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 2011-2012 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.
35 #include <sys/types.h>
37 #include <sys/param.h>
43 #ifdef HAVE_SYS_FILE_H
60 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
61 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
64 #if defined(__APPLE__) || defined (BSD)
65 # define MDB_USE_POSIX_SEM 1
66 # define MDB_FDATASYNC fsync
67 #elif defined(ANDROID)
68 # define MDB_FDATASYNC fsync
73 #ifdef MDB_USE_POSIX_SEM
74 #include <semaphore.h>
79 #include <valgrind/memcheck.h>
80 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
81 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
82 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
83 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
84 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
86 #define VGMEMP_CREATE(h,r,z)
87 #define VGMEMP_ALLOC(h,a,s)
88 #define VGMEMP_FREE(h,a)
89 #define VGMEMP_DESTROY(h)
90 #define VGMEMP_DEFINED(a,s)
94 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
95 /* Solaris just defines one or the other */
96 # define LITTLE_ENDIAN 1234
97 # define BIG_ENDIAN 4321
98 # ifdef _LITTLE_ENDIAN
99 # define BYTE_ORDER LITTLE_ENDIAN
101 # define BYTE_ORDER BIG_ENDIAN
104 # define BYTE_ORDER __BYTE_ORDER
108 #ifndef LITTLE_ENDIAN
109 #define LITTLE_ENDIAN __LITTLE_ENDIAN
112 #define BIG_ENDIAN __BIG_ENDIAN
115 #if defined(__i386) || defined(__x86_64)
116 #define MISALIGNED_OK 1
122 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
123 # error "Unknown or unsupported endianness (BYTE_ORDER)"
124 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
125 # error "Two's complement, reasonably sized integer types, please"
128 /** @defgroup internal MDB Internals
131 /** @defgroup compat Windows Compatibility Macros
132 * A bunch of macros to minimize the amount of platform-specific ifdefs
133 * needed throughout the rest of the code. When the features this library
134 * needs are similar enough to POSIX to be hidden in a one-or-two line
135 * replacement, this macro approach is used.
139 #define pthread_t DWORD
140 #define pthread_mutex_t HANDLE
141 #define pthread_key_t DWORD
142 #define pthread_self() GetCurrentThreadId()
143 #define pthread_key_create(x,y) (*(x) = TlsAlloc())
144 #define pthread_key_delete(x) TlsFree(x)
145 #define pthread_getspecific(x) TlsGetValue(x)
146 #define pthread_setspecific(x,y) TlsSetValue(x,y)
147 #define pthread_mutex_unlock(x) ReleaseMutex(x)
148 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
149 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
150 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
151 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
152 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
153 #define getpid() GetCurrentProcessId()
154 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
155 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
156 #define ErrCode() GetLastError()
157 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
158 #define close(fd) CloseHandle(fd)
159 #define munmap(ptr,len) UnmapViewOfFile(ptr)
162 #ifdef MDB_USE_POSIX_SEM
164 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
165 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
166 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
167 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
170 mdb_sem_wait(sem_t *sem)
173 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
178 /** Lock the reader mutex.
180 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
181 /** Unlock the reader mutex.
183 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
185 /** Lock the writer mutex.
186 * Only a single write transaction is allowed at a time. Other writers
187 * will block waiting for this mutex.
189 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
190 /** Unlock the writer mutex.
192 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
193 #endif /* MDB_USE_POSIX_SEM */
195 /** Get the error code for the last failed system function.
197 #define ErrCode() errno
199 /** An abstraction for a file handle.
200 * On POSIX systems file handles are small integers. On Windows
201 * they're opaque pointers.
205 /** A value for an invalid file handle.
206 * Mainly used to initialize file variables and signify that they are
209 #define INVALID_HANDLE_VALUE (-1)
211 /** Get the size of a memory page for the system.
212 * This is the basic size that the platform's memory manager uses, and is
213 * fundamental to the use of memory-mapped files.
215 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
218 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
221 #define MNAME_LEN (sizeof(pthread_mutex_t))
227 /** A flag for opening a file and requesting synchronous data writes.
228 * This is only used when writing a meta page. It's not strictly needed;
229 * we could just do a normal write and then immediately perform a flush.
230 * But if this flag is available it saves us an extra system call.
232 * @note If O_DSYNC is undefined but exists in /usr/include,
233 * preferably set some compiler flag to get the definition.
234 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
237 # define MDB_DSYNC O_DSYNC
241 /** Function for flushing the data of a file. Define this to fsync
242 * if fdatasync() is not supported.
244 #ifndef MDB_FDATASYNC
245 # define MDB_FDATASYNC fdatasync
249 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
260 /** A page number in the database.
261 * Note that 64 bit page numbers are overkill, since pages themselves
262 * already represent 12-13 bits of addressable memory, and the OS will
263 * always limit applications to a maximum of 63 bits of address space.
265 * @note In the #MDB_node structure, we only store 48 bits of this value,
266 * which thus limits us to only 60 bits of addressable data.
268 typedef MDB_ID pgno_t;
270 /** A transaction ID.
271 * See struct MDB_txn.mt_txnid for details.
273 typedef MDB_ID txnid_t;
275 /** @defgroup debug Debug Macros
279 /** Enable debug output.
280 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
281 * read from and written to the database (used for free space management).
286 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
287 # define DPRINTF (void) /* Vararg macros may be unsupported */
289 static int mdb_debug;
290 static txnid_t mdb_debug_start;
292 /** Print a debug message with printf formatting. */
293 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
294 ((void) ((mdb_debug) && \
295 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
297 # define DPRINTF(fmt, ...) ((void) 0)
299 /** Print a debug string.
300 * The string is printed literally, with no format processing.
302 #define DPUTS(arg) DPRINTF("%s", arg)
305 /** A default memory page size.
306 * The actual size is platform-dependent, but we use this for
307 * boot-strapping. We probably should not be using this any more.
308 * The #GET_PAGESIZE() macro is used to get the actual size.
310 * Note that we don't currently support Huge pages. On Linux,
311 * regular data files cannot use Huge pages, and in general
312 * Huge pages aren't actually pageable. We rely on the OS
313 * demand-pager to read our data and page it out when memory
314 * pressure from other processes is high. So until OSs have
315 * actual paging support for Huge pages, they're not viable.
317 #define MDB_PAGESIZE 4096
319 /** The minimum number of keys required in a database page.
320 * Setting this to a larger value will place a smaller bound on the
321 * maximum size of a data item. Data items larger than this size will
322 * be pushed into overflow pages instead of being stored directly in
323 * the B-tree node. This value used to default to 4. With a page size
324 * of 4096 bytes that meant that any item larger than 1024 bytes would
325 * go into an overflow page. That also meant that on average 2-3KB of
326 * each overflow page was wasted space. The value cannot be lower than
327 * 2 because then there would no longer be a tree structure. With this
328 * value, items larger than 2KB will go into overflow pages, and on
329 * average only 1KB will be wasted.
331 #define MDB_MINKEYS 2
333 /** A stamp that identifies a file as an MDB file.
334 * There's nothing special about this value other than that it is easily
335 * recognizable, and it will reflect any byte order mismatches.
337 #define MDB_MAGIC 0xBEEFC0DE
339 /** The version number for a database's file format. */
340 #define MDB_VERSION 1
342 /** The maximum size of a key in the database.
343 * While data items have essentially unbounded size, we require that
344 * keys all fit onto a regular page. This limit could be raised a bit
345 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
347 #define MAXKEYSIZE 511
352 * This is used for printing a hex dump of a key's contents.
354 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
355 /** Display a key in hex.
357 * Invoke a function to display a key in hex.
359 #define DKEY(x) mdb_dkey(x, kbuf)
361 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
365 /** An invalid page number.
366 * Mainly used to denote an empty tree.
368 #define P_INVALID (~(pgno_t)0)
370 /** Test if a flag \b f is set in a flag word \b w. */
371 #define F_ISSET(w, f) (((w) & (f)) == (f))
373 /** Used for offsets within a single page.
374 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
377 typedef uint16_t indx_t;
379 /** Default size of memory map.
380 * This is certainly too small for any actual applications. Apps should always set
381 * the size explicitly using #mdb_env_set_mapsize().
383 #define DEFAULT_MAPSIZE 1048576
385 /** @defgroup readers Reader Lock Table
386 * Readers don't acquire any locks for their data access. Instead, they
387 * simply record their transaction ID in the reader table. The reader
388 * mutex is needed just to find an empty slot in the reader table. The
389 * slot's address is saved in thread-specific data so that subsequent read
390 * transactions started by the same thread need no further locking to proceed.
392 * Since the database uses multi-version concurrency control, readers don't
393 * actually need any locking. This table is used to keep track of which
394 * readers are using data from which old transactions, so that we'll know
395 * when a particular old transaction is no longer in use. Old transactions
396 * that have discarded any data pages can then have those pages reclaimed
397 * for use by a later write transaction.
399 * The lock table is constructed such that reader slots are aligned with the
400 * processor's cache line size. Any slot is only ever used by one thread.
