1641 mc->mc_pg[mc->mc_top] = mp;
1646 assert(mc->mc_txn->mt_u.dirty_list[0].mid < MDB_IDL_UM_MAX);
1648 np = mdb_page_malloc(mc);
1651 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1652 mid.mid = np->mp_pgno;
1654 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1662 mdb_env_sync(MDB_env *env, int force)
1665 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1666 if (env->me_flags & MDB_WRITEMAP) {
1667 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1668 ? MS_ASYNC : MS_SYNC;
1669 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1672 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1676 if (MDB_FDATASYNC(env->me_fd))
1683 /** Make shadow copies of all of parent txn's cursors */
1685 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1687 MDB_cursor *mc, *m2;
1688 unsigned int i, j, size;
1690 for (i=0;i<src->mt_numdbs; i++) {
1691 if (src->mt_cursors[i]) {
1692 size = sizeof(MDB_cursor);
1693 if (src->mt_cursors[i]->mc_xcursor)
1694 size += sizeof(MDB_xcursor);
1695 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1702 mc->mc_db = &dst->mt_dbs[i];
1703 mc->mc_dbx = m2->mc_dbx;
1704 mc->mc_dbflag = &dst->mt_dbflags[i];
1705 mc->mc_snum = m2->mc_snum;
1706 mc->mc_top = m2->mc_top;
1707 mc->mc_flags = m2->mc_flags | C_SHADOW;
1708 for (j=0; j<mc->mc_snum; j++) {
1709 mc->mc_pg[j] = m2->mc_pg[j];
1710 mc->mc_ki[j] = m2->mc_ki[j];
1712 if (m2->mc_xcursor) {
1713 MDB_xcursor *mx, *mx2;
1714 mx = (MDB_xcursor *)(mc+1);
1715 mc->mc_xcursor = mx;
1716 mx2 = m2->mc_xcursor;
1717 mx->mx_db = mx2->mx_db;
1718 mx->mx_dbx = mx2->mx_dbx;
1719 mx->mx_dbflag = mx2->mx_dbflag;
1720 mx->mx_cursor.mc_txn = dst;
1721 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1722 mx->mx_cursor.mc_db = &mx->mx_db;
1723 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1724 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1725 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1726 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1727 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1728 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1729 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1730 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1733 mc->mc_xcursor = NULL;
1735 mc->mc_next = dst->mt_cursors[i];
1736 dst->mt_cursors[i] = mc;
1743 /** Merge shadow cursors back into parent's */
1745 mdb_cursor_merge(MDB_txn *txn)
1748 for (i=0; i<txn->mt_numdbs; i++) {
1749 if (txn->mt_cursors[i]) {
1751 while ((mc = txn->mt_cursors[i])) {
1752 txn->mt_cursors[i] = mc->mc_next;
1753 if (mc->mc_flags & C_SHADOW) {
1754 MDB_cursor *m2 = mc->mc_orig;
1756 m2->mc_snum = mc->mc_snum;
1757 m2->mc_top = mc->mc_top;
1758 for (j=0; j<mc->mc_snum; j++) {
1759 m2->mc_pg[j] = mc->mc_pg[j];
1760 m2->mc_ki[j] = mc->mc_ki[j];
1763 if (mc->mc_flags & C_ALLOCD)
1771 mdb_txn_reset0(MDB_txn *txn);
1773 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1774 * @param[in] txn the transaction handle to initialize
1775 * @return 0 on success, non-zero on failure.
1778 mdb_txn_renew0(MDB_txn *txn)
1780 MDB_env *env = txn->mt_env;
1786 txn->mt_numdbs = env->me_numdbs;
1787 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1789 if (txn->mt_flags & MDB_TXN_RDONLY) {
1790 if (!env->me_txns) {
1791 i = mdb_env_pick_meta(env);
1792 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1793 txn->mt_u.reader = NULL;
1795 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1796 pthread_getspecific(env->me_txkey);
1798 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1799 return MDB_BAD_RSLOT;
1801 pid_t pid = env->me_pid;
1802 pthread_t tid = pthread_self();
1805 for (i=0; i<env->me_txns->mti_numreaders; i++)
1806 if (env->me_txns->mti_readers[i].mr_pid == 0)
1808 if (i == env->me_maxreaders) {
1809 UNLOCK_MUTEX_R(env);
1810 return MDB_READERS_FULL;
1812 env->me_txns->mti_readers[i].mr_pid = pid;
1813 env->me_txns->mti_readers[i].mr_tid = tid;
1814 if (i >= env->me_txns->mti_numreaders)
1815 env->me_txns->mti_numreaders = i+1;
1816 /* Save numreaders for un-mutexed mdb_env_close() */
1817 env->me_numreaders = env->me_txns->mti_numreaders;
1818 UNLOCK_MUTEX_R(env);
1819 r = &env->me_txns->mti_readers[i];
1820 if (!(env->me_flags & MDB_NOTLS) &&
1821 (rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1822 env->me_txns->mti_readers[i].mr_pid = 0;
1826 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1827 txn->mt_u.reader = r;
1829 txn->mt_toggle = txn->mt_txnid & 1;
1830 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1834 txn->mt_txnid = env->me_txns->mti_txnid;
1835 txn->mt_toggle = txn->mt_txnid & 1;
1836 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1839 if (txn->mt_txnid == mdb_debug_start)
1842 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1843 txn->mt_u.dirty_list = env->me_dirty_list;
1844 txn->mt_u.dirty_list[0].mid = 0;
1845 txn->mt_free_pgs = env->me_free_pgs;
1846 txn->mt_free_pgs[0] = 0;
1850 /* Copy the DB info and flags */
1851 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1852 for (i=2; i<txn->mt_numdbs; i++) {
1853 x = env->me_dbflags[i];
1854 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1855 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1857 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1859 if (env->me_maxpg < txn->mt_next_pgno) {
1860 mdb_txn_reset0(txn);
1861 return MDB_MAP_RESIZED;
1868 mdb_txn_renew(MDB_txn *txn)
1872 if (!txn || txn->mt_numdbs || !(txn->mt_flags & MDB_TXN_RDONLY))
1875 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1876 DPUTS("environment had fatal error, must shutdown!");
1880 rc = mdb_txn_renew0(txn);
1881 if (rc == MDB_SUCCESS) {
1882 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1883 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1884 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1890 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1894 int rc, size, tsize = sizeof(MDB_txn);
1896 if (env->me_flags & MDB_FATAL_ERROR) {
1897 DPUTS("environment had fatal error, must shutdown!");
1900 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1903 /* Nested transactions: Max 1 child, write txns only, no writemap */
1904 if (parent->mt_child ||
1905 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1906 (env->me_flags & MDB_WRITEMAP))
1910 tsize = sizeof(MDB_ntxn);
1912 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1913 if (!(flags & MDB_RDONLY))
1914 size += env->me_maxdbs * sizeof(MDB_cursor *);
1916 if ((txn = calloc(1, size)) == NULL) {
1917 DPRINTF("calloc: %s", strerror(ErrCode()));
1920 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1921 if (flags & MDB_RDONLY) {
1922 txn->mt_flags |= MDB_TXN_RDONLY;
1923 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1925 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1926 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1932 txn->mt_free_pgs = mdb_midl_alloc();
1933 if (!txn->mt_free_pgs) {
1937 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1938 if (!txn->mt_u.dirty_list) {
1939 free(txn->mt_free_pgs);
1943 txn->mt_txnid = parent->mt_txnid;
1944 txn->mt_toggle = parent->mt_toggle;
1945 txn->mt_dirty_room = parent->mt_dirty_room;
1946 txn->mt_u.dirty_list[0].mid = 0;
1947 txn->mt_free_pgs[0] = 0;
1948 txn->mt_next_pgno = parent->mt_next_pgno;
1949 parent->mt_child = txn;
1950 txn->mt_parent = parent;
1951 txn->mt_numdbs = parent->mt_numdbs;
1952 txn->mt_dbxs = parent->mt_dbxs;
1953 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1954 /* Copy parent's mt_dbflags, but clear DB_NEW */
1955 for (i=0; i<txn->mt_numdbs; i++)
1956 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1958 ntxn = (MDB_ntxn *)txn;
1959 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1960 if (env->me_pghead) {
1961 size = MDB_IDL_SIZEOF(env->me_pghead);
1962 env->me_pghead = malloc(size);
1964 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1968 env->me_pgfree = env->me_pghead;
1970 rc = mdb_cursor_shadow(parent, txn);
1972 mdb_txn_reset0(txn);
1974 rc = mdb_txn_renew0(txn);
1980 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1981 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1982 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1988 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1989 * May be called twice for readonly txns: First reset it, then abort.
1990 * @param[in] txn the transaction handle to reset
1993 mdb_txn_reset0(MDB_txn *txn)
1995 MDB_env *env = txn->mt_env;
1998 /* Close any DBI handles opened in this txn */
1999 for (i=2; i<txn->mt_numdbs; i++) {
2000 if (txn->mt_dbflags[i] & DB_NEW) {
2001 char *ptr = env->me_dbxs[i].md_name.mv_data;
2002 env->me_dbxs[i].md_name.mv_data = NULL;
2003 env->me_dbxs[i].md_name.mv_size = 0;
2008 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2009 if (txn->mt_u.reader) {
2010 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2011 if (!(env->me_flags & MDB_NOTLS))
2012 txn->mt_u.reader = NULL; /* txn does not own reader */
2014 txn->mt_numdbs = 0; /* mark txn as reset, do not close DBs again */
2018 /* close(free) all cursors */
2019 for (i=0; i<txn->mt_numdbs; i++) {
2020 if (txn->mt_cursors[i]) {
2022 while ((mc = txn->mt_cursors[i])) {
2023 txn->mt_cursors[i] = mc->mc_next;
2024 if (mc->mc_flags & C_ALLOCD)
2030 if (!(env->me_flags & MDB_WRITEMAP)) {
2031 /* return all dirty pages to dpage list */
2032 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2033 dp = txn->mt_u.dirty_list[i].mptr;
2034 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2035 mdb_page_free(txn->mt_env, dp);
2037 /* large pages just get freed directly */
2038 VGMEMP_FREE(txn->mt_env, dp);
2044 free(env->me_pgfree);
2046 if (txn->mt_parent) {
2047 txn->mt_parent->mt_child = NULL;
2048 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2049 mdb_midl_free(txn->mt_free_pgs);
2050 free(txn->mt_u.dirty_list);
2053 if (mdb_midl_shrink(&txn->mt_free_pgs))
2054 env->me_free_pgs = txn->mt_free_pgs;
2057 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
2058 txn->mt_env->me_pglast = 0;
2061 /* The writer mutex was locked in mdb_txn_begin. */
2062 UNLOCK_MUTEX_W(env);
2067 mdb_txn_reset(MDB_txn *txn)
2072 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2073 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2074 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2076 /* This call is only valid for read-only txns */
2077 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2080 mdb_txn_reset0(txn);
2084 mdb_txn_abort(MDB_txn *txn)
2089 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2090 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2091 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2094 mdb_txn_abort(txn->mt_child);
2096 mdb_txn_reset0(txn);
2097 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2098 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2099 txn->mt_u.reader->mr_pid = 0;
2105 mdb_txn_commit(MDB_txn *txn)
2113 pgno_t next, freecnt;
2114 txnid_t oldpg_txnid, id;
2117 assert(txn != NULL);
2118 assert(txn->mt_env != NULL);
2120 if (txn->mt_child) {
2121 mdb_txn_commit(txn->mt_child);
2122 txn->mt_child = NULL;
2127 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2128 /* update the DB flags */
2129 for (i = 2; i<txn->mt_numdbs; i++) {
2130 if (txn->mt_dbflags[i] & DB_NEW)
2131 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2133 if (txn->mt_numdbs > env->me_numdbs)
2134 env->me_numdbs = txn->mt_numdbs;
2135 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2140 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2141 DPUTS("error flag is set, can't commit");
2143 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2148 if (txn->mt_parent) {
2149 MDB_txn *parent = txn->mt_parent;
2153 /* Append our free list to parent's */
2154 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2158 mdb_midl_free(txn->mt_free_pgs);
2160 parent->mt_next_pgno = txn->mt_next_pgno;
2161 parent->mt_flags = txn->mt_flags;
2163 /* Merge (and close) our cursors with parent's */
2164 mdb_cursor_merge(txn);
2166 /* Update parent's DB table. */
2167 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2168 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2169 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2170 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2171 for (i=2; i<txn->mt_numdbs; i++) {
2172 /* preserve parent's DB_NEW status */
2173 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2174 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2177 dst = txn->mt_parent->mt_u.dirty_list;
2178 src = txn->mt_u.dirty_list;
2179 /* Find len = length of merging our dirty list with parent's */
2181 dst[0].mid = 0; /* simplify loops */
2182 if (parent->mt_parent) {
2183 len = x + src[0].mid;
2184 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2185 for (i = x; y && i; y--) {
2186 pgno_t yp = src[y].mid;
2187 while (yp < dst[i].mid)
2189 if (yp == dst[i].mid) {
2194 } else { /* Simplify the above for single-ancestor case */
2195 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2197 /* Merge our dirty list with parent's */
2199 for (i = len; y; dst[i--] = src[y--]) {
2200 pgno_t yp = src[y].mid;
2201 while (yp < dst[x].mid)
2202 dst[i--] = dst[x--];
2203 if (yp == dst[x].mid)
2204 free(dst[x--].mptr);
2208 free(txn->mt_u.dirty_list);
2209 parent->mt_dirty_room = txn->mt_dirty_room;
2211 txn->mt_parent->mt_child = NULL;
2212 free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pgfree);
2217 if (txn != env->me_txn) {
2218 DPUTS("attempt to commit unknown transaction");
2223 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2226 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2227 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2229 /* Update DB root pointers */
2230 if (txn->mt_numdbs > 2) {
2233 data.mv_size = sizeof(MDB_db);
2235 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2236 for (i = 2; i < txn->mt_numdbs; i++) {
2237 if (txn->mt_dbflags[i] & DB_DIRTY) {
2238 data.mv_data = &txn->mt_dbs[i];
2239 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2246 /* Save the freelist as of this transaction to the freeDB. This
2247 * can change the freelist, so keep trying until it stabilizes.
2249 * env->me_pglast and the length of txn->mt_free_pgs cannot decrease,
2250 * except the code below can decrease env->me_pglast to split pghead.
2251 * Page numbers cannot disappear from txn->mt_free_pgs. New pages
2252 * can only appear in env->me_pghead when env->me_pglast increases.
2253 * Until then, the me_pghead pointer won't move but can become NULL.
2256 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2257 oldpg_txnid = id = 0;
2260 /* should only be one record now */
2261 if (env->me_pghead || env->me_pglast) {
2262 /* make sure first page of freeDB is touched and on freelist */
2263 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2264 if (rc && rc != MDB_NOTFOUND) {
2271 /* Delete IDLs we used from the free list */
2272 if (env->me_pglast) {
2277 rc = mdb_cursor_first(&mc, &key, NULL);
2280 oldpg_txnid = *(txnid_t *)key.mv_data;
2282 assert(oldpg_txnid <= env->me_pglast);
2284 rc = mdb_cursor_del(&mc, 0);
2287 } while (oldpg_txnid < env->me_pglast);
2290 /* Save IDL of pages freed by this txn, to freeDB */
2292 if (freecnt != txn->mt_free_pgs[0]) {
2295 /* make sure last page of freeDB is touched and on freelist */
2296 key.mv_size = MDB_MAXKEYSIZE+1;
2298 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2299 if (rc && rc != MDB_NOTFOUND)
2305 MDB_IDL idl = txn->mt_free_pgs;
2306 mdb_midl_sort(txn->mt_free_pgs);
2307 DPRINTF("IDL write txn %zu root %zu num %zu",
2308 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2309 for (i=1; i<=idl[0]; i++) {
2310 DPRINTF("IDL %zu", idl[i]);
2314 /* write to last page of freeDB */
2315 key.mv_size = sizeof(pgno_t);
2316 key.mv_data = &txn->mt_txnid;
2317 /* The free list can still grow during this call,
2318 * despite the pre-emptive touches above. So retry
2319 * until the reserved space remains big enough.
2322 assert(freecnt < txn->mt_free_pgs[0]);
2323 freecnt = txn->mt_free_pgs[0];
2324 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2325 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2328 } while (freecnt != txn->mt_free_pgs[0]);
2329 mdb_midl_sort(txn->mt_free_pgs);
2330 memcpy(data.mv_data, txn->mt_free_pgs, data.mv_size);
2331 if (oldpg_txnid < env->me_pglast || (!env->me_pghead && id))
2332 goto free_pgfirst; /* used up freeDB[oldpg_txnid] */
2335 /* Put back page numbers we took from freeDB but did not use */
2336 if (env->me_pghead) {
2341 mop = env->me_pghead;
2342 id = env->me_pglast;
2343 key.mv_size = sizeof(id);
2345 /* These steps may grow the freelist again
2346 * due to freed overflow pages...
2351 if (orig > env->me_maxfree_1pg && id > 4)
2352 orig = env->me_maxfree_1pg; /* Do not use more than 1 page */
2353 data.mv_size = (orig + 1) * sizeof(pgno_t);
2354 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2357 assert(!env->me_pghead || env->me_pglast);
2358 /* mop could have been used again here */
2359 if (id != env->me_pglast || env->me_pghead == NULL)
2360 goto again; /* was completely used up */
2361 assert(mop == env->me_pghead);
2362 } while (mop[0] < orig && --i);
2363 memcpy(data.mv_data, mop, data.mv_size);
2366 *(pgno_t *)data.mv_data = orig;
2367 mop[orig] = mop[0] - orig;
2368 env->me_pghead = mop += orig;
2369 /* Save more oldpages at the previous txnid. */
2370 assert(env->me_pglast == id && id == oldpg_txnid);
2371 env->me_pglast = --oldpg_txnid;
2375 /* Check for growth of freelist again */
2376 if (freecnt != txn->mt_free_pgs[0])
2379 free(env->me_pgfree);
2380 env->me_pghead = env->me_pgfree = NULL;
2382 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2383 if (mdb_midl_shrink(&txn->mt_free_pgs))
2384 env->me_free_pgs = txn->mt_free_pgs;
2391 if (env->me_flags & MDB_WRITEMAP) {
2392 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2393 dp = txn->mt_u.dirty_list[i].mptr;
2394 /* clear dirty flag */
2395 dp->mp_flags &= ~P_DIRTY;
2396 txn->mt_u.dirty_list[i].mid = 0;
2398 txn->mt_u.dirty_list[0].mid = 0;
2402 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2408 /* Windows actually supports scatter/gather I/O, but only on
2409 * unbuffered file handles. Since we're relying on the OS page
2410 * cache for all our data, that's self-defeating. So we just
2411 * write pages one at a time. We use the ov structure to set
2412 * the write offset, to at least save the overhead of a Seek
2416 memset(&ov, 0, sizeof(ov));
2417 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2419 dp = txn->mt_u.dirty_list[i].mptr;
2420 DPRINTF("committing page %zu", dp->mp_pgno);
2421 size = dp->mp_pgno * env->me_psize;
2422 ov.Offset = size & 0xffffffff;
2423 ov.OffsetHigh = size >> 16;
2424 ov.OffsetHigh >>= 16;
2425 /* clear dirty flag */
2426 dp->mp_flags &= ~P_DIRTY;
2427 wsize = env->me_psize;
2428 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2429 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2432 DPRINTF("WriteFile: %d", n);
2439 struct iovec iov[MDB_COMMIT_PAGES];
2443 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2444 dp = txn->mt_u.dirty_list[i].mptr;
2445 if (dp->mp_pgno != next) {
2447 rc = writev(env->me_fd, iov, n);
2451 DPUTS("short write, filesystem full?");
2453 DPRINTF("writev: %s", strerror(n));
2460 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2463 DPRINTF("committing page %zu", dp->mp_pgno);
2464 iov[n].iov_len = env->me_psize;
2465 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2466 iov[n].iov_base = (char *)dp;
2467 size += iov[n].iov_len;
2468 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2469 /* clear dirty flag */
2470 dp->mp_flags &= ~P_DIRTY;
2471 if (++n >= MDB_COMMIT_PAGES) {
2481 rc = writev(env->me_fd, iov, n);
2485 DPUTS("short write, filesystem full?");
2487 DPRINTF("writev: %s", strerror(n));
2494 /* Drop the dirty pages.
2496 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2497 dp = txn->mt_u.dirty_list[i].mptr;
2498 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2499 mdb_page_free(txn->mt_env, dp);
2501 VGMEMP_FREE(txn->mt_env, dp);
2504 txn->mt_u.dirty_list[i].mid = 0;
2506 txn->mt_u.dirty_list[0].mid = 0;
2509 if ((n = mdb_env_sync(env, 0)) != 0 ||
2510 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2518 /* update the DB flags */
2519 for (i = 2; i<txn->mt_numdbs; i++) {
2520 if (txn->mt_dbflags[i] & DB_NEW)
2521 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2523 if (txn->mt_numdbs > env->me_numdbs)
2524 env->me_numdbs = txn->mt_numdbs;
2526 UNLOCK_MUTEX_W(env);
2532 /** Read the environment parameters of a DB environment before
2533 * mapping it into memory.
2534 * @param[in] env the environment handle
2535 * @param[out] meta address of where to store the meta information
2536 * @return 0 on success, non-zero on failure.
2539 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2546 /* We don't know the page size yet, so use a minimum value.
2547 * Read both meta pages so we can use the latest one.
2550 for (i=0; i<2; i++) {
2552 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2554 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2559 else if (rc != MDB_PAGESIZE) {
2563 DPRINTF("read: %s", strerror(err));
2567 p = (MDB_page *)&pbuf;
2569 if (!F_ISSET(p->mp_flags, P_META)) {
2570 DPRINTF("page %zu not a meta page", p->mp_pgno);
2575 if (m->mm_magic != MDB_MAGIC) {
2576 DPUTS("meta has invalid magic");
2580 if (m->mm_version != MDB_VERSION) {
2581 DPRINTF("database is version %u, expected version %u",
2582 m->mm_version, MDB_VERSION);
2583 return MDB_VERSION_MISMATCH;
2587 if (m->mm_txnid > meta->mm_txnid)
2588 memcpy(meta, m, sizeof(*m));
2590 memcpy(meta, m, sizeof(*m));
2592 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2594 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2602 /** Write the environment parameters of a freshly created DB environment.
2603 * @param[in] env the environment handle
2604 * @param[out] meta address of where to store the meta information
2605 * @return 0 on success, non-zero on failure.
2608 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2615 DPUTS("writing new meta page");
2617 GET_PAGESIZE(psize);
2619 meta->mm_magic = MDB_MAGIC;
2620 meta->mm_version = MDB_VERSION;
2621 meta->mm_mapsize = env->me_mapsize;
2622 meta->mm_psize = psize;
2623 meta->mm_last_pg = 1;
2624 meta->mm_flags = env->me_flags & 0xffff;
2625 meta->mm_flags |= MDB_INTEGERKEY;
2626 meta->mm_dbs[0].md_root = P_INVALID;
2627 meta->mm_dbs[1].md_root = P_INVALID;
2629 p = calloc(2, psize);
2631 p->mp_flags = P_META;
2634 memcpy(m, meta, sizeof(*meta));
2636 q = (MDB_page *)((char *)p + psize);
2639 q->mp_flags = P_META;
2642 memcpy(m, meta, sizeof(*meta));
2647 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2648 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2649 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2652 lseek(env->me_fd, 0, SEEK_SET);
2653 rc = write(env->me_fd, p, psize * 2);
2654 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2660 /** Update the environment info to commit a transaction.
2661 * @param[in] txn the transaction that's being committed
2662 * @return 0 on success, non-zero on failure.
2665 mdb_env_write_meta(MDB_txn *txn)
2668 MDB_meta meta, metab, *mp;
2670 int rc, len, toggle;
2677 assert(txn != NULL);
2678 assert(txn->mt_env != NULL);
2680 toggle = !txn->mt_toggle;
2681 DPRINTF("writing meta page %d for root page %zu",
2682 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2685 mp = env->me_metas[toggle];
2687 if (env->me_flags & MDB_WRITEMAP) {
2688 /* Persist any increases of mapsize config */
2689 if (env->me_mapsize > mp->mm_mapsize)
2690 mp->mm_mapsize = env->me_mapsize;
2691 mp->mm_dbs[0] = txn->mt_dbs[0];
2692 mp->mm_dbs[1] = txn->mt_dbs[1];
2693 mp->mm_last_pg = txn->mt_next_pgno - 1;
2694 mp->mm_txnid = txn->mt_txnid;
2695 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2696 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2699 ptr += env->me_psize;
2700 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2707 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2708 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2710 ptr = (char *)&meta;
2711 if (env->me_mapsize > mp->mm_mapsize) {
2712 /* Persist any increases of mapsize config */
2713 meta.mm_mapsize = env->me_mapsize;
2714 off = offsetof(MDB_meta, mm_mapsize);
2716 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2718 len = sizeof(MDB_meta) - off;
2721 meta.mm_dbs[0] = txn->mt_dbs[0];
2722 meta.mm_dbs[1] = txn->mt_dbs[1];
2723 meta.mm_last_pg = txn->mt_next_pgno - 1;
2724 meta.mm_txnid = txn->mt_txnid;
2727 off += env->me_psize;
2730 /* Write to the SYNC fd */
2731 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2732 env->me_fd : env->me_mfd;
2735 memset(&ov, 0, sizeof(ov));
2737 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2740 rc = pwrite(mfd, ptr, len, off);
2745 DPUTS("write failed, disk error?");
2746 /* On a failure, the pagecache still contains the new data.
2747 * Write some old data back, to prevent it from being used.
2748 * Use the non-SYNC fd; we know it will fail anyway.
2750 meta.mm_last_pg = metab.mm_last_pg;
2751 meta.mm_txnid = metab.mm_txnid;
2753 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2755 r2 = pwrite(env->me_fd, ptr, len, off);
2758 env->me_flags |= MDB_FATAL_ERROR;
2762 /* Memory ordering issues are irrelevant; since the entire writer
2763 * is wrapped by wmutex, all of these changes will become visible
2764 * after the wmutex is unlocked. Since the DB is multi-version,
2765 * readers will get consistent data regardless of how fresh or
2766 * how stale their view of these values is.
2768 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2773 /** Check both meta pages to see which one is newer.
2774 * @param[in] env the environment handle
2775 * @return meta toggle (0 or 1).
2778 mdb_env_pick_meta(const MDB_env *env)
2780 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2784 mdb_env_create(MDB_env **env)
2788 e = calloc(1, sizeof(MDB_env));
2792 e->me_maxreaders = DEFAULT_READERS;
2793 e->me_maxdbs = e->me_numdbs = 2;
2794 e->me_fd = INVALID_HANDLE_VALUE;
2795 e->me_lfd = INVALID_HANDLE_VALUE;
2796 e->me_mfd = INVALID_HANDLE_VALUE;
2797 #ifdef MDB_USE_POSIX_SEM
2798 e->me_rmutex = SEM_FAILED;
2799 e->me_wmutex = SEM_FAILED;
2801 e->me_pid = getpid();
2802 VGMEMP_CREATE(e,0,0);
2808 mdb_env_set_mapsize(MDB_env *env, size_t size)
2812 env->me_mapsize = size;
2814 env->me_maxpg = env->me_mapsize / env->me_psize;
2819 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2823 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2828 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2830 if (env->me_map || readers < 1)
2832 env->me_maxreaders = readers;
2837 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2839 if (!env || !readers)
2841 *readers = env->me_maxreaders;
2845 /** Further setup required for opening an MDB environment
2848 mdb_env_open2(MDB_env *env)
2850 unsigned int flags = env->me_flags;
2851 int i, newenv = 0, prot;
2855 memset(&meta, 0, sizeof(meta));
2857 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2860 DPUTS("new mdbenv");
2864 /* Was a mapsize configured? */
2865 if (!env->me_mapsize) {
2866 /* If this is a new environment, take the default,
2867 * else use the size recorded in the existing env.
2869 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2870 } else if (env->me_mapsize < meta.mm_mapsize) {
2871 /* If the configured size is smaller, make sure it's
2872 * still big enough. Silently round up to minimum if not.
2874 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2875 if (env->me_mapsize < minsize)
2876 env->me_mapsize = minsize;
2882 LONG sizelo, sizehi;
2883 sizelo = env->me_mapsize & 0xffffffff;
2884 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2886 /* Windows won't create mappings for zero length files.
2887 * Just allocate the maxsize right now.
2890 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2891 if (!SetEndOfFile(env->me_fd))
2893 SetFilePointer(env->me_fd, 0, NULL, 0);
2895 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2896 PAGE_READWRITE : PAGE_READONLY,
2897 sizehi, sizelo, NULL);
2900 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2901 FILE_MAP_WRITE : FILE_MAP_READ,
2902 0, 0, env->me_mapsize, meta.mm_address);
2910 if (flags & MDB_WRITEMAP) {
2912 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2915 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2917 if (env->me_map == MAP_FAILED) {
2921 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2923 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2925 #ifdef POSIX_MADV_RANDOM
2926 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2927 #endif /* POSIX_MADV_RANDOM */
2928 #endif /* MADV_RANDOM */
2932 if (flags & MDB_FIXEDMAP)
2933 meta.mm_address = env->me_map;
2934 i = mdb_env_init_meta(env, &meta);
2935 if (i != MDB_SUCCESS) {
2938 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2939 /* Can happen because the address argument to mmap() is just a
2940 * hint. mmap() can pick another, e.g. if the range is in use.
2941 * The MAP_FIXED flag would prevent that, but then mmap could
2942 * instead unmap existing pages to make room for the new map.