401 * This alignment guarantees that there will be no contention or cache
402 * thrashing as threads update their own slot info, and also eliminates
403 * any need for locking when accessing a slot.
405 * A writer thread will scan every slot in the table to determine the oldest
406 * outstanding reader transaction. Any freed pages older than this will be
407 * reclaimed by the writer. The writer doesn't use any locks when scanning
408 * this table. This means that there's no guarantee that the writer will
409 * see the most up-to-date reader info, but that's not required for correct
410 * operation - all we need is to know the upper bound on the oldest reader,
411 * we don't care at all about the newest reader. So the only consequence of
412 * reading stale information here is that old pages might hang around a
413 * while longer before being reclaimed. That's actually good anyway, because
414 * the longer we delay reclaiming old pages, the more likely it is that a
415 * string of contiguous pages can be found after coalescing old pages from
416 * many old transactions together.
418 * @todo We don't actually do such coalescing yet, we grab pages from one
419 * old transaction at a time.
422 /** Number of slots in the reader table.
423 * This value was chosen somewhat arbitrarily. 126 readers plus a
424 * couple mutexes fit exactly into 8KB on my development machine.
425 * Applications should set the table size using #mdb_env_set_maxreaders().
427 #define DEFAULT_READERS 126
429 /** The size of a CPU cache line in bytes. We want our lock structures
430 * aligned to this size to avoid false cache line sharing in the
432 * This value works for most CPUs. For Itanium this should be 128.
438 /** The information we store in a single slot of the reader table.
439 * In addition to a transaction ID, we also record the process and
440 * thread ID that owns a slot, so that we can detect stale information,
441 * e.g. threads or processes that went away without cleaning up.
442 * @note We currently don't check for stale records. We simply re-init
443 * the table when we know that we're the only process opening the
446 typedef struct MDB_rxbody {
447 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
448 * Multiple readers that start at the same time will probably have the
449 * same ID here. Again, it's not important to exclude them from
450 * anything; all we need to know is which version of the DB they
451 * started from so we can avoid overwriting any data used in that
452 * particular version.
455 /** The process ID of the process owning this reader txn. */
457 /** The thread ID of the thread owning this txn. */
461 /** The actual reader record, with cacheline padding. */
462 typedef struct MDB_reader {
465 /** shorthand for mrb_txnid */
466 #define mr_txnid mru.mrx.mrb_txnid
467 #define mr_pid mru.mrx.mrb_pid
468 #define mr_tid mru.mrx.mrb_tid
469 /** cache line alignment */
470 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
474 /** The header for the reader table.
475 * The table resides in a memory-mapped file. (This is a different file
476 * than is used for the main database.)
478 * For POSIX the actual mutexes reside in the shared memory of this
479 * mapped file. On Windows, mutexes are named objects allocated by the
480 * kernel; we store the mutex names in this mapped file so that other
481 * processes can grab them. This same approach is also used on
482 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
483 * process-shared POSIX mutexes. For these cases where a named object
484 * is used, the object name is derived from a 64 bit FNV hash of the
485 * environment pathname. As such, naming collisions are extremely
486 * unlikely. If a collision occurs, the results are unpredictable.
488 typedef struct MDB_txbody {
489 /** Stamp identifying this as an MDB file. It must be set
492 /** Version number of this lock file. Must be set to #MDB_VERSION. */
493 uint32_t mtb_version;
494 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
495 char mtb_rmname[MNAME_LEN];
497 /** Mutex protecting access to this table.
498 * This is the reader lock that #LOCK_MUTEX_R acquires.
500 pthread_mutex_t mtb_mutex;
502 /** The ID of the last transaction committed to the database.
503 * This is recorded here only for convenience; the value can always
504 * be determined by reading the main database meta pages.
507 /** The number of slots that have been used in the reader table.
508 * This always records the maximum count, it is not decremented
509 * when readers release their slots.
511 unsigned mtb_numreaders;
514 /** The actual reader table definition. */
515 typedef struct MDB_txninfo {
518 #define mti_magic mt1.mtb.mtb_magic
519 #define mti_version mt1.mtb.mtb_version
520 #define mti_mutex mt1.mtb.mtb_mutex
521 #define mti_rmname mt1.mtb.mtb_rmname
522 #define mti_txnid mt1.mtb.mtb_txnid
523 #define mti_numreaders mt1.mtb.mtb_numreaders
524 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
527 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
528 char mt2_wmname[MNAME_LEN];
529 #define mti_wmname mt2.mt2_wmname
531 pthread_mutex_t mt2_wmutex;
532 #define mti_wmutex mt2.mt2_wmutex
534 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
536 MDB_reader mti_readers[1];
540 /** Common header for all page types.
541 * Overflow records occupy a number of contiguous pages with no
542 * headers on any page after the first.
544 typedef struct MDB_page {
545 #define mp_pgno mp_p.p_pgno
546 #define mp_next mp_p.p_next
548 pgno_t p_pgno; /**< page number */
549 void * p_next; /**< for in-memory list of freed structs */
552 /** @defgroup mdb_page Page Flags
554 * Flags for the page headers.
557 #define P_BRANCH 0x01 /**< branch page */
558 #define P_LEAF 0x02 /**< leaf page */
559 #define P_OVERFLOW 0x04 /**< overflow page */
560 #define P_META 0x08 /**< meta page */
561 #define P_DIRTY 0x10 /**< dirty page */
562 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
563 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
565 uint16_t mp_flags; /**< @ref mdb_page */
566 #define mp_lower mp_pb.pb.pb_lower
567 #define mp_upper mp_pb.pb.pb_upper
568 #define mp_pages mp_pb.pb_pages
571 indx_t pb_lower; /**< lower bound of free space */
572 indx_t pb_upper; /**< upper bound of free space */
574 uint32_t pb_pages; /**< number of overflow pages */
576 indx_t mp_ptrs[1]; /**< dynamic size */
579 /** Size of the page header, excluding dynamic data at the end */
580 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
582 /** Address of first usable data byte in a page, after the header */
583 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
585 /** Number of nodes on a page */
586 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
588 /** The amount of space remaining in the page */
589 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
591 /** The percentage of space used in the page, in tenths of a percent. */
592 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
593 ((env)->me_psize - PAGEHDRSZ))
594 /** The minimum page fill factor, in tenths of a percent.
595 * Pages emptier than this are candidates for merging.
597 #define FILL_THRESHOLD 250
599 /** Test if a page is a leaf page */
600 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
601 /** Test if a page is a LEAF2 page */
602 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
603 /** Test if a page is a branch page */
604 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
605 /** Test if a page is an overflow page */
606 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
607 /** Test if a page is a sub page */
608 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
610 /** The number of overflow pages needed to store the given size. */
611 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
613 /** Header for a single key/data pair within a page.
614 * We guarantee 2-byte alignment for nodes.
616 typedef struct MDB_node {
617 /** lo and hi are used for data size on leaf nodes and for
618 * child pgno on branch nodes. On 64 bit platforms, flags
619 * is also used for pgno. (Branch nodes have no flags).
620 * They are in host byte order in case that lets some
621 * accesses be optimized into a 32-bit word access.
623 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
624 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
625 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
626 /** @defgroup mdb_node Node Flags
628 * Flags for node headers.
631 #define F_BIGDATA 0x01 /**< data put on overflow page */
632 #define F_SUBDATA 0x02 /**< data is a sub-database */
633 #define F_DUPDATA 0x04 /**< data has duplicates */
635 /** valid flags for #mdb_node_add() */
636 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
639 unsigned short mn_flags; /**< @ref mdb_node */
640 unsigned short mn_ksize; /**< key size */
641 char mn_data[1]; /**< key and data are appended here */
644 /** Size of the node header, excluding dynamic data at the end */
645 #define NODESIZE offsetof(MDB_node, mn_data)
647 /** Bit position of top word in page number, for shifting mn_flags */
648 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
650 /** Size of a node in a branch page with a given key.
651 * This is just the node header plus the key, there is no data.
653 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
655 /** Size of a node in a leaf page with a given key and data.
656 * This is node header plus key plus data size.
658 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
660 /** Address of node \b i in page \b p */
661 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
663 /** Address of the key for the node */
664 #define NODEKEY(node) (void *)((node)->mn_data)
666 /** Address of the data for a node */
667 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
669 /** Get the page number pointed to by a branch node */
670 #define NODEPGNO(node) \
671 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
672 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
673 /** Set the page number in a branch node */
674 #define SETPGNO(node,pgno) do { \
675 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
676 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
678 /** Get the size of the data in a leaf node */
679 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
680 /** Set the size of the data for a leaf node */
681 #define SETDSZ(node,size) do { \
682 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
683 /** The size of a key in a node */
684 #define NODEKSZ(node) ((node)->mn_ksize)
686 /** Copy a page number from src to dst */
688 #define COPY_PGNO(dst,src) dst = src
690 #if SIZE_MAX > 4294967295UL
691 #define COPY_PGNO(dst,src) do { \
692 unsigned short *s, *d; \
693 s = (unsigned short *)&(src); \
694 d = (unsigned short *)&(dst); \
701 #define COPY_PGNO(dst,src) do { \
702 unsigned short *s, *d; \
703 s = (unsigned short *)&(src); \
704 d = (unsigned short *)&(dst); \
710 /** The address of a key in a LEAF2 page.
711 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
712 * There are no node headers, keys are stored contiguously.
714 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
716 /** Set the \b node's key into \b key, if requested. */
717 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
718 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
720 /** Information about a single database in the environment. */
721 typedef struct MDB_db {
722 uint32_t md_pad; /**< also ksize for LEAF2 pages */
723 uint16_t md_flags; /**< @ref mdb_open */
724 uint16_t md_depth; /**< depth of this tree */
725 pgno_t md_branch_pages; /**< number of internal pages */
726 pgno_t md_leaf_pages; /**< number of leaf pages */
727 pgno_t md_overflow_pages; /**< number of overflow pages */
728 size_t md_entries; /**< number of data items */
729 pgno_t md_root; /**< the root page of this tree */
732 /** Handle for the DB used to track free pages. */
734 /** Handle for the default DB. */
737 /** Meta page content. */
738 typedef struct MDB_meta {
739 /** Stamp identifying this as an MDB file. It must be set
742 /** Version number of this lock file. Must be set to #MDB_VERSION. */
744 void *mm_address; /**< address for fixed mapping */
745 size_t mm_mapsize; /**< size of mmap region */
746 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
747 /** The size of pages used in this DB */
748 #define mm_psize mm_dbs[0].md_pad
749 /** Any persistent environment flags. @ref mdb_env */
750 #define mm_flags mm_dbs[0].md_flags
751 pgno_t mm_last_pg; /**< last used page in file */
752 txnid_t mm_txnid; /**< txnid that committed this page */
755 /** Buffer for a stack-allocated dirty page.
756 * The members define size and alignment, and silence type
757 * aliasing warnings. They are not used directly; that could
758 * mean incorrectly using several union members in parallel.
760 typedef union MDB_pagebuf {
761 char mb_raw[MDB_PAGESIZE];
764 char mm_pad[PAGEHDRSZ];
769 /** Auxiliary DB info.
770 * The information here is mostly static/read-only. There is
771 * only a single copy of this record in the environment.
773 typedef struct MDB_dbx {
774 MDB_val md_name; /**< name of the database */
775 MDB_cmp_func *md_cmp; /**< function for comparing keys */
776 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
777 MDB_rel_func *md_rel; /**< user relocate function */
778 void *md_relctx; /**< user-provided context for md_rel */
781 /** A database transaction.
782 * Every operation requires a transaction handle.
785 MDB_txn *mt_parent; /**< parent of a nested txn */
786 MDB_txn *mt_child; /**< nested txn under this txn */
787 pgno_t mt_next_pgno; /**< next unallocated page */
788 /** The ID of this transaction. IDs are integers incrementing from 1.
789 * Only committed write transactions increment the ID. If a transaction
790 * aborts, the ID may be re-used by the next writer.
793 MDB_env *mt_env; /**< the DB environment */
794 /** The list of pages that became unused during this transaction.
798 MDB_ID2L dirty_list; /**< modified pages */
799 MDB_reader *reader; /**< this thread's slot in the reader table */
801 /** Array of records for each DB known in the environment. */
803 /** Array of MDB_db records for each known DB */
805 /** @defgroup mt_dbflag Transaction DB Flags
809 #define DB_DIRTY 0x01 /**< DB was written in this txn */
810 #define DB_STALE 0x02 /**< DB record is older than txnID */
812 /** Array of cursors for each DB */
813 MDB_cursor **mt_cursors;
814 /** Array of flags for each DB */
815 unsigned char *mt_dbflags;
816 /** Number of DB records in use. This number only ever increments;
817 * we don't decrement it when individual DB handles are closed.
821 /** @defgroup mdb_txn Transaction Flags
825 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
826 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
828 unsigned int mt_flags; /**< @ref mdb_txn */
829 /** Tracks which of the two meta pages was used at the start
830 * of this transaction.
832 unsigned int mt_toggle;
835 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
836 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
837 * raise this on a 64 bit machine.
839 #define CURSOR_STACK 32
843 /** Cursors are used for all DB operations */
845 /** Next cursor on this DB in this txn */
847 /** Original cursor if this is a shadow */
849 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
850 struct MDB_xcursor *mc_xcursor;
851 /** The transaction that owns this cursor */
853 /** The database handle this cursor operates on */
855 /** The database record for this cursor */
857 /** The database auxiliary record for this cursor */
859 /** The @ref mt_dbflag for this database */
860 unsigned char *mc_dbflag;
861 unsigned short mc_snum; /**< number of pushed pages */
862 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
863 /** @defgroup mdb_cursor Cursor Flags
865 * Cursor state flags.
868 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
869 #define C_EOF 0x02 /**< No more data */
870 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
871 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
872 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
873 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
875 unsigned int mc_flags; /**< @ref mdb_cursor */
876 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
877 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
880 /** Context for sorted-dup records.
881 * We could have gone to a fully recursive design, with arbitrarily
882 * deep nesting of sub-databases. But for now we only handle these
883 * levels - main DB, optional sub-DB, sorted-duplicate DB.
885 typedef struct MDB_xcursor {
886 /** A sub-cursor for traversing the Dup DB */
887 MDB_cursor mx_cursor;
888 /** The database record for this Dup DB */
890 /** The auxiliary DB record for this Dup DB */
892 /** The @ref mt_dbflag for this Dup DB */
893 unsigned char mx_dbflag;
896 /** A set of pages freed by an earlier transaction. */
897 typedef struct MDB_oldpages {
898 /** Usually we only read one record from the FREEDB at a time, but
899 * in case we read more, this will chain them together.
901 struct MDB_oldpages *mo_next;
902 /** The ID of the transaction in which these pages were freed. */
904 /** An #MDB_IDL of the pages */
905 pgno_t mo_pages[1]; /* dynamic */
908 /** The database environment. */
910 HANDLE me_fd; /**< The main data file */
911 HANDLE me_lfd; /**< The lock file */
912 HANDLE me_mfd; /**< just for writing the meta pages */
913 /** Failed to update the meta page. Probably an I/O error. */
914 #define MDB_FATAL_ERROR 0x80000000U
915 uint32_t me_flags; /**< @ref mdb_env */
916 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
917 unsigned int me_maxreaders; /**< size of the reader table */
918 unsigned int me_numreaders; /**< max numreaders set by this env */
919 MDB_dbi me_numdbs; /**< number of DBs opened */
920 MDB_dbi me_maxdbs; /**< size of the DB table */
921 pid_t me_pid; /**< process ID of this env */
922 char *me_path; /**< path to the DB files */
923 char *me_map; /**< the memory map of the data file */
924 MDB_txninfo *me_txns; /**< the memory map of the lock file */
925 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
926 MDB_txn *me_txn; /**< current write transaction */
927 size_t me_mapsize; /**< size of the data memory map */
928 off_t me_size; /**< current file size */
929 pgno_t me_maxpg; /**< me_mapsize / me_psize */
930 txnid_t me_pgfirst; /**< ID of first old page record we used */
931 txnid_t me_pglast; /**< ID of last old page record we used */
932 MDB_dbx *me_dbxs; /**< array of static DB info */
933 uint16_t *me_dbflags; /**< array of DB flags */
934 MDB_oldpages *me_pghead; /**< list of old page records */
935 MDB_oldpages *me_pgfree; /**< list of page records to free */
936 pthread_key_t me_txkey; /**< thread-key for readers */
937 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
938 /** IDL of pages that became unused in a write txn */
940 /** ID2L of pages that were written during a write txn */
941 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
943 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
945 #elif defined(MDB_USE_POSIX_SEM)
946 sem_t *me_rmutex; /* Shared mutexes are not supported */
950 /** max number of pages to commit in one writev() call */
951 #define MDB_COMMIT_PAGES 64
952 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
953 #undef MDB_COMMIT_PAGES
954 #define MDB_COMMIT_PAGES IOV_MAX
957 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
958 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
959 static int mdb_page_touch(MDB_cursor *mc);
961 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
962 static int mdb_page_search_root(MDB_cursor *mc,
963 MDB_val *key, int modify);
964 #define MDB_PS_MODIFY 1
965 #define MDB_PS_ROOTONLY 2
966 static int mdb_page_search(MDB_cursor *mc,
967 MDB_val *key, int flags);
968 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
970 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
971 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
972 pgno_t newpgno, unsigned int nflags);
974 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
975 static int mdb_env_pick_meta(const MDB_env *env);
976 static int mdb_env_write_meta(MDB_txn *txn);
978 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
979 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
980 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
981 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
982 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
983 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
984 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
985 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
986 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
988 static int mdb_rebalance(MDB_cursor *mc);
989 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
991 static void mdb_cursor_pop(MDB_cursor *mc);
992 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
994 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
995 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
996 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
997 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
998 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1000 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1001 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1003 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1004 static void mdb_xcursor_init0(MDB_cursor *mc);
1005 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1007 static int mdb_drop0(MDB_cursor *mc, int subs);
1008 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1011 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1015 static SECURITY_DESCRIPTOR mdb_null_sd;
1016 static SECURITY_ATTRIBUTES mdb_all_sa;
1017 static int mdb_sec_inited;
1020 /** Return the library version info. */
1022 mdb_version(int *major, int *minor, int *patch)
1024 if (major) *major = MDB_VERSION_MAJOR;
1025 if (minor) *minor = MDB_VERSION_MINOR;
1026 if (patch) *patch = MDB_VERSION_PATCH;
1027 return MDB_VERSION_STRING;
1030 /** Table of descriptions for MDB @ref errors */
1031 static char *const mdb_errstr[] = {
1032 "MDB_KEYEXIST: Key/data pair already exists",
1033 "MDB_NOTFOUND: No matching key/data pair found",
1034 "MDB_PAGE_NOTFOUND: Requested page not found",
1035 "MDB_CORRUPTED: Located page was wrong type",
1036 "MDB_PANIC: Update of meta page failed",
1037 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1038 "MDB_INVALID: File is not an MDB file",
1039 "MDB_MAP_FULL: Environment mapsize limit reached",
1040 "MDB_DBS_FULL: Environment maxdbs limit reached",
1041 "MDB_READERS_FULL: Environment maxreaders limit reached",
1042 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1043 "MDB_TXN_FULL: Nested transaction has too many dirty pages - transaction too big",
1044 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1045 "MDB_PAGE_FULL: Internal error - page has no more space"
1049 mdb_strerror(int err)
1053 return ("Successful return: 0");
1055 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1056 i = err - MDB_KEYEXIST;
1057 return mdb_errstr[i];
1060 return strerror(err);
1064 /** Display a key in hexadecimal and return the address of the result.