2944 return EBUSY; /* TODO: Make a new MDB_* error code? */
2946 env->me_psize = meta.mm_psize;
2947 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2948 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2950 env->me_maxpg = env->me_mapsize / env->me_psize;
2952 p = (MDB_page *)env->me_map;
2953 env->me_metas[0] = METADATA(p);
2954 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2958 int toggle = mdb_env_pick_meta(env);
2959 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2961 DPRINTF("opened database version %u, pagesize %u",
2962 env->me_metas[0]->mm_version, env->me_psize);
2963 DPRINTF("using meta page %d", toggle);
2964 DPRINTF("depth: %u", db->md_depth);
2965 DPRINTF("entries: %zu", db->md_entries);
2966 DPRINTF("branch pages: %zu", db->md_branch_pages);
2967 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2968 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2969 DPRINTF("root: %zu", db->md_root);
2977 /** Release a reader thread's slot in the reader lock table.
2978 * This function is called automatically when a thread exits.
2979 * @param[in] ptr This points to the slot in the reader lock table.
2982 mdb_env_reader_dest(void *ptr)
2984 MDB_reader *reader = ptr;
2990 /** Junk for arranging thread-specific callbacks on Windows. This is
2991 * necessarily platform and compiler-specific. Windows supports up
2992 * to 1088 keys. Let's assume nobody opens more than 64 environments
2993 * in a single process, for now. They can override this if needed.
2995 #ifndef MAX_TLS_KEYS
2996 #define MAX_TLS_KEYS 64
2998 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2999 static int mdb_tls_nkeys;
3001 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3005 case DLL_PROCESS_ATTACH: break;
3006 case DLL_THREAD_ATTACH: break;
3007 case DLL_THREAD_DETACH:
3008 for (i=0; i<mdb_tls_nkeys; i++) {
3009 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3010 mdb_env_reader_dest(r);
3013 case DLL_PROCESS_DETACH: break;
3018 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3020 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3024 /* Force some symbol references.
3025 * _tls_used forces the linker to create the TLS directory if not already done
3026 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3028 #pragma comment(linker, "/INCLUDE:_tls_used")
3029 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3030 #pragma const_seg(".CRT$XLB")
3031 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3032 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3035 #pragma comment(linker, "/INCLUDE:__tls_used")
3036 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3037 #pragma data_seg(".CRT$XLB")
3038 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3040 #endif /* WIN 32/64 */
3041 #endif /* !__GNUC__ */
3044 /** Downgrade the exclusive lock on the region back to shared */
3046 mdb_env_share_locks(MDB_env *env, int *excl)
3048 int rc = 0, toggle = mdb_env_pick_meta(env);
3050 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3055 /* First acquire a shared lock. The Unlock will
3056 * then release the existing exclusive lock.
3058 memset(&ov, 0, sizeof(ov));
3059 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3062 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3068 struct flock lock_info;
3069 /* The shared lock replaces the existing lock */
3070 memset((void *)&lock_info, 0, sizeof(lock_info));
3071 lock_info.l_type = F_RDLCK;
3072 lock_info.l_whence = SEEK_SET;
3073 lock_info.l_start = 0;
3074 lock_info.l_len = 1;
3075 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3076 (rc = ErrCode()) == EINTR) ;
3077 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3084 /** Try to get exlusive lock, otherwise shared.
3085 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3088 mdb_env_excl_lock(MDB_env *env, int *excl)
3092 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3096 memset(&ov, 0, sizeof(ov));
3097 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3104 struct flock lock_info;
3105 memset((void *)&lock_info, 0, sizeof(lock_info));
3106 lock_info.l_type = F_WRLCK;
3107 lock_info.l_whence = SEEK_SET;
3108 lock_info.l_start = 0;
3109 lock_info.l_len = 1;
3110 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3111 (rc = ErrCode()) == EINTR) ;
3115 # ifdef MDB_USE_POSIX_SEM
3116 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3119 lock_info.l_type = F_RDLCK;
3120 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3121 (rc = ErrCode()) == EINTR) ;
3129 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3131 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3133 * @(#) $Revision: 5.1 $
3134 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3135 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3137 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3141 * Please do not copyright this code. This code is in the public domain.
3143 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3144 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3145 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3146 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3147 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3148 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3149 * PERFORMANCE OF THIS SOFTWARE.
3152 * chongo <Landon Curt Noll> /\oo/\
3153 * http://www.isthe.com/chongo/
3155 * Share and Enjoy! :-)
3158 typedef unsigned long long mdb_hash_t;
3159 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3161 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3162 * @param[in] str string to hash
3163 * @param[in] hval initial value for hash
3164 * @return 64 bit hash
3166 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3167 * hval arg on the first call.
3170 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3172 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3173 unsigned char *end = s + val->mv_size;
3175 * FNV-1a hash each octet of the string
3178 /* xor the bottom with the current octet */
3179 hval ^= (mdb_hash_t)*s++;
3181 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3182 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3183 (hval << 7) + (hval << 8) + (hval << 40);
3185 /* return our new hash value */
3189 /** Hash the string and output the hash in hex.
3190 * @param[in] str string to hash
3191 * @param[out] hexbuf an array of 17 chars to hold the hash
3194 mdb_hash_hex(MDB_val *val, char *hexbuf)
3197 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3198 for (i=0; i<8; i++) {
3199 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3205 /** Open and/or initialize the lock region for the environment.
3206 * @param[in] env The MDB environment.
3207 * @param[in] lpath The pathname of the file used for the lock region.
3208 * @param[in] mode The Unix permissions for the file, if we create it.
3209 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3210 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3211 * @return 0 on success, non-zero on failure.
3214 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3217 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3219 # define MDB_ERRCODE_ROFS EROFS
3220 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3221 # define MDB_CLOEXEC O_CLOEXEC
3224 # define MDB_CLOEXEC 0
3231 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3232 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3233 FILE_ATTRIBUTE_NORMAL, NULL);
3235 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3237 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3239 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3244 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3245 /* Lose record locks when exec*() */
3246 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3247 fcntl(env->me_lfd, F_SETFD, fdflags);
3250 if (!(env->me_flags & MDB_NOTLS)) {
3251 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3254 env->me_flags |= MDB_ENV_TXKEY;
3256 /* Windows TLS callbacks need help finding their TLS info. */
3257 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3261 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3265 /* Try to get exclusive lock. If we succeed, then
3266 * nobody is using the lock region and we should initialize it.
3268 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3271 size = GetFileSize(env->me_lfd, NULL);
3273 size = lseek(env->me_lfd, 0, SEEK_END);
3275 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3276 if (size < rsize && *excl > 0) {
3278 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3279 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3281 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3285 size = rsize - sizeof(MDB_txninfo);
3286 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3291 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3293 if (!mh) goto fail_errno;
3294 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3296 if (!env->me_txns) goto fail_errno;
3298 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3300 if (m == MAP_FAILED) goto fail_errno;
3306 BY_HANDLE_FILE_INFORMATION stbuf;
3315 if (!mdb_sec_inited) {
3316 InitializeSecurityDescriptor(&mdb_null_sd,
3317 SECURITY_DESCRIPTOR_REVISION);
3318 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3319 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3320 mdb_all_sa.bInheritHandle = FALSE;
3321 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3324 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3325 idbuf.volume = stbuf.dwVolumeSerialNumber;
3326 idbuf.nhigh = stbuf.nFileIndexHigh;
3327 idbuf.nlow = stbuf.nFileIndexLow;
3328 val.mv_data = &idbuf;
3329 val.mv_size = sizeof(idbuf);
3330 mdb_hash_hex(&val, hexbuf);
3331 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3332 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3333 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3334 if (!env->me_rmutex) goto fail_errno;
3335 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3336 if (!env->me_wmutex) goto fail_errno;
3337 #elif defined(MDB_USE_POSIX_SEM)
3346 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3347 idbuf.dev = stbuf.st_dev;
3348 idbuf.ino = stbuf.st_ino;
3349 val.mv_data = &idbuf;
3350 val.mv_size = sizeof(idbuf);
3351 mdb_hash_hex(&val, hexbuf);
3352 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3353 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3354 /* Clean up after a previous run, if needed: Try to
3355 * remove both semaphores before doing anything else.
3357 sem_unlink(env->me_txns->mti_rmname);
3358 sem_unlink(env->me_txns->mti_wmname);
3359 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3360 O_CREAT|O_EXCL, mode, 1);
3361 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3362 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3363 O_CREAT|O_EXCL, mode, 1);
3364 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3365 #else /* MDB_USE_POSIX_SEM */
3366 pthread_mutexattr_t mattr;
3368 if ((rc = pthread_mutexattr_init(&mattr))
3369 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3370 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3371 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3373 pthread_mutexattr_destroy(&mattr);
3374 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3376 env->me_txns->mti_version = MDB_VERSION;
3377 env->me_txns->mti_magic = MDB_MAGIC;
3378 env->me_txns->mti_txnid = 0;
3379 env->me_txns->mti_numreaders = 0;
3382 if (env->me_txns->mti_magic != MDB_MAGIC) {
3383 DPUTS("lock region has invalid magic");
3387 if (env->me_txns->mti_version != MDB_VERSION) {
3388 DPRINTF("lock region is version %u, expected version %u",
3389 env->me_txns->mti_version, MDB_VERSION);
3390 rc = MDB_VERSION_MISMATCH;
3394 if (rc != EACCES && rc != EAGAIN) {
3398 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3399 if (!env->me_rmutex) goto fail_errno;
3400 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3401 if (!env->me_wmutex) goto fail_errno;
3402 #elif defined(MDB_USE_POSIX_SEM)
3403 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3404 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3405 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3406 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3417 /** The name of the lock file in the DB environment */
3418 #define LOCKNAME "/lock.mdb"
3419 /** The name of the data file in the DB environment */
3420 #define DATANAME "/data.mdb"
3421 /** The suffix of the lock file when no subdir is used */
3422 #define LOCKSUFF "-lock"
3423 /** Only a subset of the @ref mdb_env flags can be changed
3424 * at runtime. Changing other flags requires closing the
3425 * environment and re-opening it with the new flags.
3427 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3428 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3431 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3433 int oflags, rc, len, excl = -1;
3434 char *lpath, *dpath;
3436 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3440 if (flags & MDB_NOSUBDIR) {
3441 rc = len + sizeof(LOCKSUFF) + len + 1;
3443 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3448 if (flags & MDB_NOSUBDIR) {
3449 dpath = lpath + len + sizeof(LOCKSUFF);
3450 sprintf(lpath, "%s" LOCKSUFF, path);
3451 strcpy(dpath, path);
3453 dpath = lpath + len + sizeof(LOCKNAME);
3454 sprintf(lpath, "%s" LOCKNAME, path);
3455 sprintf(dpath, "%s" DATANAME, path);
3459 flags |= env->me_flags;
3460 if (flags & MDB_RDONLY) {
3461 /* silently ignore WRITEMAP when we're only getting read access */
3462 flags &= ~MDB_WRITEMAP;
3464 if (!((env->me_free_pgs = mdb_midl_alloc()) &&
3465 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3468 env->me_flags = flags |= MDB_ENV_ACTIVE;
3472 env->me_path = strdup(path);
3473 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3474 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3475 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3480 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3485 if (F_ISSET(flags, MDB_RDONLY)) {
3486 oflags = GENERIC_READ;
3487 len = OPEN_EXISTING;
3489 oflags = GENERIC_READ|GENERIC_WRITE;
3492 mode = FILE_ATTRIBUTE_NORMAL;
3493 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3494 NULL, len, mode, NULL);
3496 if (F_ISSET(flags, MDB_RDONLY))
3499 oflags = O_RDWR | O_CREAT;
3501 env->me_fd = open(dpath, oflags, mode);
3503 if (env->me_fd == INVALID_HANDLE_VALUE) {
3508 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3509 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3510 env->me_mfd = env->me_fd;
3512 /* Synchronous fd for meta writes. Needed even with
3513 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3516 env->me_mfd = CreateFile(dpath, oflags,
3517 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3518 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3520 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3522 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3527 DPRINTF("opened dbenv %p", (void *) env);
3529 rc = mdb_env_share_locks(env, &excl);
3535 mdb_env_close0(env, excl);
3541 /** Destroy resources from mdb_env_open() and clear our readers */
3543 mdb_env_close0(MDB_env *env, int excl)
3547 if (!(env->me_flags & MDB_ENV_ACTIVE))
3550 free(env->me_dbflags);
3553 free(env->me_dirty_list);
3554 if (env->me_free_pgs)
3555 mdb_midl_free(env->me_free_pgs);
3557 if (env->me_flags & MDB_ENV_TXKEY) {
3558 pthread_key_delete(env->me_txkey);
3560 /* Delete our key from the global list */
3561 for (i=0; i<mdb_tls_nkeys; i++)
3562 if (mdb_tls_keys[i] == env->me_txkey) {
3563 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3571 munmap(env->me_map, env->me_mapsize);
3573 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3575 if (env->me_fd != INVALID_HANDLE_VALUE)
3578 pid_t pid = env->me_pid;
3579 /* Clearing readers is done in this function because
3580 * me_txkey with its destructor must be disabled first.
3582 for (i = env->me_numreaders; --i >= 0; )
3583 if (env->me_txns->mti_readers[i].mr_pid == pid)
3584 env->me_txns->mti_readers[i].mr_pid = 0;
3586 if (env->me_rmutex) {
3587 CloseHandle(env->me_rmutex);
3588 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3590 /* Windows automatically destroys the mutexes when
3591 * the last handle closes.
3593 #elif defined(MDB_USE_POSIX_SEM)
3594 if (env->me_rmutex != SEM_FAILED) {
3595 sem_close(env->me_rmutex);
3596 if (env->me_wmutex != SEM_FAILED)
3597 sem_close(env->me_wmutex);
3598 /* If we have the filelock: If we are the
3599 * only remaining user, clean up semaphores.
3602 mdb_env_excl_lock(env, &excl);
3604 sem_unlink(env->me_txns->mti_rmname);
3605 sem_unlink(env->me_txns->mti_wmname);
3609 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3611 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3614 /* Unlock the lockfile. Windows would have unlocked it
3615 * after closing anyway, but not necessarily at once.
3617 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3623 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3627 mdb_env_copy(MDB_env *env, const char *path)
3629 MDB_txn *txn = NULL;
3633 HANDLE newfd = INVALID_HANDLE_VALUE;
3635 if (env->me_flags & MDB_NOSUBDIR) {
3636 lpath = (char *)path;
3639 len += sizeof(DATANAME);
3640 lpath = malloc(len);
3643 sprintf(lpath, "%s" DATANAME, path);
3646 /* The destination path must exist, but the destination file must not.
3647 * We don't want the OS to cache the writes, since the source data is
3648 * already in the OS cache.
3651 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3652 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3654 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3660 if (!(env->me_flags & MDB_NOSUBDIR))
3662 if (newfd == INVALID_HANDLE_VALUE) {
3667 #ifdef F_NOCACHE /* __APPLE__ */
3668 rc = fcntl(newfd, F_NOCACHE, 1);
3675 /* Do the lock/unlock of the reader mutex before starting the
3676 * write txn. Otherwise other read txns could block writers.