1065 * @param[in] key the key to display
1066 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1067 * @return The key in hexadecimal form.
1070 mdb_dkey(MDB_val *key, char *buf)
1073 unsigned char *c = key->mv_data;
1075 if (key->mv_size > MAXKEYSIZE)
1076 return "MAXKEYSIZE";
1077 /* may want to make this a dynamic check: if the key is mostly
1078 * printable characters, print it as-is instead of converting to hex.
1082 for (i=0; i<key->mv_size; i++)
1083 ptr += sprintf(ptr, "%02x", *c++);
1085 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1090 /** Display all the keys in the page. */
1092 mdb_page_list(MDB_page *mp)
1095 unsigned int i, nkeys, nsize;
1099 nkeys = NUMKEYS(mp);
1100 fprintf(stderr, "numkeys %d\n", nkeys);
1101 for (i=0; i<nkeys; i++) {
1102 node = NODEPTR(mp, i);
1103 key.mv_size = node->mn_ksize;
1104 key.mv_data = node->mn_data;
1105 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1106 if (F_ISSET(node->mn_flags, F_BIGDATA))
1107 nsize += sizeof(pgno_t);
1109 nsize += NODEDSZ(node);
1110 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1115 mdb_cursor_chk(MDB_cursor *mc)
1121 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1122 for (i=0; i<mc->mc_top; i++) {
1124 node = NODEPTR(mp, mc->mc_ki[i]);
1125 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1128 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1134 /** Count all the pages in each DB and in the freelist
1135 * and make sure it matches the actual number of pages
1138 static void mdb_audit(MDB_txn *txn)
1142 MDB_ID freecount, count;
1147 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1148 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1149 freecount += *(MDB_ID *)data.mv_data;
1152 for (i = 0; i<txn->mt_numdbs; i++) {
1153 MDB_xcursor mx, *mxp;
1154 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1155 mdb_cursor_init(&mc, txn, i, mxp);
1156 if (txn->mt_dbs[i].md_root == P_INVALID)
1158 count += txn->mt_dbs[i].md_branch_pages +
1159 txn->mt_dbs[i].md_leaf_pages +
1160 txn->mt_dbs[i].md_overflow_pages;
1161 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1162 mdb_page_search(&mc, NULL, 0);
1166 mp = mc.mc_pg[mc.mc_top];
1167 for (j=0; j<NUMKEYS(mp); j++) {
1168 MDB_node *leaf = NODEPTR(mp, j);
1169 if (leaf->mn_flags & F_SUBDATA) {
1171 memcpy(&db, NODEDATA(leaf), sizeof(db));
1172 count += db.md_branch_pages + db.md_leaf_pages +
1173 db.md_overflow_pages;
1177 while (mdb_cursor_sibling(&mc, 1) == 0);
1180 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1181 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1182 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1188 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1190 return txn->mt_dbxs[dbi].md_cmp(a, b);
1194 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1196 if (txn->mt_dbxs[dbi].md_dcmp)
1197 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1199 return EINVAL; /* too bad you can't distinguish this from a valid result */
1202 /** Allocate a single page.
1203 * Re-use old malloc'd pages first, otherwise just malloc.
1206 mdb_page_malloc(MDB_cursor *mc) {
1208 size_t sz = mc->mc_txn->mt_env->me_psize;
1209 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1210 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1211 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1212 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1213 } else if ((ret = malloc(sz)) != NULL) {
1214 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1219 /** Allocate pages for writing.
1220 * If there are free pages available from older transactions, they
1221 * will be re-used first. Otherwise a new page will be allocated.
1222 * @param[in] mc cursor A cursor handle identifying the transaction and
1223 * database for which we are allocating.
1224 * @param[in] num the number of pages to allocate.
1225 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1226 * will always be satisfied by a single contiguous chunk of memory.
1227 * @return 0 on success, non-zero on failure.
1230 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1232 MDB_txn *txn = mc->mc_txn;
1234 pgno_t pgno = P_INVALID;
1239 /* The free list won't have any content at all until txn 2 has
1240 * committed. The pages freed by txn 2 will be unreferenced
1241 * after txn 3 commits, and so will be safe to re-use in txn 4.
1243 if (txn->mt_txnid > 3) {
1245 if (!txn->mt_env->me_pghead &&
1246 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1247 /* See if there's anything in the free DB */
1253 txnid_t *kptr, last;
1255 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1256 if (!txn->mt_env->me_pgfirst) {
1257 mdb_page_search(&m2, NULL, 0);
1258 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1259 kptr = (txnid_t *)NODEKEY(leaf);
1266 last = txn->mt_env->me_pglast + 1;
1268 key.mv_data = &last;
1269 key.mv_size = sizeof(last);
1270 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1273 last = *(txnid_t *)key.mv_data;
1276 /* Unusable if referred by a meta page or reader... */
1278 if (last < txn->mt_txnid-1) {
1279 j = txn->mt_env->me_txns->mti_numreaders;
1280 r = txn->mt_env->me_txns->mti_readers + j;
1281 for (j = -j; j && (last<r[j].mr_txnid || !r[j].mr_pid); j++) ;
1285 /* It's usable, grab it.
1290 if (!txn->mt_env->me_pgfirst) {
1291 mdb_node_read(txn, leaf, &data);
1293 txn->mt_env->me_pglast = last;
1294 if (!txn->mt_env->me_pgfirst)
1295 txn->mt_env->me_pgfirst = last;
1296 idl = (MDB_ID *) data.mv_data;
1297 /* We might have a zero-length IDL due to freelist growth
1298 * during a prior commit
1300 if (!idl[0]) goto again;
1301 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1304 mop->mo_next = txn->mt_env->me_pghead;
1305 mop->mo_txnid = last;
1306 txn->mt_env->me_pghead = mop;
1307 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1312 DPRINTF("IDL read txn %zu root %zu num %zu",
1313 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1314 for (i=0; i<idl[0]; i++) {
1315 DPRINTF("IDL %zu", idl[i+1]);
1322 if (txn->mt_env->me_pghead) {
1323 MDB_oldpages *mop = txn->mt_env->me_pghead;
1325 /* FIXME: For now, always use fresh pages. We
1326 * really ought to search the free list for a
1331 /* peel pages off tail, so we only have to truncate the list */
1332 pgno = MDB_IDL_LAST(mop->mo_pages);
1333 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1335 if (mop->mo_pages[2] > mop->mo_pages[1])
1336 mop->mo_pages[0] = 0;
1340 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1341 txn->mt_env->me_pghead = mop->mo_next;
1342 if (mc->mc_dbi == FREE_DBI) {
1343 mop->mo_next = txn->mt_env->me_pgfree;
1344 txn->mt_env->me_pgfree = mop;
1353 if (pgno == P_INVALID) {
1354 /* DB size is maxed out */
1355 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1356 DPUTS("DB size maxed out");
1357 return MDB_MAP_FULL;
1360 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1361 if (pgno == P_INVALID) {
1362 pgno = txn->mt_next_pgno;
1363 txn->mt_next_pgno += num;
1365 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1368 if (txn->mt_env->me_dpages && num == 1) {
1369 np = txn->mt_env->me_dpages;
1370 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1371 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1372 txn->mt_env->me_dpages = np->mp_next;
1374 size_t sz = txn->mt_env->me_psize * num;
1375 if ((np = malloc(sz)) == NULL)
1377 VGMEMP_ALLOC(txn->mt_env, np, sz);
1379 if (pgno == P_INVALID) {
1380 np->mp_pgno = txn->mt_next_pgno;
1381 txn->mt_next_pgno += num;
1386 mid.mid = np->mp_pgno;
1388 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1389 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1391 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1398 /** Copy a page: avoid copying unused portions of the page.