3678 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3683 /* We must start the actual read txn after blocking writers */
3684 mdb_txn_reset0(txn);
3686 /* Temporarily block writers until we snapshot the meta pages */
3689 rc = mdb_txn_renew0(txn);
3691 UNLOCK_MUTEX_W(env);
3696 wsize = env->me_psize * 2;
3700 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3701 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3704 rc = write(newfd, env->me_map, wsize);
3705 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3708 UNLOCK_MUTEX_W(env);
3713 ptr = env->me_map + wsize;
3714 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3715 #define MAX_WRITE 2147483648U
3719 if (wsize > MAX_WRITE)
3723 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3724 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3733 if (wsize > MAX_WRITE)
3737 wres = write(newfd, ptr, w2);
3738 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3747 if (newfd != INVALID_HANDLE_VALUE)
3754 mdb_env_close(MDB_env *env)
3762 for (i = env->me_numdbs; --i > MAIN_DBI; )
3763 free(env->me_dbxs[i].md_name.mv_data);
3765 VGMEMP_DESTROY(env);
3766 while ((dp = env->me_dpages) != NULL) {
3767 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3768 env->me_dpages = dp->mp_next;
3772 mdb_env_close0(env, 0);
3776 /** Compare two items pointing at aligned size_t's */
3778 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3780 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3781 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3784 /** Compare two items pointing at aligned int's */
3786 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3788 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3789 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3792 /** Compare two items pointing at ints of unknown alignment.
3793 * Nodes and keys are guaranteed to be 2-byte aligned.
3796 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3798 #if BYTE_ORDER == LITTLE_ENDIAN
3799 unsigned short *u, *c;
3802 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3803 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3806 } while(!x && u > (unsigned short *)a->mv_data);
3809 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3813 /** Compare two items lexically */
3815 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3822 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3828 diff = memcmp(a->mv_data, b->mv_data, len);
3829 return diff ? diff : len_diff<0 ? -1 : len_diff;
3832 /** Compare two items in reverse byte order */
3834 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3836 const unsigned char *p1, *p2, *p1_lim;
3840 p1_lim = (const unsigned char *)a->mv_data;
3841 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3842 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3844 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3850 while (p1 > p1_lim) {
3851 diff = *--p1 - *--p2;
3855 return len_diff<0 ? -1 : len_diff;
3858 /** Search for key within a page, using binary search.
3859 * Returns the smallest entry larger or equal to the key.
3860 * If exactp is non-null, stores whether the found entry was an exact match
3861 * in *exactp (1 or 0).
3862 * Updates the cursor index with the index of the found entry.
3863 * If no entry larger or equal to the key is found, returns NULL.
3866 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3868 unsigned int i = 0, nkeys;
3871 MDB_page *mp = mc->mc_pg[mc->mc_top];
3872 MDB_node *node = NULL;
3877 nkeys = NUMKEYS(mp);
3882 COPY_PGNO(pgno, mp->mp_pgno);
3883 DPRINTF("searching %u keys in %s %spage %zu",
3884 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3891 low = IS_LEAF(mp) ? 0 : 1;
3893 cmp = mc->mc_dbx->md_cmp;
3895 /* Branch pages have no data, so if using integer keys,
3896 * alignment is guaranteed. Use faster mdb_cmp_int.
3898 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3899 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3906 nodekey.mv_size = mc->mc_db->md_pad;
3907 node = NODEPTR(mp, 0); /* fake */
3908 while (low <= high) {
3909 i = (low + high) >> 1;
3910 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3911 rc = cmp(key, &nodekey);
3912 DPRINTF("found leaf index %u [%s], rc = %i",
3913 i, DKEY(&nodekey), rc);
3922 while (low <= high) {
3923 i = (low + high) >> 1;
3925 node = NODEPTR(mp, i);
3926 nodekey.mv_size = NODEKSZ(node);
3927 nodekey.mv_data = NODEKEY(node);
3929 rc = cmp(key, &nodekey);
3932 DPRINTF("found leaf index %u [%s], rc = %i",
3933 i, DKEY(&nodekey), rc);
3935 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3936 i, DKEY(&nodekey), NODEPGNO(node), rc);
3947 if (rc > 0) { /* Found entry is less than the key. */
3948 i++; /* Skip to get the smallest entry larger than key. */
3950 node = NODEPTR(mp, i);
3953 *exactp = (rc == 0);
3954 /* store the key index */
3955 mc->mc_ki[mc->mc_top] = i;
3957 /* There is no entry larger or equal to the key. */
3960 /* nodeptr is fake for LEAF2 */
3966 mdb_cursor_adjust(MDB_cursor *mc, func)
3970 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3971 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3978 /** Pop a page off the top of the cursor's stack. */
3980 mdb_cursor_pop(MDB_cursor *mc)
3983 #ifndef MDB_DEBUG_SKIP
3984 MDB_page *top = mc->mc_pg[mc->mc_top];
3990 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3991 mc->mc_dbi, (void *) mc);
3995 /** Push a page onto the top of the cursor's stack. */
3997 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3999 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
4000 mc->mc_dbi, (void *) mc);
4002 if (mc->mc_snum >= CURSOR_STACK) {
4003 assert(mc->mc_snum < CURSOR_STACK);
4004 return MDB_CURSOR_FULL;
4007 mc->mc_top = mc->mc_snum++;
4008 mc->mc_pg[mc->mc_top] = mp;
4009 mc->mc_ki[mc->mc_top] = 0;
4014 /** Find the address of the page corresponding to a given page number.
4015 * @param[in] txn the transaction for this access.
4016 * @param[in] pgno the page number for the page to retrieve.
4017 * @param[out] ret address of a pointer where the page's address will be stored.
4018 * @return 0 on success, non-zero on failure.
4021 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
4025 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4026 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4030 MDB_ID2L dl = tx2->mt_u.dirty_list;
4032 unsigned x = mdb_mid2l_search(dl, pgno);
4033 if (x <= dl[0].mid && dl[x].mid == pgno) {
4038 } while ((tx2 = tx2->mt_parent) != NULL);
4041 if (pgno < txn->mt_next_pgno) {
4042 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4044 DPRINTF("page %zu not found", pgno);
4050 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
4053 /** Search for the page a given key should be in.
4054 * Pushes parent pages on the cursor stack. This function continues a
4055 * search on a cursor that has already been initialized. (Usually by
4056 * #mdb_page_search() but also by #mdb_node_move().)
4057 * @param[in,out] mc the cursor for this operation.
4058 * @param[in] key the key to search for. If NULL, search for the lowest
4059 * page. (This is used by #mdb_cursor_first().)
4060 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4061 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4062 * @return 0 on success, non-zero on failure.
4065 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4067 MDB_page *mp = mc->mc_pg[mc->mc_top];
4072 while (IS_BRANCH(mp)) {
4076 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4077 assert(NUMKEYS(mp) > 1);
4078 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4080 if (key == NULL) /* Initialize cursor to first page. */
4082 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4083 /* cursor to last page */
4087 node = mdb_node_search(mc, key, &exact);
4089 i = NUMKEYS(mp) - 1;
4091 i = mc->mc_ki[mc->mc_top];
4100 DPRINTF("following index %u for key [%s]",
4102 assert(i < NUMKEYS(mp));
4103 node = NODEPTR(mp, i);
4105 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
4108 mc->mc_ki[mc->mc_top] = i;
4109 if ((rc = mdb_cursor_push(mc, mp)))
4113 if ((rc = mdb_page_touch(mc)) != 0)
4115 mp = mc->mc_pg[mc->mc_top];
4120 DPRINTF("internal error, index points to a %02X page!?",
4122 return MDB_CORRUPTED;
4125 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4126 key ? DKEY(key) : NULL);
4131 /** Search for the page a given key should be in.
4132 * Pushes parent pages on the cursor stack. This function just sets up
4133 * the search; it finds the root page for \b mc's database and sets this
4134 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4135 * called to complete the search.
4136 * @param[in,out] mc the cursor for this operation.
4137 * @param[in] key the key to search for. If NULL, search for the lowest
4138 * page. (This is used by #mdb_cursor_first().)
4139 * @param[in] modify If true, visited pages are updated with new page numbers.
4140 * @return 0 on success, non-zero on failure.
4143 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4148 /* Make sure the txn is still viable, then find the root from
4149 * the txn's db table.
4151 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4152 DPUTS("transaction has failed, must abort");
4155 /* Make sure we're using an up-to-date root */
4156 if (mc->mc_dbi > MAIN_DBI) {
4157 if ((*mc->mc_dbflag & DB_STALE) ||
4158 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4160 unsigned char dbflag = 0;
4161 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4162 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4165 if (*mc->mc_dbflag & DB_STALE) {
4169 MDB_node *leaf = mdb_node_search(&mc2,
4170 &mc->mc_dbx->md_name, &exact);
4172 return MDB_NOTFOUND;
4173 mdb_node_read(mc->mc_txn, leaf, &data);
4174 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4176 /* The txn may not know this DBI, or another process may
4177 * have dropped and recreated the DB with other flags.
4179 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4180 return MDB_INCOMPATIBLE;
4181 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4183 if (flags & MDB_PS_MODIFY)
4185 *mc->mc_dbflag &= ~DB_STALE;
4186 *mc->mc_dbflag |= dbflag;
4189 root = mc->mc_db->md_root;
4191 if (root == P_INVALID) { /* Tree is empty. */
4192 DPUTS("tree is empty");
4193 return MDB_NOTFOUND;
4198 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4199 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
4205 DPRINTF("db %u root page %zu has flags 0x%X",
4206 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4208 if (flags & MDB_PS_MODIFY) {
4209 if ((rc = mdb_page_touch(mc)))
4213 if (flags & MDB_PS_ROOTONLY)
4216 return mdb_page_search_root(mc, key, flags);
4219 /** Return the data associated with a given node.
4220 * @param[in] txn The transaction for this operation.
4221 * @param[in] leaf The node being read.
4222 * @param[out] data Updated to point to the node's data.
4223 * @return 0 on success, non-zero on failure.
4226 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4228 MDB_page *omp; /* overflow page */
4232 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4233 data->mv_size = NODEDSZ(leaf);
4234 data->mv_data = NODEDATA(leaf);
4238 /* Read overflow data.
4240 data->mv_size = NODEDSZ(leaf);
4241 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4242 if ((rc = mdb_page_get(txn, pgno, &omp))) {
4243 DPRINTF("read overflow page %zu failed", pgno);
4246 data->mv_data = METADATA(omp);
4252 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4253 MDB_val *key, MDB_val *data)
4262 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4264 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4267 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4271 mdb_cursor_init(&mc, txn, dbi, &mx);
4272 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4275 /** Find a sibling for a page.
4276 * Replaces the page at the top of the cursor's stack with the
4277 * specified sibling, if one exists.
4278 * @param[in] mc The cursor for this operation.
4279 * @param[in] move_right Non-zero if the right sibling is requested,
4280 * otherwise the left sibling.
4281 * @return 0 on success, non-zero on failure.
4284 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4290 if (mc->mc_snum < 2) {
4291 return MDB_NOTFOUND; /* root has no siblings */
4295 DPRINTF("parent page is page %zu, index %u",
4296 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4298 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4299 : (mc->mc_ki[mc->mc_top] == 0)) {
4300 DPRINTF("no more keys left, moving to %s sibling",
4301 move_right ? "right" : "left");
4302 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4303 /* undo cursor_pop before returning */
4310 mc->mc_ki[mc->mc_top]++;
4312 mc->mc_ki[mc->mc_top]--;
4313 DPRINTF("just moving to %s index key %u",
4314 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4316 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4318 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4319 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4322 mdb_cursor_push(mc, mp);
4324 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4329 /** Move the cursor to the next data item. */
4331 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4337 if (mc->mc_flags & C_EOF) {
4338 return MDB_NOTFOUND;
4341 assert(mc->mc_flags & C_INITIALIZED);
4343 mp = mc->mc_pg[mc->mc_top];
4345 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4346 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4347 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4348 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4349 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4350 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4354 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4355 if (op == MDB_NEXT_DUP)
4356 return MDB_NOTFOUND;
4360 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4362 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4363 DPUTS("=====> move to next sibling page");
4364 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4365 mc->mc_flags |= C_EOF;
4366 mc->mc_flags &= ~C_INITIALIZED;
4367 return MDB_NOTFOUND;
4369 mp = mc->mc_pg[mc->mc_top];
4370 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4372 mc->mc_ki[mc->mc_top]++;
4374 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4375 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4378 key->mv_size = mc->mc_db->md_pad;
4379 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4383 assert(IS_LEAF(mp));
4384 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4386 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4387 mdb_xcursor_init1(mc, leaf);
4390 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4393 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4394 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4395 if (rc != MDB_SUCCESS)
4400 MDB_GET_KEY(leaf, key);
4404 /** Move the cursor to the previous data item. */
4406 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4412 assert(mc->mc_flags & C_INITIALIZED);
4414 mp = mc->mc_pg[mc->mc_top];
4416 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4417 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4418 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4419 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4420 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4421 if (op != MDB_PREV || rc == MDB_SUCCESS)
4424 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4425 if (op == MDB_PREV_DUP)
4426 return MDB_NOTFOUND;
4431 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4433 if (mc->mc_ki[mc->mc_top] == 0) {
4434 DPUTS("=====> move to prev sibling page");
4435 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4436 mc->mc_flags &= ~C_INITIALIZED;
4437 return MDB_NOTFOUND;
4439 mp = mc->mc_pg[mc->mc_top];
4440 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4441 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4443 mc->mc_ki[mc->mc_top]--;
4445 mc->mc_flags &= ~C_EOF;
4447 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4448 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4451 key->mv_size = mc->mc_db->md_pad;
4452 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4456 assert(IS_LEAF(mp));
4457 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4459 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4460 mdb_xcursor_init1(mc, leaf);
4463 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4466 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4467 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4468 if (rc != MDB_SUCCESS)
4473 MDB_GET_KEY(leaf, key);
4477 /** Set the cursor on a specific data item. */
4479 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4480 MDB_cursor_op op, int *exactp)
4484 MDB_node *leaf = NULL;
4489 assert(key->mv_size > 0);
4491 /* See if we're already on the right page */
4492 if (mc->mc_flags & C_INITIALIZED) {
4495 mp = mc->mc_pg[mc->mc_top];
4497 mc->mc_ki[mc->mc_top] = 0;
4498 return MDB_NOTFOUND;
4500 if (mp->mp_flags & P_LEAF2) {
4501 nodekey.mv_size = mc->mc_db->md_pad;
4502 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4504 leaf = NODEPTR(mp, 0);
4505 MDB_GET_KEY(leaf, &nodekey);
4507 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4509 /* Probably happens rarely, but first node on the page
4510 * was the one we wanted.