1399 * @param[in] dst page to copy into
1400 * @param[in] src page to copy from
1403 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1405 dst->mp_flags = src->mp_flags | P_DIRTY;
1406 dst->mp_pages = src->mp_pages;
1408 if (IS_LEAF2(src)) {
1409 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1411 unsigned int i, nkeys = NUMKEYS(src);
1412 for (i=0; i<nkeys; i++)
1413 dst->mp_ptrs[i] = src->mp_ptrs[i];
1414 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1415 psize - src->mp_upper);
1419 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1420 * @param[in] mc cursor pointing to the page to be touched
1421 * @return 0 on success, non-zero on failure.
1424 mdb_page_touch(MDB_cursor *mc)
1426 MDB_page *mp = mc->mc_pg[mc->mc_top];
1430 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1432 if ((rc = mdb_page_alloc(mc, 1, &np)))
1434 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1435 assert(mp->mp_pgno != np->mp_pgno);
1436 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1438 /* If page isn't full, just copy the used portion */
1439 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1442 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1444 np->mp_flags |= P_DIRTY;
1449 /* Adjust other cursors pointing to mp */
1450 if (mc->mc_flags & C_SUB) {
1451 MDB_cursor *m2, *m3;
1452 MDB_dbi dbi = mc->mc_dbi-1;
1454 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1455 if (m2 == mc) continue;
1456 m3 = &m2->mc_xcursor->mx_cursor;
1457 if (m3->mc_snum < mc->mc_snum) continue;
1458 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1459 m3->mc_pg[mc->mc_top] = mp;
1465 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1466 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1467 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1468 m2->mc_pg[mc->mc_top] = mp;
1472 mc->mc_pg[mc->mc_top] = mp;
1473 /** If this page has a parent, update the parent to point to
1477 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1479 mc->mc_db->md_root = mp->mp_pgno;
1480 } else if (mc->mc_txn->mt_parent) {
1483 /* If txn has a parent, make sure the page is in our
1486 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1487 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1488 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1489 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1490 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1491 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1492 mc->mc_pg[mc->mc_top] = mp;
1498 np = mdb_page_malloc(mc);
1501 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1502 mid.mid = np->mp_pgno;
1504 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1512 mdb_env_sync(MDB_env *env, int force)
1515 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1516 if (env->me_flags & MDB_WRITEMAP) {
1517 int flags = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
1518 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1521 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1525 if (MDB_FDATASYNC(env->me_fd))
1532 /** Make shadow copies of all of parent txn's cursors */
1534 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1536 MDB_cursor *mc, *m2;
1537 unsigned int i, j, size;
1539 for (i=0;i<src->mt_numdbs; i++) {
1540 if (src->mt_cursors[i]) {
1541 size = sizeof(MDB_cursor);
1542 if (src->mt_cursors[i]->mc_xcursor)
1543 size += sizeof(MDB_xcursor);
1544 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1551 mc->mc_db = &dst->mt_dbs[i];
1552 mc->mc_dbx = m2->mc_dbx;
1553 mc->mc_dbflag = &dst->mt_dbflags[i];
1554 mc->mc_snum = m2->mc_snum;
1555 mc->mc_top = m2->mc_top;
1556 mc->mc_flags = m2->mc_flags | C_SHADOW;
1557 for (j=0; j<mc->mc_snum; j++) {
1558 mc->mc_pg[j] = m2->mc_pg[j];
1559 mc->mc_ki[j] = m2->mc_ki[j];
1561 if (m2->mc_xcursor) {
1562 MDB_xcursor *mx, *mx2;
1563 mx = (MDB_xcursor *)(mc+1);
1564 mc->mc_xcursor = mx;
1565 mx2 = m2->mc_xcursor;
1566 mx->mx_db = mx2->mx_db;
1567 mx->mx_dbx = mx2->mx_dbx;
1568 mx->mx_dbflag = mx2->mx_dbflag;
1569 mx->mx_cursor.mc_txn = dst;
1570 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1571 mx->mx_cursor.mc_db = &mx->mx_db;
1572 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1573 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1574 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1575 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1576 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1577 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1578 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1579 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1582 mc->mc_xcursor = NULL;
1584 mc->mc_next = dst->mt_cursors[i];
1585 dst->mt_cursors[i] = mc;
1592 /** Merge shadow cursors back into parent's */
1594 mdb_cursor_merge(MDB_txn *txn)
1597 for (i=0; i<txn->mt_numdbs; i++) {
1598 if (txn->mt_cursors[i]) {
1600 while ((mc = txn->mt_cursors[i])) {
1601 txn->mt_cursors[i] = mc->mc_next;
1602 if (mc->mc_flags & C_SHADOW) {
1603 MDB_cursor *m2 = mc->mc_orig;
1605 m2->mc_snum = mc->mc_snum;
1606 m2->mc_top = mc->mc_top;
1607 for (j=0; j<mc->mc_snum; j++) {
1608 m2->mc_pg[j] = mc->mc_pg[j];
1609 m2->mc_ki[j] = mc->mc_ki[j];
1612 if (mc->mc_flags & C_ALLOCD)
1620 mdb_txn_reset0(MDB_txn *txn);
1622 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1623 * @param[in] txn the transaction handle to initialize
1624 * @return 0 on success, non-zero on failure. This can only
1625 * fail for read-only transactions, and then only if the
1626 * reader table is full.
1629 mdb_txn_renew0(MDB_txn *txn)
1631 MDB_env *env = txn->mt_env;
1635 txn->mt_numdbs = env->me_numdbs;
1636 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1638 if (txn->mt_flags & MDB_TXN_RDONLY) {
1639 MDB_reader *r = pthread_getspecific(env->me_txkey);
1641 pid_t pid = env->me_pid;
1642 pthread_t tid = pthread_self();
1645 for (i=0; i<env->me_txns->mti_numreaders; i++)
1646 if (env->me_txns->mti_readers[i].mr_pid == 0)
1648 if (i == env->me_maxreaders) {
1649 UNLOCK_MUTEX_R(env);
1650 return MDB_READERS_FULL;
1652 env->me_txns->mti_readers[i].mr_pid = pid;
1653 env->me_txns->mti_readers[i].mr_tid = tid;
1654 if (i >= env->me_txns->mti_numreaders)
1655 env->me_txns->mti_numreaders = i+1;
1656 /* Save numreaders for un-mutexed mdb_env_close() */
1657 env->me_numreaders = env->me_txns->mti_numreaders;
1658 UNLOCK_MUTEX_R(env);
1659 r = &env->me_txns->mti_readers[i];
1660 pthread_setspecific(env->me_txkey, r);
1662 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1663 txn->mt_toggle = txn->mt_txnid & 1;
1664 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1665 txn->mt_u.reader = r;
1669 txn->mt_txnid = env->me_txns->mti_txnid;
1670 txn->mt_toggle = txn->mt_txnid & 1;
1671 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1674 if (txn->mt_txnid == mdb_debug_start)
1677 txn->mt_u.dirty_list = env->me_dirty_list;
1678 txn->mt_u.dirty_list[0].mid = 0;
1679 txn->mt_free_pgs = env->me_free_pgs;
1680 txn->mt_free_pgs[0] = 0;
1684 /* Copy the DB info and flags */
1685 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1686 for (i=2; i<txn->mt_numdbs; i++)
1687 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1688 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1689 if (txn->mt_numdbs > 2)
1690 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1696 mdb_txn_renew(MDB_txn *txn)
1703 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1704 DPUTS("environment had fatal error, must shutdown!");
1708 rc = mdb_txn_renew0(txn);
1709 if (rc == MDB_SUCCESS) {
1710 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1711 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1712 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1718 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1723 if (env->me_flags & MDB_FATAL_ERROR) {
1724 DPUTS("environment had fatal error, must shutdown!");
1727 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1730 /* parent already has an active child txn */
1731 if (parent->mt_child) {
1734 /* nested TXNs not supported here */
1735 if (env->me_flags & MDB_WRITEMAP)
1738 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1739 if (!(flags & MDB_RDONLY))
1740 size += env->me_maxdbs * sizeof(MDB_cursor *);
1742 if ((txn = calloc(1, size)) == NULL) {
1743 DPRINTF("calloc: %s", strerror(ErrCode()));
1746 txn->mt_dbs = (MDB_db *)(txn+1);
1747 if (flags & MDB_RDONLY) {
1748 txn->mt_flags |= MDB_TXN_RDONLY;
1749 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1751 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1752 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1757 txn->mt_free_pgs = mdb_midl_alloc();
1758 if (!txn->mt_free_pgs) {
1762 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1763 if (!txn->mt_u.dirty_list) {
1764 free(txn->mt_free_pgs);
1768 txn->mt_txnid = parent->mt_txnid;
1769 txn->mt_toggle = parent->mt_toggle;
1770 txn->mt_u.dirty_list[0].mid = 0;
1771 txn->mt_free_pgs[0] = 0;
1772 txn->mt_next_pgno = parent->mt_next_pgno;
1773 parent->mt_child = txn;
1774 txn->mt_parent = parent;
1775 txn->mt_numdbs = parent->mt_numdbs;
1776 txn->mt_dbxs = parent->mt_dbxs;
1777 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1778 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1779 mdb_cursor_shadow(parent, txn);
1782 rc = mdb_txn_renew0(txn);
1788 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1789 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1790 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1796 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1797 * @param[in] txn the transaction handle to reset
1800 mdb_txn_reset0(MDB_txn *txn)
1802 MDB_env *env = txn->mt_env;
1804 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1805 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1811 /* close(free) all cursors */
1812 for (i=0; i<txn->mt_numdbs; i++) {
1813 if (txn->mt_cursors[i]) {
1815 while ((mc = txn->mt_cursors[i])) {
1816 txn->mt_cursors[i] = mc->mc_next;
1817 if (mc->mc_flags & C_ALLOCD)
1823 if (!(env->me_flags & MDB_WRITEMAP)) {
1824 /* return all dirty pages to dpage list */
1825 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1826 dp = txn->mt_u.dirty_list[i].mptr;
1827 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1828 dp->mp_next = txn->mt_env->me_dpages;
1829 VGMEMP_FREE(txn->mt_env, dp);
1830 txn->mt_env->me_dpages = dp;
1832 /* large pages just get freed directly */
1833 VGMEMP_FREE(txn->mt_env, dp);
1839 if (txn->mt_parent) {
1840 txn->mt_parent->mt_child = NULL;
1841 mdb_midl_free(txn->mt_free_pgs);
1842 free(txn->mt_u.dirty_list);
1845 if (mdb_midl_shrink(&txn->mt_free_pgs))
1846 env->me_free_pgs = txn->mt_free_pgs;
1849 while ((mop = txn->mt_env->me_pghead)) {
1850 txn->mt_env->me_pghead = mop->mo_next;
1853 txn->mt_env->me_pgfirst = 0;
1854 txn->mt_env->me_pglast = 0;
1857 /* The writer mutex was locked in mdb_txn_begin. */
1858 UNLOCK_MUTEX_W(env);
1863 mdb_txn_reset(MDB_txn *txn)
1868 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1869 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1870 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1872 mdb_txn_reset0(txn);
1876 mdb_txn_abort(MDB_txn *txn)
1881 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1882 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1883 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1886 mdb_txn_abort(txn->mt_child);
1888 mdb_txn_reset0(txn);
1893 mdb_txn_commit(MDB_txn *txn)
1901 pgno_t next, freecnt;
1904 assert(txn != NULL);
1905 assert(txn->mt_env != NULL);
1907 if (txn->mt_child) {
1908 mdb_txn_commit(txn->mt_child);
1909 txn->mt_child = NULL;
1914 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1915 if (txn->mt_numdbs > env->me_numdbs) {
1916 /* update the DB flags */
1918 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
1919 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
1926 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1927 DPUTS("error flag is set, can't commit");
1929 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1934 /* Merge (and close) our cursors with parent's */
1935 mdb_cursor_merge(txn);
1937 if (txn->mt_parent) {
1943 /* Update parent's DB table */
1944 ip = &txn->mt_parent->mt_dbs[2];
1945 jp = &txn->mt_dbs[2];
1946 for (i = 2; i < txn->mt_numdbs; i++) {
1947 if (ip->md_root != jp->md_root)
1951 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1953 /* Append our free list to parent's */
1954 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1956 mdb_midl_free(txn->mt_free_pgs);
1958 /* Merge our dirty list with parent's */
1959 dst = txn->mt_parent->mt_u.dirty_list;
1960 src = txn->mt_u.dirty_list;
1961 x = mdb_mid2l_search(dst, src[1].mid);
1962 for (y=1; y<=src[0].mid; y++) {
1963 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1967 dst[x].mptr = src[y].mptr;
1970 for (; y<=src[0].mid; y++) {
1971 if (++x >= MDB_IDL_UM_MAX) {
1973 return MDB_TXN_FULL;
1978 free(txn->mt_u.dirty_list);
1979 txn->mt_parent->mt_child = NULL;
1984 if (txn != env->me_txn) {
1985 DPUTS("attempt to commit unknown transaction");
1990 if (!txn->mt_u.dirty_list[0].mid)
1993 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1994 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1996 /* Update DB root pointers. Their pages have already been
1997 * touched so this is all in-place and cannot fail.
1999 if (txn->mt_numdbs > 2) {
2002 data.mv_size = sizeof(MDB_db);
2004 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2005 for (i = 2; i < txn->mt_numdbs; i++) {
2006 if (txn->mt_dbflags[i] & DB_DIRTY) {
2007 data.mv_data = &txn->mt_dbs[i];
2008 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2013 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2015 /* should only be one record now */
2016 if (env->me_pghead) {
2017 /* make sure first page of freeDB is touched and on freelist */
2018 mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2021 /* Delete IDLs we used from the free list */
2022 if (env->me_pgfirst) {
2027 key.mv_size = sizeof(cur);
2028 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
2031 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
2032 rc = mdb_cursor_del(&mc, 0);
2038 env->me_pgfirst = 0;
2042 /* save to free list */
2044 freecnt = txn->mt_free_pgs[0];
2045 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2048 /* make sure last page of freeDB is touched and on freelist */
2049 key.mv_size = MAXKEYSIZE+1;
2051 mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2053 mdb_midl_sort(txn->mt_free_pgs);
2057 MDB_IDL idl = txn->mt_free_pgs;
2058 DPRINTF("IDL write txn %zu root %zu num %zu",
2059 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2060 for (i=0; i<idl[0]; i++) {
2061 DPRINTF("IDL %zu", idl[i+1]);
2065 /* write to last page of freeDB */
2066 key.mv_size = sizeof(pgno_t);
2067 key.mv_data = &txn->mt_txnid;
2068 data.mv_data = txn->mt_free_pgs;
2069 /* The free list can still grow during this call,
2070 * despite the pre-emptive touches above. So check
2071 * and make sure the entire thing got written.
2074 freecnt = txn->mt_free_pgs[0];
2075 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2076 rc = mdb_cursor_put(&mc, &key, &data, 0);
2081 } while (freecnt != txn->mt_free_pgs[0]);
2083 /* should only be one record now */
2085 if (env->me_pghead) {
2091 mop = env->me_pghead;
2093 key.mv_size = sizeof(id);
2095 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2096 data.mv_data = mop->mo_pages;
2097 orig = mop->mo_pages[0];
2098 /* These steps may grow the freelist again
2099 * due to freed overflow pages...
2101 mdb_cursor_put(&mc, &key, &data, 0);
2102 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2103 /* could have been used again here */
2104 if (mop->mo_pages[0] != orig) {
2105 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2106 data.mv_data = mop->mo_pages;
2108 mdb_cursor_put(&mc, &key, &data, 0);
2110 env->me_pghead = NULL;
2113 /* was completely used up */
2114 mdb_cursor_del(&mc, 0);
2118 env->me_pgfirst = 0;
2122 while (env->me_pgfree) {
2123 MDB_oldpages *mop = env->me_pgfree;
2124 env->me_pgfree = mop->mo_next;
2128 /* Check for growth of freelist again */
2129 if (freecnt != txn->mt_free_pgs[0])
2132 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2133 if (mdb_midl_shrink(&txn->mt_free_pgs))
2134 env->me_free_pgs = txn->mt_free_pgs;
2141 if (env->me_flags & MDB_WRITEMAP) {
2142 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2143 dp = txn->mt_u.dirty_list[i].mptr;
2144 /* clear dirty flag */
2145 dp->mp_flags &= ~P_DIRTY;
2146 txn->mt_u.dirty_list[i].mid = 0;
2148 txn->mt_u.dirty_list[0].mid = 0;
2152 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2158 /* Windows actually supports scatter/gather I/O, but only on
2159 * unbuffered file handles. Since we're relying on the OS page
2160 * cache for all our data, that's self-defeating. So we just
2161 * write pages one at a time. We use the ov structure to set
2162 * the write offset, to at least save the overhead of a Seek
2166 memset(&ov, 0, sizeof(ov));
2167 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2169 dp = txn->mt_u.dirty_list[i].mptr;
2170 DPRINTF("committing page %zu", dp->mp_pgno);
2171 size = dp->mp_pgno * env->me_psize;
2172 ov.Offset = size & 0xffffffff;
2173 ov.OffsetHigh = size >> 16;
2174 ov.OffsetHigh >>= 16;
2175 /* clear dirty flag */
2176 dp->mp_flags &= ~P_DIRTY;
2177 wsize = env->me_psize;
2178 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2179 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2182 DPRINTF("WriteFile: %d", n);
2189 struct iovec iov[MDB_COMMIT_PAGES];
2193 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2194 dp = txn->mt_u.dirty_list[i].mptr;
2195 if (dp->mp_pgno != next) {
2197 rc = writev(env->me_fd, iov, n);
2201 DPUTS("short write, filesystem full?");
2203 DPRINTF("writev: %s", strerror(n));
2210 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2213 DPRINTF("committing page %zu", dp->mp_pgno);
2214 iov[n].iov_len = env->me_psize;
2215 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2216 iov[n].iov_base = (char *)dp;
2217 size += iov[n].iov_len;
2218 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2219 /* clear dirty flag */
2220 dp->mp_flags &= ~P_DIRTY;
2221 if (++n >= MDB_COMMIT_PAGES) {
2231 rc = writev(env->me_fd, iov, n);
2235 DPUTS("short write, filesystem full?");
2237 DPRINTF("writev: %s", strerror(n));
2244 /* Drop the dirty pages.