4512 mc->mc_ki[mc->mc_top] = 0;
4519 unsigned int nkeys = NUMKEYS(mp);
4521 if (mp->mp_flags & P_LEAF2) {
4522 nodekey.mv_data = LEAF2KEY(mp,
4523 nkeys-1, nodekey.mv_size);
4525 leaf = NODEPTR(mp, nkeys-1);
4526 MDB_GET_KEY(leaf, &nodekey);
4528 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4530 /* last node was the one we wanted */
4531 mc->mc_ki[mc->mc_top] = nkeys-1;
4537 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4538 /* This is definitely the right page, skip search_page */
4539 if (mp->mp_flags & P_LEAF2) {
4540 nodekey.mv_data = LEAF2KEY(mp,
4541 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4543 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4544 MDB_GET_KEY(leaf, &nodekey);
4546 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4548 /* current node was the one we wanted */
4558 /* If any parents have right-sibs, search.
4559 * Otherwise, there's nothing further.
4561 for (i=0; i<mc->mc_top; i++)
4563 NUMKEYS(mc->mc_pg[i])-1)
4565 if (i == mc->mc_top) {
4566 /* There are no other pages */
4567 mc->mc_ki[mc->mc_top] = nkeys;
4568 return MDB_NOTFOUND;
4572 /* There are no other pages */
4573 mc->mc_ki[mc->mc_top] = 0;
4574 return MDB_NOTFOUND;
4578 rc = mdb_page_search(mc, key, 0);
4579 if (rc != MDB_SUCCESS)
4582 mp = mc->mc_pg[mc->mc_top];
4583 assert(IS_LEAF(mp));
4586 leaf = mdb_node_search(mc, key, exactp);
4587 if (exactp != NULL && !*exactp) {
4588 /* MDB_SET specified and not an exact match. */
4589 return MDB_NOTFOUND;
4593 DPUTS("===> inexact leaf not found, goto sibling");
4594 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4595 return rc; /* no entries matched */
4596 mp = mc->mc_pg[mc->mc_top];
4597 assert(IS_LEAF(mp));
4598 leaf = NODEPTR(mp, 0);
4602 mc->mc_flags |= C_INITIALIZED;
4603 mc->mc_flags &= ~C_EOF;
4606 key->mv_size = mc->mc_db->md_pad;
4607 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4611 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4612 mdb_xcursor_init1(mc, leaf);
4615 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4616 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4617 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4620 if (op == MDB_GET_BOTH) {
4626 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4627 if (rc != MDB_SUCCESS)
4630 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4632 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4634 rc = mc->mc_dbx->md_dcmp(data, &d2);
4636 if (op == MDB_GET_BOTH || rc > 0)
4637 return MDB_NOTFOUND;
4642 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4643 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4648 /* The key already matches in all other cases */
4649 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4650 MDB_GET_KEY(leaf, key);
4651 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4656 /** Move the cursor to the first item in the database. */
4658 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4663 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4664 rc = mdb_page_search(mc, NULL, 0);
4665 if (rc != MDB_SUCCESS)
4668 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4670 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4671 mc->mc_flags |= C_INITIALIZED;
4672 mc->mc_flags &= ~C_EOF;
4674 mc->mc_ki[mc->mc_top] = 0;
4676 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4677 key->mv_size = mc->mc_db->md_pad;
4678 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4683 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4684 mdb_xcursor_init1(mc, leaf);
4685 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4690 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4691 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4695 MDB_GET_KEY(leaf, key);
4699 /** Move the cursor to the last item in the database. */
4701 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4706 if (!(mc->mc_flags & C_EOF)) {
4708 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4711 lkey.mv_size = MDB_MAXKEYSIZE+1;
4712 lkey.mv_data = NULL;
4713 rc = mdb_page_search(mc, &lkey, 0);
4714 if (rc != MDB_SUCCESS)
4717 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4719 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4721 mc->mc_flags |= C_INITIALIZED|C_EOF;
4722 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4724 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4725 key->mv_size = mc->mc_db->md_pad;
4726 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4731 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4732 mdb_xcursor_init1(mc, leaf);
4733 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4738 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4739 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4744 MDB_GET_KEY(leaf, key);
4749 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4758 case MDB_GET_CURRENT:
4759 if (!(mc->mc_flags & C_INITIALIZED)) {
4762 MDB_page *mp = mc->mc_pg[mc->mc_top];
4764 mc->mc_ki[mc->mc_top] = 0;
4770 key->mv_size = mc->mc_db->md_pad;
4771 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4773 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4774 MDB_GET_KEY(leaf, key);
4776 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4777 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4779 rc = mdb_node_read(mc->mc_txn, leaf, data);
4786 case MDB_GET_BOTH_RANGE:
4787 if (data == NULL || mc->mc_xcursor == NULL) {
4795 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4797 } else if (op == MDB_SET_RANGE)
4798 rc = mdb_cursor_set(mc, key, data, op, NULL);
4800 rc = mdb_cursor_set(mc, key, data, op, &exact);
4802 case MDB_GET_MULTIPLE:
4804 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4805 !(mc->mc_flags & C_INITIALIZED)) {
4810 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4811 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4814 case MDB_NEXT_MULTIPLE:
4816 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4820 if (!(mc->mc_flags & C_INITIALIZED))
4821 rc = mdb_cursor_first(mc, key, data);
4823 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4824 if (rc == MDB_SUCCESS) {
4825 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4828 mx = &mc->mc_xcursor->mx_cursor;
4829 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4831 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4832 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4840 case MDB_NEXT_NODUP:
4841 if (!(mc->mc_flags & C_INITIALIZED))
4842 rc = mdb_cursor_first(mc, key, data);
4844 rc = mdb_cursor_next(mc, key, data, op);
4848 case MDB_PREV_NODUP:
4849 if (!(mc->mc_flags & C_INITIALIZED)) {
4850 rc = mdb_cursor_last(mc, key, data);
4851 mc->mc_flags |= C_INITIALIZED;
4852 mc->mc_ki[mc->mc_top]++;
4854 rc = mdb_cursor_prev(mc, key, data, op);
4857 rc = mdb_cursor_first(mc, key, data);
4861 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4862 !(mc->mc_flags & C_INITIALIZED) ||
4863 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4867 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4870 rc = mdb_cursor_last(mc, key, data);
4874 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4875 !(mc->mc_flags & C_INITIALIZED) ||
4876 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4880 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4883 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4891 /** Touch all the pages in the cursor stack.
4892 * Makes sure all the pages are writable, before attempting a write operation.
4893 * @param[in] mc The cursor to operate on.
4896 mdb_cursor_touch(MDB_cursor *mc)
4900 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4903 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4904 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4905 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4908 *mc->mc_dbflag |= DB_DIRTY;
4910 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4911 rc = mdb_page_touch(mc);
4915 mc->mc_top = mc->mc_snum-1;
4920 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4923 MDB_node *leaf = NULL;
4924 MDB_val xdata, *rdata, dkey;
4927 int do_sub = 0, insert = 0;
4928 unsigned int mcount = 0;
4932 char dbuf[MDB_MAXKEYSIZE+1];
4933 unsigned int nflags;
4936 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4939 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4942 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4945 #if SIZE_MAX > MAXDATASIZE
4946 if (data->mv_size > MAXDATASIZE)
4950 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4951 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4955 if (flags == MDB_CURRENT) {
4956 if (!(mc->mc_flags & C_INITIALIZED))
4959 } else if (mc->mc_db->md_root == P_INVALID) {
4961 /* new database, write a root leaf page */
4962 DPUTS("allocating new root leaf page");
4963 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4967 mdb_cursor_push(mc, np);
4968 mc->mc_db->md_root = np->mp_pgno;
4969 mc->mc_db->md_depth++;
4970 *mc->mc_dbflag |= DB_DIRTY;
4971 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4973 np->mp_flags |= P_LEAF2;
4974 mc->mc_flags |= C_INITIALIZED;
4980 if (flags & MDB_APPEND) {
4982 rc = mdb_cursor_last(mc, &k2, &d2);
4984 rc = mc->mc_dbx->md_cmp(key, &k2);
4987 mc->mc_ki[mc->mc_top]++;
4989 /* new key is <= last key */
4994 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4996 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4997 DPRINTF("duplicate key [%s]", DKEY(key));
4999 return MDB_KEYEXIST;
5001 if (rc && rc != MDB_NOTFOUND)
5005 /* Cursor is positioned, now make sure all pages are writable */
5006 rc2 = mdb_cursor_touch(mc);
5011 /* The key already exists */
5012 if (rc == MDB_SUCCESS) {
5013 /* there's only a key anyway, so this is a no-op */
5014 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5015 unsigned int ksize = mc->mc_db->md_pad;
5016 if (key->mv_size != ksize)
5018 if (flags == MDB_CURRENT) {
5019 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5020 memcpy(ptr, key->mv_data, ksize);
5025 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5028 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5029 /* Was a single item before, must convert now */
5031 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5032 /* Just overwrite the current item */
5033 if (flags == MDB_CURRENT)
5036 dkey.mv_size = NODEDSZ(leaf);
5037 dkey.mv_data = NODEDATA(leaf);
5038 #if UINT_MAX < SIZE_MAX
5039 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5040 #ifdef MISALIGNED_OK
5041 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5043 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5046 /* if data matches, ignore it */
5047 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5048 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5050 /* create a fake page for the dup items */
5051 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5052 dkey.mv_data = dbuf;
5053 fp = (MDB_page *)&pbuf;
5054 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5055 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5056 fp->mp_lower = PAGEHDRSZ;
5057 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5058 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5059 fp->mp_flags |= P_LEAF2;
5060 fp->mp_pad = data->mv_size;
5061 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5063 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5064 (dkey.mv_size & 1) + (data->mv_size & 1);
5066 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5069 xdata.mv_size = fp->mp_upper;
5074 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5075 /* See if we need to convert from fake page to subDB */
5077 unsigned int offset;
5080 fp = NODEDATA(leaf);
5081 if (flags == MDB_CURRENT) {
5083 fp->mp_flags |= P_DIRTY;
5084 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5085 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5089 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5090 offset = fp->mp_pad;
5091 if (SIZELEFT(fp) >= offset)
5093 offset *= 4; /* space for 4 more */
5095 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5097 offset += offset & 1;
5098 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5099 offset >= mc->mc_txn->mt_env->me_nodemax) {
5100 /* yes, convert it */
5102 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5103 dummy.md_pad = fp->mp_pad;
5104 dummy.md_flags = MDB_DUPFIXED;
5105 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5106 dummy.md_flags |= MDB_INTEGERKEY;
5109 dummy.md_branch_pages = 0;
5110 dummy.md_leaf_pages = 1;
5111 dummy.md_overflow_pages = 0;
5112 dummy.md_entries = NUMKEYS(fp);
5114 xdata.mv_size = sizeof(MDB_db);
5115 xdata.mv_data = &dummy;
5116 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5118 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5119 flags |= F_DUPDATA|F_SUBDATA;
5120 dummy.md_root = mp->mp_pgno;
5122 /* no, just grow it */
5124 xdata.mv_size = NODEDSZ(leaf) + offset;
5125 xdata.mv_data = &pbuf;
5126 mp = (MDB_page *)&pbuf;
5127 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5130 mp->mp_flags = fp->mp_flags | P_DIRTY;
5131 mp->mp_pad = fp->mp_pad;
5132 mp->mp_lower = fp->mp_lower;
5133 mp->mp_upper = fp->mp_upper + offset;
5135 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5137 nsize = NODEDSZ(leaf) - fp->mp_upper;
5138 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5139 for (i=0; i<NUMKEYS(fp); i++)
5140 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5142 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5146 /* data is on sub-DB, just store it */
5147 flags |= F_DUPDATA|F_SUBDATA;
5151 /* overflow page overwrites need special handling */
5152 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5155 int ovpages, dpages;
5157 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5158 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5159 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5160 mdb_page_get(mc->mc_txn, pg, &omp);
5161 /* Is the ov page writable and large enough? */
5162 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5163 /* yes, overwrite it. Note in this case we don't
5164 * bother to try shrinking the node if the new data
5165 * is smaller than the overflow threshold.
5167 if (F_ISSET(flags, MDB_RESERVE))
5168 data->mv_data = METADATA(omp);
5170 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5173 /* no, free ovpages */
5175 mc->mc_db->md_overflow_pages -= ovpages;
5176 for (i=0; i<ovpages; i++) {
5177 DPRINTF("freed ov page %zu", pg);
5178 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5182 } else if (NODEDSZ(leaf) == data->mv_size) {
5183 /* same size, just replace it. Note that we could
5184 * also reuse this node if the new data is smaller,
5185 * but instead we opt to shrink the node in that case.
5187 if (F_ISSET(flags, MDB_RESERVE))
5188 data->mv_data = NODEDATA(leaf);
5189 else if (data->mv_size)
5190 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5192 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5195 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5196 mc->mc_db->md_entries--;
5198 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5205 nflags = flags & NODE_ADD_FLAGS;
5206 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5207 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5208 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5209 nflags &= ~MDB_APPEND;
5211 nflags |= MDB_SPLIT_REPLACE;
5212 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5214 /* There is room already in this leaf page. */
5215 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5216 if (rc == 0 && !do_sub && insert) {
5217 /* Adjust other cursors pointing to mp */
5218 MDB_cursor *m2, *m3;
5219 MDB_dbi dbi = mc->mc_dbi;
5220 unsigned i = mc->mc_top;
5221 MDB_page *mp = mc->mc_pg[i];
5223 if (mc->mc_flags & C_SUB)
5226 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5227 if (mc->mc_flags & C_SUB)
5228 m3 = &m2->mc_xcursor->mx_cursor;
5231 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5232 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5239 if (rc != MDB_SUCCESS)
5240 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5242 /* Now store the actual data in the child DB. Note that we're
5243 * storing the user data in the keys field, so there are strict
5244 * size limits on dupdata. The actual data fields of the child
5245 * DB are all zero size.
5252 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5253 if (flags & MDB_CURRENT) {
5254 xflags = MDB_CURRENT;
5256 mdb_xcursor_init1(mc, leaf);
5257 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5259 /* converted, write the original data first */
5261 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5265 /* Adjust other cursors pointing to mp */
5267 unsigned i = mc->mc_top;
5268 MDB_page *mp = mc->mc_pg[i];
5270 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5271 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5272 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5273 mdb_xcursor_init1(m2, leaf);
5277 /* we've done our job */
5280 if (flags & MDB_APPENDDUP)
5281 xflags |= MDB_APPEND;
5282 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5283 if (flags & F_SUBDATA) {
5284 void *db = NODEDATA(leaf);
5285 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5288 /* sub-writes might have failed so check rc again.
5289 * Don't increment count if we just replaced an existing item.
5291 if (!rc && !(flags & MDB_CURRENT))
5292 mc->mc_db->md_entries++;
5293 if (flags & MDB_MULTIPLE) {
5295 if (mcount < data[1].mv_size) {
5296 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5297 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5303 /* If we succeeded and the key didn't exist before, make sure
5304 * the cursor is marked valid.