2246 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2247 dp = txn->mt_u.dirty_list[i].mptr;
2248 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2249 dp->mp_next = txn->mt_env->me_dpages;
2250 VGMEMP_FREE(txn->mt_env, dp);
2251 txn->mt_env->me_dpages = dp;
2253 VGMEMP_FREE(txn->mt_env, dp);
2256 txn->mt_u.dirty_list[i].mid = 0;
2258 txn->mt_u.dirty_list[0].mid = 0;
2261 if ((n = mdb_env_sync(env, 0)) != 0 ||
2262 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2269 if (txn->mt_numdbs > env->me_numdbs) {
2270 /* update the DB flags */
2272 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2273 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2277 UNLOCK_MUTEX_W(env);
2283 /** Read the environment parameters of a DB environment before
2284 * mapping it into memory.
2285 * @param[in] env the environment handle
2286 * @param[out] meta address of where to store the meta information
2287 * @return 0 on success, non-zero on failure.
2290 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2297 /* We don't know the page size yet, so use a minimum value.
2301 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2303 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2308 else if (rc != MDB_PAGESIZE) {
2312 DPRINTF("read: %s", strerror(err));
2316 p = (MDB_page *)&pbuf;
2318 if (!F_ISSET(p->mp_flags, P_META)) {
2319 DPRINTF("page %zu not a meta page", p->mp_pgno);
2324 if (m->mm_magic != MDB_MAGIC) {
2325 DPUTS("meta has invalid magic");
2329 if (m->mm_version != MDB_VERSION) {
2330 DPRINTF("database is version %u, expected version %u",
2331 m->mm_version, MDB_VERSION);
2332 return MDB_VERSION_MISMATCH;
2335 memcpy(meta, m, sizeof(*m));
2339 /** Write the environment parameters of a freshly created DB environment.
2340 * @param[in] env the environment handle
2341 * @param[out] meta address of where to store the meta information
2342 * @return 0 on success, non-zero on failure.
2345 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2352 DPUTS("writing new meta page");
2354 GET_PAGESIZE(psize);
2356 meta->mm_magic = MDB_MAGIC;
2357 meta->mm_version = MDB_VERSION;
2358 meta->mm_psize = psize;
2359 meta->mm_last_pg = 1;
2360 meta->mm_flags = env->me_flags & 0xffff;
2361 meta->mm_flags |= MDB_INTEGERKEY;
2362 meta->mm_dbs[0].md_root = P_INVALID;
2363 meta->mm_dbs[1].md_root = P_INVALID;
2365 p = calloc(2, psize);
2367 p->mp_flags = P_META;
2370 memcpy(m, meta, sizeof(*meta));
2372 q = (MDB_page *)((char *)p + psize);
2375 q->mp_flags = P_META;
2378 memcpy(m, meta, sizeof(*meta));
2383 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2384 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2387 rc = write(env->me_fd, p, psize * 2);
2388 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2394 /** Update the environment info to commit a transaction.
2395 * @param[in] txn the transaction that's being committed
2396 * @return 0 on success, non-zero on failure.
2399 mdb_env_write_meta(MDB_txn *txn)
2402 MDB_meta meta, metab;
2404 int rc, len, toggle;
2410 assert(txn != NULL);
2411 assert(txn->mt_env != NULL);
2413 toggle = !txn->mt_toggle;
2414 DPRINTF("writing meta page %d for root page %zu",
2415 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2419 if (env->me_flags & MDB_WRITEMAP) {
2420 MDB_meta *mp = env->me_metas[toggle];
2421 mp->mm_dbs[0] = txn->mt_dbs[0];
2422 mp->mm_dbs[1] = txn->mt_dbs[1];
2423 mp->mm_last_pg = txn->mt_next_pgno - 1;
2424 mp->mm_txnid = txn->mt_txnid;
2425 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2426 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2429 ptr += env->me_psize;
2430 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2437 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2438 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2440 ptr = (char *)&meta;
2441 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2442 len = sizeof(MDB_meta) - off;
2445 meta.mm_dbs[0] = txn->mt_dbs[0];
2446 meta.mm_dbs[1] = txn->mt_dbs[1];
2447 meta.mm_last_pg = txn->mt_next_pgno - 1;
2448 meta.mm_txnid = txn->mt_txnid;
2451 off += env->me_psize;
2454 /* Write to the SYNC fd */
2457 memset(&ov, 0, sizeof(ov));
2459 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2462 rc = pwrite(env->me_mfd, ptr, len, off);
2467 DPUTS("write failed, disk error?");
2468 /* On a failure, the pagecache still contains the new data.
2469 * Write some old data back, to prevent it from being used.
2470 * Use the non-SYNC fd; we know it will fail anyway.
2472 meta.mm_last_pg = metab.mm_last_pg;
2473 meta.mm_txnid = metab.mm_txnid;
2475 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2477 r2 = pwrite(env->me_fd, ptr, len, off);
2480 env->me_flags |= MDB_FATAL_ERROR;
2484 /* Memory ordering issues are irrelevant; since the entire writer
2485 * is wrapped by wmutex, all of these changes will become visible
2486 * after the wmutex is unlocked. Since the DB is multi-version,
2487 * readers will get consistent data regardless of how fresh or
2488 * how stale their view of these values is.
2490 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2495 /** Check both meta pages to see which one is newer.
2496 * @param[in] env the environment handle
2497 * @return meta toggle (0 or 1).
2500 mdb_env_pick_meta(const MDB_env *env)
2502 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2506 mdb_env_create(MDB_env **env)
2510 e = calloc(1, sizeof(MDB_env));
2514 e->me_free_pgs = mdb_midl_alloc();
2515 if (!e->me_free_pgs) {
2519 e->me_maxreaders = DEFAULT_READERS;
2521 e->me_fd = INVALID_HANDLE_VALUE;
2522 e->me_lfd = INVALID_HANDLE_VALUE;
2523 e->me_mfd = INVALID_HANDLE_VALUE;
2524 e->me_pid = getpid();
2525 VGMEMP_CREATE(e,0,0);
2531 mdb_env_set_mapsize(MDB_env *env, size_t size)
2535 env->me_mapsize = size;
2537 env->me_maxpg = env->me_mapsize / env->me_psize;
2542 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2546 env->me_maxdbs = dbs;
2551 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2553 if (env->me_map || readers < 1)
2555 env->me_maxreaders = readers;
2560 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2562 if (!env || !readers)
2564 *readers = env->me_maxreaders;
2568 /** Further setup required for opening an MDB environment
2571 mdb_env_open2(MDB_env *env, unsigned int flags)
2573 int i, newenv = 0, prot;
2577 env->me_flags = flags;
2579 memset(&meta, 0, sizeof(meta));
2581 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2584 DPUTS("new mdbenv");
2588 if (!env->me_mapsize) {
2589 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2595 LONG sizelo, sizehi;
2596 sizelo = env->me_mapsize & 0xffffffff;
2597 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2599 /* Windows won't create mappings for zero length files.
2600 * Just allocate the maxsize right now.
2603 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2604 if (!SetEndOfFile(env->me_fd))
2606 SetFilePointer(env->me_fd, 0, NULL, 0);
2608 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2609 PAGE_READWRITE : PAGE_READONLY,
2610 sizehi, sizelo, NULL);
2613 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2614 FILE_MAP_WRITE : FILE_MAP_READ,
2615 0, 0, env->me_mapsize, meta.mm_address);
2622 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2625 if (flags & MDB_WRITEMAP) {
2627 ftruncate(env->me_fd, env->me_mapsize);
2629 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2631 if (env->me_map == MAP_FAILED) {
2638 meta.mm_mapsize = env->me_mapsize;
2639 if (flags & MDB_FIXEDMAP)
2640 meta.mm_address = env->me_map;
2641 i = mdb_env_init_meta(env, &meta);
2642 if (i != MDB_SUCCESS) {
2643 munmap(env->me_map, env->me_mapsize);
2647 env->me_psize = meta.mm_psize;
2649 env->me_maxpg = env->me_mapsize / env->me_psize;
2651 p = (MDB_page *)env->me_map;
2652 env->me_metas[0] = METADATA(p);
2653 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2657 int toggle = mdb_env_pick_meta(env);
2658 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2660 DPRINTF("opened database version %u, pagesize %u",
2661 env->me_metas[0]->mm_version, env->me_psize);
2662 DPRINTF("using meta page %d", toggle);
2663 DPRINTF("depth: %u", db->md_depth);
2664 DPRINTF("entries: %zu", db->md_entries);
2665 DPRINTF("branch pages: %zu", db->md_branch_pages);
2666 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2667 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2668 DPRINTF("root: %zu", db->md_root);
2676 /** Release a reader thread's slot in the reader lock table.