5307 mc->mc_flags |= C_INITIALIZED;
5312 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5317 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5320 if (!(mc->mc_flags & C_INITIALIZED))
5323 rc = mdb_cursor_touch(mc);
5327 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5329 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5330 if (flags != MDB_NODUPDATA) {
5331 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5332 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5334 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5335 /* If sub-DB still has entries, we're done */
5336 if (mc->mc_xcursor->mx_db.md_entries) {
5337 if (leaf->mn_flags & F_SUBDATA) {
5338 /* update subDB info */
5339 void *db = NODEDATA(leaf);
5340 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5342 /* shrink fake page */
5343 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5345 mc->mc_db->md_entries--;
5348 /* otherwise fall thru and delete the sub-DB */
5351 if (leaf->mn_flags & F_SUBDATA) {
5352 /* add all the child DB's pages to the free list */
5353 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5354 if (rc == MDB_SUCCESS) {
5355 mc->mc_db->md_entries -=
5356 mc->mc_xcursor->mx_db.md_entries;
5361 return mdb_cursor_del0(mc, leaf);
5364 /** Allocate and initialize new pages for a database.
5365 * @param[in] mc a cursor on the database being added to.
5366 * @param[in] flags flags defining what type of page is being allocated.
5367 * @param[in] num the number of pages to allocate. This is usually 1,
5368 * unless allocating overflow pages for a large record.
5369 * @param[out] mp Address of a page, or NULL on failure.
5370 * @return 0 on success, non-zero on failure.
5373 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5378 if ((rc = mdb_page_alloc(mc, num, &np)))
5380 DPRINTF("allocated new mpage %zu, page size %u",
5381 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5382 np->mp_flags = flags | P_DIRTY;
5383 np->mp_lower = PAGEHDRSZ;
5384 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5387 mc->mc_db->md_branch_pages++;
5388 else if (IS_LEAF(np))
5389 mc->mc_db->md_leaf_pages++;
5390 else if (IS_OVERFLOW(np)) {
5391 mc->mc_db->md_overflow_pages += num;
5399 /** Calculate the size of a leaf node.
5400 * The size depends on the environment's page size; if a data item
5401 * is too large it will be put onto an overflow page and the node
5402 * size will only include the key and not the data. Sizes are always
5403 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5404 * of the #MDB_node headers.
5405 * @param[in] env The environment handle.
5406 * @param[in] key The key for the node.
5407 * @param[in] data The data for the node.
5408 * @return The number of bytes needed to store the node.
5411 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5415 sz = LEAFSIZE(key, data);
5416 if (sz >= env->me_nodemax) {
5417 /* put on overflow page */
5418 sz -= data->mv_size - sizeof(pgno_t);
5422 return sz + sizeof(indx_t);
5425 /** Calculate the size of a branch node.
5426 * The size should depend on the environment's page size but since
5427 * we currently don't support spilling large keys onto overflow
5428 * pages, it's simply the size of the #MDB_node header plus the
5429 * size of the key. Sizes are always rounded up to an even number
5430 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5431 * @param[in] env The environment handle.
5432 * @param[in] key The key for the node.
5433 * @return The number of bytes needed to store the node.
5436 mdb_branch_size(MDB_env *env, MDB_val *key)
5441 if (sz >= env->me_nodemax) {
5442 /* put on overflow page */
5443 /* not implemented */
5444 /* sz -= key->size - sizeof(pgno_t); */
5447 return sz + sizeof(indx_t);
5450 /** Add a node to the page pointed to by the cursor.
5451 * @param[in] mc The cursor for this operation.
5452 * @param[in] indx The index on the page where the new node should be added.
5453 * @param[in] key The key for the new node.
5454 * @param[in] data The data for the new node, if any.
5455 * @param[in] pgno The page number, if adding a branch node.
5456 * @param[in] flags Flags for the node.
5457 * @return 0 on success, non-zero on failure. Possible errors are:
5459 * <li>ENOMEM - failed to allocate overflow pages for the node.
5460 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5461 * should never happen since all callers already calculate the
5462 * page's free space before calling this function.
5466 mdb_node_add(MDB_cursor *mc, indx_t indx,
5467 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5470 size_t node_size = NODESIZE;
5473 MDB_page *mp = mc->mc_pg[mc->mc_top];
5474 MDB_page *ofp = NULL; /* overflow page */
5477 assert(mp->mp_upper >= mp->mp_lower);
5479 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5480 IS_LEAF(mp) ? "leaf" : "branch",
5481 IS_SUBP(mp) ? "sub-" : "",
5482 mp->mp_pgno, indx, data ? data->mv_size : 0,
5483 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5486 /* Move higher keys up one slot. */
5487 int ksize = mc->mc_db->md_pad, dif;
5488 char *ptr = LEAF2KEY(mp, indx, ksize);
5489 dif = NUMKEYS(mp) - indx;
5491 memmove(ptr+ksize, ptr, dif*ksize);
5492 /* insert new key */
5493 memcpy(ptr, key->mv_data, ksize);
5495 /* Just using these for counting */
5496 mp->mp_lower += sizeof(indx_t);
5497 mp->mp_upper -= ksize - sizeof(indx_t);
5502 node_size += key->mv_size;
5506 if (F_ISSET(flags, F_BIGDATA)) {
5507 /* Data already on overflow page. */
5508 node_size += sizeof(pgno_t);
5509 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5510 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5512 /* Put data on overflow page. */
5513 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5514 data->mv_size, node_size+data->mv_size);
5515 node_size += sizeof(pgno_t);
5516 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5518 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5521 node_size += data->mv_size;
5524 node_size += node_size & 1;
5526 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5527 DPRINTF("not enough room in page %zu, got %u ptrs",
5528 mp->mp_pgno, NUMKEYS(mp));
5529 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5530 mp->mp_upper - mp->mp_lower);
5531 DPRINTF("node size = %zu", node_size);
5532 return MDB_PAGE_FULL;
5535 /* Move higher pointers up one slot. */
5536 for (i = NUMKEYS(mp); i > indx; i--)
5537 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5539 /* Adjust free space offsets. */
5540 ofs = mp->mp_upper - node_size;
5541 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5542 mp->mp_ptrs[indx] = ofs;
5544 mp->mp_lower += sizeof(indx_t);
5546 /* Write the node data. */
5547 node = NODEPTR(mp, indx);
5548 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5549 node->mn_flags = flags;
5551 SETDSZ(node,data->mv_size);
5556 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5561 if (F_ISSET(flags, F_BIGDATA))
5562 memcpy(node->mn_data + key->mv_size, data->mv_data,
5564 else if (F_ISSET(flags, MDB_RESERVE))
5565 data->mv_data = node->mn_data + key->mv_size;
5567 memcpy(node->mn_data + key->mv_size, data->mv_data,
5570 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5572 if (F_ISSET(flags, MDB_RESERVE))
5573 data->mv_data = METADATA(ofp);
5575 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5582 /** Delete the specified node from a page.
5583 * @param[in] mp The page to operate on.
5584 * @param[in] indx The index of the node to delete.
5585 * @param[in] ksize The size of a node. Only used if the page is
5586 * part of a #MDB_DUPFIXED database.
5589 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5592 indx_t i, j, numkeys, ptr;
5599 COPY_PGNO(pgno, mp->mp_pgno);
5600 DPRINTF("delete node %u on %s page %zu", indx,
5601 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5604 assert(indx < NUMKEYS(mp));
5607 int x = NUMKEYS(mp) - 1 - indx;
5608 base = LEAF2KEY(mp, indx, ksize);
5610 memmove(base, base + ksize, x * ksize);
5611 mp->mp_lower -= sizeof(indx_t);
5612 mp->mp_upper += ksize - sizeof(indx_t);
5616 node = NODEPTR(mp, indx);
5617 sz = NODESIZE + node->mn_ksize;
5619 if (F_ISSET(node->mn_flags, F_BIGDATA))
5620 sz += sizeof(pgno_t);
5622 sz += NODEDSZ(node);
5626 ptr = mp->mp_ptrs[indx];
5627 numkeys = NUMKEYS(mp);
5628 for (i = j = 0; i < numkeys; i++) {
5630 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5631 if (mp->mp_ptrs[i] < ptr)
5632 mp->mp_ptrs[j] += sz;
5637 base = (char *)mp + mp->mp_upper;
5638 memmove(base + sz, base, ptr - mp->mp_upper);
5640 mp->mp_lower -= sizeof(indx_t);
5644 /** Compact the main page after deleting a node on a subpage.
5645 * @param[in] mp The main page to operate on.
5646 * @param[in] indx The index of the subpage on the main page.
5649 mdb_node_shrink(MDB_page *mp, indx_t indx)
5656 indx_t i, numkeys, ptr;
5658 node = NODEPTR(mp, indx);
5659 sp = (MDB_page *)NODEDATA(node);
5660 osize = NODEDSZ(node);
5662 delta = sp->mp_upper - sp->mp_lower;
5663 SETDSZ(node, osize - delta);
5664 xp = (MDB_page *)((char *)sp + delta);
5666 /* shift subpage upward */
5668 nsize = NUMKEYS(sp) * sp->mp_pad;
5669 memmove(METADATA(xp), METADATA(sp), nsize);
5672 nsize = osize - sp->mp_upper;
5673 numkeys = NUMKEYS(sp);
5674 for (i=numkeys-1; i>=0; i--)
5675 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5677 xp->mp_upper = sp->mp_lower;
5678 xp->mp_lower = sp->mp_lower;
5679 xp->mp_flags = sp->mp_flags;
5680 xp->mp_pad = sp->mp_pad;
5681 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5683 /* shift lower nodes upward */
5684 ptr = mp->mp_ptrs[indx];
5685 numkeys = NUMKEYS(mp);
5686 for (i = 0; i < numkeys; i++) {
5687 if (mp->mp_ptrs[i] <= ptr)
5688 mp->mp_ptrs[i] += delta;
5691 base = (char *)mp + mp->mp_upper;
5692 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5693 mp->mp_upper += delta;
5696 /** Initial setup of a sorted-dups cursor.
5697 * Sorted duplicates are implemented as a sub-database for the given key.
5698 * The duplicate data items are actually keys of the sub-database.
5699 * Operations on the duplicate data items are performed using a sub-cursor
5700 * initialized when the sub-database is first accessed. This function does
5701 * the preliminary setup of the sub-cursor, filling in the fields that
5702 * depend only on the parent DB.
5703 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5706 mdb_xcursor_init0(MDB_cursor *mc)
5708 MDB_xcursor *mx = mc->mc_xcursor;
5710 mx->mx_cursor.mc_xcursor = NULL;
5711 mx->mx_cursor.mc_txn = mc->mc_txn;
5712 mx->mx_cursor.mc_db = &mx->mx_db;
5713 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5714 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5715 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5716 mx->mx_cursor.mc_snum = 0;
5717 mx->mx_cursor.mc_top = 0;
5718 mx->mx_cursor.mc_flags = C_SUB;
5719 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5720 mx->mx_dbx.md_dcmp = NULL;
5721 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5724 /** Final setup of a sorted-dups cursor.
5725 * Sets up the fields that depend on the data from the main cursor.
5726 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5727 * @param[in] node The data containing the #MDB_db record for the
5728 * sorted-dup database.
5731 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5733 MDB_xcursor *mx = mc->mc_xcursor;
5735 if (node->mn_flags & F_SUBDATA) {
5736 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5737 mx->mx_cursor.mc_pg[0] = 0;
5738 mx->mx_cursor.mc_snum = 0;
5739 mx->mx_cursor.mc_flags = C_SUB;
5741 MDB_page *fp = NODEDATA(node);
5742 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5743 mx->mx_db.md_flags = 0;
5744 mx->mx_db.md_depth = 1;
5745 mx->mx_db.md_branch_pages = 0;
5746 mx->mx_db.md_leaf_pages = 1;
5747 mx->mx_db.md_overflow_pages = 0;
5748 mx->mx_db.md_entries = NUMKEYS(fp);
5749 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5750 mx->mx_cursor.mc_snum = 1;
5751 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5752 mx->mx_cursor.mc_top = 0;
5753 mx->mx_cursor.mc_pg[0] = fp;
5754 mx->mx_cursor.mc_ki[0] = 0;
5755 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5756 mx->mx_db.md_flags = MDB_DUPFIXED;
5757 mx->mx_db.md_pad = fp->mp_pad;
5758 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5759 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5762 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5764 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5766 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5767 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5768 #if UINT_MAX < SIZE_MAX
5769 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5770 #ifdef MISALIGNED_OK
5771 mx->mx_dbx.md_cmp = mdb_cmp_long;
5773 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5778 /** Initialize a cursor for a given transaction and database. */
5780 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5785 mc->mc_db = &txn->mt_dbs[dbi];
5786 mc->mc_dbx = &txn->mt_dbxs[dbi];
5787 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5792 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5794 mc->mc_xcursor = mx;
5795 mdb_xcursor_init0(mc);
5797 mc->mc_xcursor = NULL;
5799 if (*mc->mc_dbflag & DB_STALE) {
5800 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5805 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5808 MDB_xcursor *mx = NULL;
5809 size_t size = sizeof(MDB_cursor);
5811 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5814 /* Allow read access to the freelist */
5815 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5818 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5819 size += sizeof(MDB_xcursor);
5821 if ((mc = malloc(size)) != NULL) {
5822 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5823 mx = (MDB_xcursor *)(mc + 1);
5825 mdb_cursor_init(mc, txn, dbi, mx);
5826 if (txn->mt_cursors) {
5827 mc->mc_next = txn->mt_cursors[dbi];
5828 txn->mt_cursors[dbi] = mc;
5830 mc->mc_flags |= C_ALLOCD;
5841 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5843 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5846 if (txn->mt_cursors)
5849 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5853 /* Return the count of duplicate data items for the current key */
5855 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5859 if (mc == NULL || countp == NULL)
5862 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5865 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5866 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5869 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5872 *countp = mc->mc_xcursor->mx_db.md_entries;
5878 mdb_cursor_close(MDB_cursor *mc)
5881 /* remove from txn, if tracked */
5882 if (mc->mc_txn->mt_cursors) {
5883 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5884 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5886 *prev = mc->mc_next;
5888 if (mc->mc_flags & C_ALLOCD)
5894 mdb_cursor_txn(MDB_cursor *mc)
5896 if (!mc) return NULL;
5901 mdb_cursor_dbi(MDB_cursor *mc)
5907 /** Replace the key for a node with a new key.
5908 * @param[in] mp The page containing the node to operate on.
5909 * @param[in] indx The index of the node to operate on.
5910 * @param[in] key The new key to use.
5911 * @return 0 on success, non-zero on failure.
5914 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5921 indx_t ptr, i, numkeys, indx;
5924 indx = mc->mc_ki[mc->mc_top];
5925 mp = mc->mc_pg[mc->mc_top];
5926 node = NODEPTR(mp, indx);
5927 ptr = mp->mp_ptrs[indx];
5931 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5932 k2.mv_data = NODEKEY(node);
5933 k2.mv_size = node->mn_ksize;
5934 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5936 mdb_dkey(&k2, kbuf2),
5942 delta0 = delta = key->mv_size - node->mn_ksize;
5944 /* Must be 2-byte aligned. If new key is
5945 * shorter by 1, the shift will be skipped.
5947 delta += (delta & 1);
5949 if (delta > 0 && SIZELEFT(mp) < delta) {
5951 /* not enough space left, do a delete and split */
5952 DPRINTF("Not enough room, delta = %d, splitting...", delta);
5953 pgno = NODEPGNO(node);
5954 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5955 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
5958 numkeys = NUMKEYS(mp);
5959 for (i = 0; i < numkeys; i++) {
5960 if (mp->mp_ptrs[i] <= ptr)
5961 mp->mp_ptrs[i] -= delta;
5964 base = (char *)mp + mp->mp_upper;
5965 len = ptr - mp->mp_upper + NODESIZE;
5966 memmove(base - delta, base, len);
5967 mp->mp_upper -= delta;
5969 node = NODEPTR(mp, indx);
5972 /* But even if no shift was needed, update ksize */
5974 node->mn_ksize = key->mv_size;
5977 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5983 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
5985 /** Move a node from csrc to cdst.
5988 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5995 unsigned short flags;
5999 /* Mark src and dst as dirty. */
6000 if ((rc = mdb_page_touch(csrc)) ||
6001 (rc = mdb_page_touch(cdst)))
6004 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6005 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6006 key.mv_size = csrc->mc_db->md_pad;
6007 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6009 data.mv_data = NULL;
6013 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6014 assert(!((long)srcnode&1));
6015 srcpg = NODEPGNO(srcnode);
6016 flags = srcnode->mn_flags;
6017 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6018 unsigned int snum = csrc->mc_snum;
6020 /* must find the lowest key below src */
6021 mdb_page_search_root(csrc, NULL, 0);
6022 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6023 key.mv_size = csrc->mc_db->md_pad;
6024 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6026 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6027 key.mv_size = NODEKSZ(s2);
6028 key.mv_data = NODEKEY(s2);
6030 csrc->mc_snum = snum--;
6031 csrc->mc_top = snum;
6033 key.mv_size = NODEKSZ(srcnode);
6034 key.mv_data = NODEKEY(srcnode);
6036 data.mv_size = NODEDSZ(srcnode);
6037 data.mv_data = NODEDATA(srcnode);
6039 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6040 unsigned int snum = cdst->mc_snum;
6043 /* must find the lowest key below dst */
6044 mdb_page_search_root(cdst, NULL, 0);
6045 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6046 bkey.mv_size = cdst->mc_db->md_pad;
6047 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6049 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6050 bkey.mv_size = NODEKSZ(s2);
6051 bkey.mv_data = NODEKEY(s2);
6053 cdst->mc_snum = snum--;
6054 cdst->mc_top = snum;
6055 mdb_cursor_copy(cdst, &mn);
6057 rc = mdb_update_key(&mn, &bkey);
6062 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6063 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6064 csrc->mc_ki[csrc->mc_top],
6066 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6067 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6069 /* Add the node to the destination page.
6071 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6072 if (rc != MDB_SUCCESS)
6075 /* Delete the node from the source page.
6077 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6080 /* Adjust other cursors pointing to mp */
6081 MDB_cursor *m2, *m3;
6082 MDB_dbi dbi = csrc->mc_dbi;
6083 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6085 if (csrc->mc_flags & C_SUB)
6088 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6089 if (m2 == csrc) continue;
6090 if (csrc->mc_flags & C_SUB)
6091 m3 = &m2->mc_xcursor->mx_cursor;
6094 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6095 csrc->mc_ki[csrc->mc_top]) {
6096 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6097 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6102 /* Update the parent separators.
6104 if (csrc->mc_ki[csrc->mc_top] == 0) {
6105 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6106 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6107 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6109 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6110 key.mv_size = NODEKSZ(srcnode);
6111 key.mv_data = NODEKEY(srcnode);
6113 DPRINTF("update separator for source page %zu to [%s]",
6114 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6115 mdb_cursor_copy(csrc, &mn);
6118 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6121 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6123 indx_t ix = csrc->mc_ki[csrc->mc_top];
6124 nullkey.mv_size = 0;
6125 csrc->mc_ki[csrc->mc_top] = 0;
6126 rc = mdb_update_key(csrc, &nullkey);
6127 csrc->mc_ki[csrc->mc_top] = ix;
6128 assert(rc == MDB_SUCCESS);
6132 if (cdst->mc_ki[cdst->mc_top] == 0) {
6133 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6134 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6135 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6137 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6138 key.mv_size = NODEKSZ(srcnode);
6139 key.mv_data = NODEKEY(srcnode);
6141 DPRINTF("update separator for destination page %zu to [%s]",
6142 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6143 mdb_cursor_copy(cdst, &mn);
6146 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6149 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6151 indx_t ix = cdst->mc_ki[cdst->mc_top];
6152 nullkey.mv_size = 0;
6153 cdst->mc_ki[cdst->mc_top] = 0;
6154 rc = mdb_update_key(cdst, &nullkey);
6155 cdst->mc_ki[cdst->mc_top] = ix;
6156 assert(rc == MDB_SUCCESS);
6163 /** Merge one page into another.
6164 * The nodes from the page pointed to by \b csrc will
6165 * be copied to the page pointed to by \b cdst and then
6166 * the \b csrc page will be freed.
6167 * @param[in] csrc Cursor pointing to the source page.
6168 * @param[in] cdst Cursor pointing to the destination page.
6171 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6179 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6180 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6182 assert(csrc->mc_snum > 1); /* can't merge root page */
6183 assert(cdst->mc_snum > 1);
6185 /* Mark dst as dirty. */
6186 if ((rc = mdb_page_touch(cdst)))
6189 /* Move all nodes from src to dst.
6191 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6192 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6193 key.mv_size = csrc->mc_db->md_pad;
6194 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6195 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6196 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6197 if (rc != MDB_SUCCESS)
6199 key.mv_data = (char *)key.mv_data + key.mv_size;
6202 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6203 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6204 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6205 unsigned int snum = csrc->mc_snum;
6207 /* must find the lowest key below src */
6208 mdb_page_search_root(csrc, NULL, 0);
6209 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6210 key.mv_size = csrc->mc_db->md_pad;
6211 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6213 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6214 key.mv_size = NODEKSZ(s2);
6215 key.mv_data = NODEKEY(s2);
6217 csrc->mc_snum = snum--;
6218 csrc->mc_top = snum;
6220 key.mv_size = srcnode->mn_ksize;
6221 key.mv_data = NODEKEY(srcnode);
6224 data.mv_size = NODEDSZ(srcnode);
6225 data.mv_data = NODEDATA(srcnode);
6226 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6227 if (rc != MDB_SUCCESS)
6232 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6233 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);
6235 /* Unlink the src page from parent and add to free list.
6237 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6238 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6241 rc = mdb_update_key(csrc, &key);
6247 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6248 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6249 csrc->mc_db->md_leaf_pages--;
6251 csrc->mc_db->md_branch_pages--;
6253 /* Adjust other cursors pointing to mp */
6254 MDB_cursor *m2, *m3;
6255 MDB_dbi dbi = csrc->mc_dbi;
6256 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6258 if (csrc->mc_flags & C_SUB)
6261 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6262 if (csrc->mc_flags & C_SUB)
6263 m3 = &m2->mc_xcursor->mx_cursor;
6266 if (m3 == csrc) continue;
6267 if (m3->mc_snum < csrc->mc_snum) continue;
6268 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6269 m3->mc_pg[csrc->mc_top] = mp;
6270 m3->mc_ki[csrc->mc_top] += nkeys;
6274 mdb_cursor_pop(csrc);
6276 return mdb_rebalance(csrc);
6279 /** Copy the contents of a cursor.
6280 * @param[in] csrc The cursor to copy from.
6281 * @param[out] cdst The cursor to copy to.
6284 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6288 cdst->mc_txn = csrc->mc_txn;
6289 cdst->mc_dbi = csrc->mc_dbi;
6290 cdst->mc_db = csrc->mc_db;
6291 cdst->mc_dbx = csrc->mc_dbx;
6292 cdst->mc_snum = csrc->mc_snum;
6293 cdst->mc_top = csrc->mc_top;
6294 cdst->mc_flags = csrc->mc_flags;
6296 for (i=0; i<csrc->mc_snum; i++) {
6297 cdst->mc_pg[i] = csrc->mc_pg[i];
6298 cdst->mc_ki[i] = csrc->mc_ki[i];
6302 /** Rebalance the tree after a delete operation.
6303 * @param[in] mc Cursor pointing to the page where rebalancing
6305 * @return 0 on success, non-zero on failure.
6308 mdb_rebalance(MDB_cursor *mc)
6312 unsigned int ptop, minkeys;
6315 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6319 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6320 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6321 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6322 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6326 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6327 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6330 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6331 DPRINTF("no need to rebalance page %zu, above fill threshold",
6337 if (mc->mc_snum < 2) {
6338 MDB_page *mp = mc->mc_pg[0];
6340 DPUTS("Can't rebalance a subpage, ignoring");
6343 if (NUMKEYS(mp) == 0) {
6344 DPUTS("tree is completely empty");
6345 mc->mc_db->md_root = P_INVALID;
6346 mc->mc_db->md_depth = 0;
6347 mc->mc_db->md_leaf_pages = 0;
6348 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6352 /* Adjust other cursors pointing to mp */
6353 MDB_cursor *m2, *m3;
6354 MDB_dbi dbi = mc->mc_dbi;
6356 if (mc->mc_flags & C_SUB)
6359 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6360 if (m2 == mc) continue;
6361 if (mc->mc_flags & C_SUB)
6362 m3 = &m2->mc_xcursor->mx_cursor;
6365 if (m3->mc_snum < mc->mc_snum) continue;
6366 if (m3->mc_pg[0] == mp) {
6372 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6373 DPUTS("collapsing root page!");
6374 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6375 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6376 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6379 mc->mc_db->md_depth--;
6380 mc->mc_db->md_branch_pages--;
6382 /* Adjust other cursors pointing to mp */
6383 MDB_cursor *m2, *m3;
6384 MDB_dbi dbi = mc->mc_dbi;
6386 if (mc->mc_flags & C_SUB)
6389 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6390 if (m2 == mc) continue;
6391 if (mc->mc_flags & C_SUB)
6392 m3 = &m2->mc_xcursor->mx_cursor;
6395 if (m3->mc_snum < mc->mc_snum) continue;
6396 if (m3->mc_pg[0] == mp) {
6397 m3->mc_pg[0] = mc->mc_pg[0];
6404 DPUTS("root page doesn't need rebalancing");
6408 /* The parent (branch page) must have at least 2 pointers,
6409 * otherwise the tree is invalid.
6411 ptop = mc->mc_top-1;
6412 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6414 /* Leaf page fill factor is below the threshold.
6415 * Try to move keys from left or right neighbor, or
6416 * merge with a neighbor page.
6421 mdb_cursor_copy(mc, &mn);
6422 mn.mc_xcursor = NULL;
6424 if (mc->mc_ki[ptop] == 0) {
6425 /* We're the leftmost leaf in our parent.
6427 DPUTS("reading right neighbor");
6429 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6430 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6432 mn.mc_ki[mn.mc_top] = 0;
6433 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6435 /* There is at least one neighbor to the left.
6437 DPUTS("reading left neighbor");
6439 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6440 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6442 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6443 mc->mc_ki[mc->mc_top] = 0;
6446 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6447 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);
6449 /* If the neighbor page is above threshold and has enough keys,
6450 * move one key from it. Otherwise we should try to merge them.
6451 * (A branch page must never have less than 2 keys.)
6453 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6454 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6455 return mdb_node_move(&mn, mc);
6457 if (mc->mc_ki[ptop] == 0)
6458 rc = mdb_page_merge(&mn, mc);
6460 rc = mdb_page_merge(mc, &mn);
6461 mc->mc_flags &= ~C_INITIALIZED;
6466 /** Complete a delete operation started by #mdb_cursor_del(). */
6468 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6472 /* add overflow pages to free list */
6473 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6477 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6478 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6479 mc->mc_db->md_overflow_pages -= ovpages;
6480 for (i=0; i<ovpages; i++) {
6481 DPRINTF("freed ov page %zu", pg);
6482 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6486 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6487 mc->mc_db->md_entries--;
6488 rc = mdb_rebalance(mc);
6489 if (rc != MDB_SUCCESS)
6490 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6491 /* if mc points past last node in page, invalidate */
6492 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6493 mc->mc_flags &= ~C_INITIALIZED;
6499 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6500 MDB_val *key, MDB_val *data)
6505 MDB_val rdata, *xdata;
6509 assert(key != NULL);
6511 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6513 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6516 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6520 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6524 mdb_cursor_init(&mc, txn, dbi, &mx);
6535 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6537 /* let mdb_page_split know about this cursor if needed:
6538 * delete will trigger a rebalance; if it needs to move
6539 * a node from one page to another, it will have to
6540 * update the parent's separator key(s). If the new sepkey
6541 * is larger than the current one, the parent page may
6542 * run out of space, triggering a split. We need this
6543 * cursor to be consistent until the end of the rebalance.
6545 mc.mc_next = txn->mt_cursors[dbi];
6546 txn->mt_cursors[dbi] = &mc;
6547 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6548 txn->mt_cursors[dbi] = mc.mc_next;
6553 /** Split a page and insert a new node.
6554 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6555 * The cursor will be updated to point to the actual page and index where
6556 * the node got inserted after the split.
6557 * @param[in] newkey The key for the newly inserted node.
6558 * @param[in] newdata The data for the newly inserted node.
6559 * @param[in] newpgno The page number, if the new node is a branch node.
6560 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6561 * @return 0 on success, non-zero on failure.
6564 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6565 unsigned int nflags)
6568 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6571 unsigned int i, j, split_indx, nkeys, pmax;
6573 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6575 MDB_page *mp, *rp, *pp;
6580 mp = mc->mc_pg[mc->mc_top];
6581 newindx = mc->mc_ki[mc->mc_top];
6583 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6584 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6585 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6587 /* Create a right sibling. */
6588 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6590 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6592 if (mc->mc_snum < 2) {
6593 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6595 /* shift current top to make room for new parent */
6596 mc->mc_pg[1] = mc->mc_pg[0];
6597 mc->mc_ki[1] = mc->mc_ki[0];
6600 mc->mc_db->md_root = pp->mp_pgno;
6601 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6602 mc->mc_db->md_depth++;
6605 /* Add left (implicit) pointer. */
6606 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6607 /* undo the pre-push */
6608 mc->mc_pg[0] = mc->mc_pg[1];
6609 mc->mc_ki[0] = mc->mc_ki[1];
6610 mc->mc_db->md_root = mp->mp_pgno;
6611 mc->mc_db->md_depth--;
6618 ptop = mc->mc_top-1;
6619 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6622 mc->mc_flags |= C_SPLITTING;
6623 mdb_cursor_copy(mc, &mn);
6624 mn.mc_pg[mn.mc_top] = rp;
6625 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6627 if (nflags & MDB_APPEND) {
6628 mn.mc_ki[mn.mc_top] = 0;
6630 split_indx = newindx;
6635 nkeys = NUMKEYS(mp);
6636 split_indx = nkeys / 2;
6637 if (newindx < split_indx)
6643 unsigned int lsize, rsize, ksize;
6644 /* Move half of the keys to the right sibling */
6646 x = mc->mc_ki[mc->mc_top] - split_indx;
6647 ksize = mc->mc_db->md_pad;
6648 split = LEAF2KEY(mp, split_indx, ksize);
6649 rsize = (nkeys - split_indx) * ksize;
6650 lsize = (nkeys - split_indx) * sizeof(indx_t);
6651 mp->mp_lower -= lsize;
6652 rp->mp_lower += lsize;
6653 mp->mp_upper += rsize - lsize;
6654 rp->mp_upper -= rsize - lsize;
6655 sepkey.mv_size = ksize;
6656 if (newindx == split_indx) {
6657 sepkey.mv_data = newkey->mv_data;
6659 sepkey.mv_data = split;
6662 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6663 memcpy(rp->mp_ptrs, split, rsize);
6664 sepkey.mv_data = rp->mp_ptrs;
6665 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6666 memcpy(ins, newkey->mv_data, ksize);
6667 mp->mp_lower += sizeof(indx_t);
6668 mp->mp_upper -= ksize - sizeof(indx_t);
6671 memcpy(rp->mp_ptrs, split, x * ksize);
6672 ins = LEAF2KEY(rp, x, ksize);
6673 memcpy(ins, newkey->mv_data, ksize);
6674 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6675 rp->mp_lower += sizeof(indx_t);
6676 rp->mp_upper -= ksize - sizeof(indx_t);
6677 mc->mc_ki[mc->mc_top] = x;
6678 mc->mc_pg[mc->mc_top] = rp;
6683 /* For leaf pages, check the split point based on what
6684 * fits where, since otherwise mdb_node_add can fail.
6686 * This check is only needed when the data items are
6687 * relatively large, such that being off by one will
6688 * make the difference between success or failure.
6690 * It's also relevant if a page happens to be laid out
6691 * such that one half of its nodes are all "small" and
6692 * the other half of its nodes are "large." If the new
6693 * item is also "large" and falls on the half with
6694 * "large" nodes, it also may not fit.
6697 unsigned int psize, nsize;
6698 /* Maximum free space in an empty page */
6699 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6700 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6701 if ((nkeys < 20) || (nsize > pmax/16)) {
6702 if (newindx <= split_indx) {
6705 for (i=0; i<split_indx; i++) {
6706 node = NODEPTR(mp, i);
6707 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6708 if (F_ISSET(node->mn_flags, F_BIGDATA))
6709 psize += sizeof(pgno_t);
6711 psize += NODEDSZ(node);
6715 split_indx = newindx;
6726 for (i=nkeys-1; i>=split_indx; i--) {
6727 node = NODEPTR(mp, i);
6728 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6729 if (F_ISSET(node->mn_flags, F_BIGDATA))
6730 psize += sizeof(pgno_t);
6732 psize += NODEDSZ(node);
6736 split_indx = newindx;
6747 /* First find the separating key between the split pages.
6748 * The case where newindx == split_indx is ambiguous; the
6749 * new item could go to the new page or stay on the original
6750 * page. If newpos == 1 it goes to the new page.
6752 if (newindx == split_indx && newpos) {
6753 sepkey.mv_size = newkey->mv_size;
6754 sepkey.mv_data = newkey->mv_data;
6756 node = NODEPTR(mp, split_indx);
6757 sepkey.mv_size = node->mn_ksize;
6758 sepkey.mv_data = NODEKEY(node);
6762 DPRINTF("separator is [%s]", DKEY(&sepkey));
6764 /* Copy separator key to the parent.
6766 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6770 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6773 if (mn.mc_snum == mc->mc_snum) {
6774 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6775 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6776 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6777 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6782 /* Right page might now have changed parent.
6783 * Check if left page also changed parent.
6785 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6786 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6787 for (i=0; i<ptop; i++) {
6788 mc->mc_pg[i] = mn.mc_pg[i];
6789 mc->mc_ki[i] = mn.mc_ki[i];
6791 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6792 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6796 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6799 mc->mc_flags ^= C_SPLITTING;
6800 if (rc != MDB_SUCCESS) {
6803 if (nflags & MDB_APPEND) {
6804 mc->mc_pg[mc->mc_top] = rp;
6805 mc->mc_ki[mc->mc_top] = 0;
6806 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6809 for (i=0; i<mc->mc_top; i++)
6810 mc->mc_ki[i] = mn.mc_ki[i];
6817 /* Move half of the keys to the right sibling. */
6819 /* grab a page to hold a temporary copy */
6820 copy = mdb_page_malloc(mc);
6824 copy->mp_pgno = mp->mp_pgno;
6825 copy->mp_flags = mp->mp_flags;
6826 copy->mp_lower = PAGEHDRSZ;
6827 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6828 mc->mc_pg[mc->mc_top] = copy;
6829 for (i = j = 0; i <= nkeys; j++) {
6830 if (i == split_indx) {
6831 /* Insert in right sibling. */
6832 /* Reset insert index for right sibling. */
6833 if (i != newindx || (newpos ^ ins_new)) {
6835 mc->mc_pg[mc->mc_top] = rp;
6839 if (i == newindx && !ins_new) {
6840 /* Insert the original entry that caused the split. */
6841 rkey.mv_data = newkey->mv_data;
6842 rkey.mv_size = newkey->mv_size;
6851 /* Update index for the new key. */
6852 mc->mc_ki[mc->mc_top] = j;
6853 } else if (i == nkeys) {
6856 node = NODEPTR(mp, i);
6857 rkey.mv_data = NODEKEY(node);
6858 rkey.mv_size = node->mn_ksize;
6860 xdata.mv_data = NODEDATA(node);
6861 xdata.mv_size = NODEDSZ(node);
6864 pgno = NODEPGNO(node);
6865 flags = node->mn_flags;
6870 if (!IS_LEAF(mp) && j == 0) {
6871 /* First branch index doesn't need key data. */
6875 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6879 nkeys = NUMKEYS(copy);
6880 for (i=0; i<nkeys; i++)
6881 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6882 mp->mp_lower = copy->mp_lower;
6883 mp->mp_upper = copy->mp_upper;
6884 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6885 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6887 /* reset back to original page */
6888 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6889 mc->mc_pg[mc->mc_top] = mp;
6890 if (nflags & MDB_RESERVE) {
6891 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6892 if (!(node->mn_flags & F_BIGDATA))
6893 newdata->mv_data = NODEDATA(node);
6899 /* return tmp page to freelist */
6900 mdb_page_free(mc->mc_txn->mt_env, copy);
6903 /* Adjust other cursors pointing to mp */
6904 MDB_cursor *m2, *m3;
6905 MDB_dbi dbi = mc->mc_dbi;
6906 int fixup = NUMKEYS(mp);
6908 if (mc->mc_flags & C_SUB)
6911 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6912 if (m2 == mc) continue;
6913 if (mc->mc_flags & C_SUB)
6914 m3 = &m2->mc_xcursor->mx_cursor;
6917 if (!(m3->mc_flags & C_INITIALIZED))
6919 if (m3->mc_flags & C_SPLITTING)
6924 for (k=m3->mc_top; k>=0; k--) {
6925 m3->mc_ki[k+1] = m3->mc_ki[k];
6926 m3->mc_pg[k+1] = m3->mc_pg[k];
6928 if (m3->mc_ki[0] >= split_indx) {
6933 m3->mc_pg[0] = mc->mc_pg[0];
6937 if (m3->mc_pg[mc->mc_top] == mp) {
6938 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6939 m3->mc_ki[mc->mc_top]++;
6940 if (m3->mc_ki[mc->mc_top] >= fixup) {
6941 m3->mc_pg[mc->mc_top] = rp;
6942 m3->mc_ki[mc->mc_top] -= fixup;
6943 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6945 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6946 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6955 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6956 MDB_val *key, MDB_val *data, unsigned int flags)
6961 assert(key != NULL);
6962 assert(data != NULL);
6964 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6967 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6971 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6975 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
6978 mdb_cursor_init(&mc, txn, dbi, &mx);
6979 return mdb_cursor_put(&mc, key, data, flags);
6983 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6985 if ((flag & CHANGEABLE) != flag)
6988 env->me_flags |= flag;
6990 env->me_flags &= ~flag;
6995 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7000 *arg = env->me_flags;
7005 mdb_env_get_path(MDB_env *env, const char **arg)
7010 *arg = env->me_path;
7014 /** Common code for #mdb_stat() and #mdb_env_stat().
7015 * @param[in] env the environment to operate in.
7016 * @param[in] db the #MDB_db record containing the stats to return.
7017 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7018 * @return 0, this function always succeeds.
7021 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7023 arg->ms_psize = env->me_psize;
7024 arg->ms_depth = db->md_depth;
7025 arg->ms_branch_pages = db->md_branch_pages;
7026 arg->ms_leaf_pages = db->md_leaf_pages;
7027 arg->ms_overflow_pages = db->md_overflow_pages;
7028 arg->ms_entries = db->md_entries;
7033 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7037 if (env == NULL || arg == NULL)
7040 toggle = mdb_env_pick_meta(env);
7042 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7046 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7050 if (env == NULL || arg == NULL)
7053 toggle = mdb_env_pick_meta(env);
7054 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7055 arg->me_mapsize = env->me_mapsize;
7056 arg->me_maxreaders = env->me_maxreaders;
7057 arg->me_numreaders = env->me_numreaders;
7058 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7059 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7063 /** Set the default comparison functions for a database.
7064 * Called immediately after a database is opened to set the defaults.
7065 * The user can then override them with #mdb_set_compare() or
7066 * #mdb_set_dupsort().
7067 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7068 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7071 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7073 uint16_t f = txn->mt_dbs[dbi].md_flags;
7075 txn->mt_dbxs[dbi].md_cmp =
7076 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7077 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7079 txn->mt_dbxs[dbi].md_dcmp =
7080 !(f & MDB_DUPSORT) ? 0 :
7081 ((f & MDB_INTEGERDUP)
7082 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7083 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7086 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7091 int rc, dbflag, exact;
7092 unsigned int unused = 0;
7095 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7096 mdb_default_cmp(txn, FREE_DBI);
7099 if ((flags & VALID_FLAGS) != flags)
7105 if (flags & PERSISTENT_FLAGS) {
7106 uint16_t f2 = flags & PERSISTENT_FLAGS;
7107 /* make sure flag changes get committed */
7108 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7109 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7110 txn->mt_flags |= MDB_TXN_DIRTY;
7113 mdb_default_cmp(txn, MAIN_DBI);
7117 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7118 mdb_default_cmp(txn, MAIN_DBI);
7121 /* Is the DB already open? */
7123 for (i=2; i<txn->mt_numdbs; i++) {
7124 if (!txn->mt_dbxs[i].md_name.mv_size) {
7125 /* Remember this free slot */
7126 if (!unused) unused = i;
7129 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7130 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7136 /* If no free slot and max hit, fail */
7137 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7138 return MDB_DBS_FULL;
7140 /* Find the DB info */
7141 dbflag = DB_NEW|DB_VALID;
7144 key.mv_data = (void *)name;
7145 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7146 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7147 if (rc == MDB_SUCCESS) {
7148 /* make sure this is actually a DB */
7149 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7150 if (!(node->mn_flags & F_SUBDATA))
7152 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7153 /* Create if requested */
7155 data.mv_size = sizeof(MDB_db);
7156 data.mv_data = &dummy;
7157 memset(&dummy, 0, sizeof(dummy));
7158 dummy.md_root = P_INVALID;
7159 dummy.md_flags = flags & PERSISTENT_FLAGS;
7160 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7164 /* OK, got info, add to table */
7165 if (rc == MDB_SUCCESS) {
7166 unsigned int slot = unused ? unused : txn->mt_numdbs;
7167 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7168 txn->mt_dbxs[slot].md_name.mv_size = len;
7169 txn->mt_dbxs[slot].md_rel = NULL;
7170 txn->mt_dbflags[slot] = dbflag;
7171 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7173 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7174 mdb_default_cmp(txn, slot);
7183 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7185 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7188 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7191 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7194 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7196 ptr = env->me_dbxs[dbi].md_name.mv_data;
7197 env->me_dbxs[dbi].md_name.mv_data = NULL;
7198 env->me_dbxs[dbi].md_name.mv_size = 0;
7202 /** Add all the DB's pages to the free list.
7203 * @param[in] mc Cursor on the DB to free.
7204 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7205 * @return 0 on success, non-zero on failure.
7208 mdb_drop0(MDB_cursor *mc, int subs)
7212 rc = mdb_page_search(mc, NULL, 0);
7213 if (rc == MDB_SUCCESS) {
7218 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7219 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7222 mdb_cursor_copy(mc, &mx);
7223 while (mc->mc_snum > 0) {
7224 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7225 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7226 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7227 if (ni->mn_flags & F_BIGDATA) {
7228 int j, ovpages = OVPAGES(NODEDSZ(ni), mc->mc_txn->mt_env->me_psize);
7230 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7231 for (j=0; j<ovpages; j++) {
7232 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7235 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7236 mdb_xcursor_init1(mc, ni);
7237 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7243 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7245 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7248 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7253 mc->mc_ki[mc->mc_top] = i;
7254 rc = mdb_cursor_sibling(mc, 1);
7256 /* no more siblings, go back to beginning
7257 * of previous level.
7261 for (i=1; i<mc->mc_snum; i++) {
7263 mc->mc_pg[i] = mx.mc_pg[i];
7268 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7269 mc->mc_db->md_root);
7274 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7279 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7282 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7285 rc = mdb_cursor_open(txn, dbi, &mc);
7289 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7293 /* Can't delete the main DB */
7294 if (del && dbi > MAIN_DBI) {
7295 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7297 txn->mt_dbflags[dbi] = DB_STALE;
7298 mdb_dbi_close(txn->mt_env, dbi);
7301 /* reset the DB record, mark it dirty */
7302 txn->mt_dbflags[dbi] |= DB_DIRTY;
7303 txn->mt_dbs[dbi].md_depth = 0;
7304 txn->mt_dbs[dbi].md_branch_pages = 0;
7305 txn->mt_dbs[dbi].md_leaf_pages = 0;
7306 txn->mt_dbs[dbi].md_overflow_pages = 0;
7307 txn->mt_dbs[dbi].md_entries = 0;
7308 txn->mt_dbs[dbi].md_root = P_INVALID;
7310 txn->mt_flags |= MDB_TXN_DIRTY;
7313 mdb_cursor_close(mc);
7317 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7319 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7322 txn->mt_dbxs[dbi].md_cmp = cmp;
7326 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7328 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7331 txn->mt_dbxs[dbi].md_dcmp = cmp;
7335 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7337 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7340 txn->mt_dbxs[dbi].md_rel = rel;
7344 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7346 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7349 txn->mt_dbxs[dbi].md_relctx = ctx;