2677 * This function is called automatically when a thread exits.
2678 * @param[in] ptr This points to the slot in the reader lock table.
2681 mdb_env_reader_dest(void *ptr)
2683 MDB_reader *reader = ptr;
2689 /** Junk for arranging thread-specific callbacks on Windows. This is
2690 * necessarily platform and compiler-specific. Windows supports up
2691 * to 1088 keys. Let's assume nobody opens more than 64 environments
2692 * in a single process, for now. They can override this if needed.
2694 #ifndef MAX_TLS_KEYS
2695 #define MAX_TLS_KEYS 64
2697 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2698 static int mdb_tls_nkeys;
2700 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2704 case DLL_PROCESS_ATTACH: break;
2705 case DLL_THREAD_ATTACH: break;
2706 case DLL_THREAD_DETACH:
2707 for (i=0; i<mdb_tls_nkeys; i++) {
2708 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2709 mdb_env_reader_dest(r);
2712 case DLL_PROCESS_DETACH: break;
2717 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2719 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2723 /* Force some symbol references.
2724 * _tls_used forces the linker to create the TLS directory if not already done
2725 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2727 #pragma comment(linker, "/INCLUDE:_tls_used")
2728 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2729 #pragma const_seg(".CRT$XLB")
2730 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2731 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2734 #pragma comment(linker, "/INCLUDE:__tls_used")
2735 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2736 #pragma data_seg(".CRT$XLB")
2737 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2739 #endif /* WIN 32/64 */
2740 #endif /* !__GNUC__ */
2743 /** Downgrade the exclusive lock on the region back to shared */
2745 mdb_env_share_locks(MDB_env *env)
2747 int rc = 0, toggle = mdb_env_pick_meta(env);
2749 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2754 /* First acquire a shared lock. The Unlock will
2755 * then release the existing exclusive lock.
2757 memset(&ov, 0, sizeof(ov));
2758 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2759 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2763 struct flock lock_info;
2764 /* The shared lock replaces the existing lock */
2765 memset((void *)&lock_info, 0, sizeof(lock_info));
2766 lock_info.l_type = F_RDLCK;
2767 lock_info.l_whence = SEEK_SET;
2768 lock_info.l_start = 0;
2769 lock_info.l_len = 1;
2770 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2771 (rc = ErrCode()) == EINTR) ;
2778 /** Try to get exlusive lock, otherwise shared.
2779 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
2782 mdb_env_excl_lock(MDB_env *env, int *excl)
2786 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2790 memset(&ov, 0, sizeof(ov));
2791 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2796 struct flock lock_info;
2797 memset((void *)&lock_info, 0, sizeof(lock_info));
2798 lock_info.l_type = F_WRLCK;
2799 lock_info.l_whence = SEEK_SET;
2800 lock_info.l_start = 0;
2801 lock_info.l_len = 1;
2802 if (!fcntl(env->me_lfd, F_SETLK, &lock_info)) {
2805 # ifdef MDB_USE_POSIX_SEM
2806 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
2809 lock_info.l_type = F_RDLCK;
2810 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
2811 (rc = ErrCode()) == EINTR) ;
2817 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
2819 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2821 * @(#) $Revision: 5.1 $
2822 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2823 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2825 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2829 * Please do not copyright this code. This code is in the public domain.
2831 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2832 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2833 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2834 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2835 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2836 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2837 * PERFORMANCE OF THIS SOFTWARE.
2840 * chongo <Landon Curt Noll> /\oo/\
2841 * http://www.isthe.com/chongo/
2843 * Share and Enjoy! :-)
2846 typedef unsigned long long mdb_hash_t;
2847 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2849 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2850 * @param[in] str string to hash
2851 * @param[in] hval initial value for hash
2852 * @return 64 bit hash
2854 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2855 * hval arg on the first call.
2858 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
2860 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
2861 unsigned char *end = s + val->mv_size;
2863 * FNV-1a hash each octet of the string
2866 /* xor the bottom with the current octet */
2867 hval ^= (mdb_hash_t)*s++;
2869 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2870 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2871 (hval << 7) + (hval << 8) + (hval << 40);
2873 /* return our new hash value */
2877 /** Hash the string and output the hash in hex.
2878 * @param[in] str string to hash
2879 * @param[out] hexbuf an array of 17 chars to hold the hash
2882 mdb_hash_hex(MDB_val *val, char *hexbuf)
2885 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
2886 for (i=0; i<8; i++) {
2887 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2893 /** Open and/or initialize the lock region for the environment.
2894 * @param[in] env The MDB environment.
2895 * @param[in] lpath The pathname of the file used for the lock region.
2896 * @param[in] mode The Unix permissions for the file, if we create it.
2897 * @param[out] excl Set to true if we got an exclusive lock on the region.
2898 * @return 0 on success, non-zero on failure.
2901 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2909 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2910 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2911 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2915 /* Try to get exclusive lock. If we succeed, then
2916 * nobody is using the lock region and we should initialize it.
2918 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2919 size = GetFileSize(env->me_lfd, NULL);
2925 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1)
2927 /* Lose record locks when exec*() */
2928 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
2929 fcntl(env->me_lfd, F_SETFD, fdflags);
2931 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
2932 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1)
2936 /* Try to get exclusive lock. If we succeed, then
2937 * nobody is using the lock region and we should initialize it.
2939 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2941 size = lseek(env->me_lfd, 0, SEEK_END);
2943 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2944 if (size < rsize && *excl > 0) {
2946 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2947 if (!SetEndOfFile(env->me_lfd)) {
2952 if (ftruncate(env->me_lfd, rsize) != 0) {
2959 size = rsize - sizeof(MDB_txninfo);
2960 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2965 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2971 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2973 if (!env->me_txns) {
2978 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2980 if (m == MAP_FAILED) {
2981 env->me_txns = NULL;
2990 BY_HANDLE_FILE_INFORMATION stbuf;
2999 if (!mdb_sec_inited) {
3000 InitializeSecurityDescriptor(&mdb_null_sd,
3001 SECURITY_DESCRIPTOR_REVISION);
3002 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3003 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3004 mdb_all_sa.bInheritHandle = FALSE;
3005 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3008 GetFileInformationByHandle(env->me_lfd, &stbuf);
3009 idbuf.volume = stbuf.dwVolumeSerialNumber;
3010 idbuf.nhigh = stbuf.nFileIndexHigh;
3011 idbuf.nlow = stbuf.nFileIndexLow;
3012 val.mv_data = &idbuf;
3013 val.mv_size = sizeof(idbuf);
3014 mdb_hash_hex(&val, hexbuf);
3015 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3016 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3017 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3018 if (!env->me_rmutex) {
3022 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3023 if (!env->me_wmutex) {
3027 #elif defined(MDB_USE_POSIX_SEM)
3036 fstat(env->me_lfd, &stbuf);
3037 idbuf.dev = stbuf.st_dev;
3038 idbuf.ino = stbuf.st_ino;
3039 val.mv_data = &idbuf;
3040 val.mv_size = sizeof(idbuf);
3041 mdb_hash_hex(&val, hexbuf);
3042 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3043 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3044 /* Clean up after a previous run, if needed: Try to
3045 * remove both semaphores before doing anything else.
3047 sem_unlink(env->me_txns->mti_rmname);
3048 sem_unlink(env->me_txns->mti_wmname);
3049 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3050 O_CREAT|O_EXCL, mode, 1);
3051 if (env->me_rmutex == SEM_FAILED) {
3055 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3056 O_CREAT|O_EXCL, mode, 1);
3057 if (env->me_wmutex == SEM_FAILED) {
3061 #else /* MDB_USE_POSIX_SEM */
3062 pthread_mutexattr_t mattr;
3064 pthread_mutexattr_init(&mattr);
3065 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
3069 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
3070 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
3071 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3072 env->me_txns->mti_version = MDB_VERSION;
3073 env->me_txns->mti_magic = MDB_MAGIC;
3074 env->me_txns->mti_txnid = 0;
3075 env->me_txns->mti_numreaders = 0;
3078 if (env->me_txns->mti_magic != MDB_MAGIC) {
3079 DPUTS("lock region has invalid magic");
3083 if (env->me_txns->mti_version != MDB_VERSION) {
3084 DPRINTF("lock region is version %u, expected version %u",
3085 env->me_txns->mti_version, MDB_VERSION);
3086 rc = MDB_VERSION_MISMATCH;
3090 if (rc != EACCES && rc != EAGAIN) {
3094 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3095 if (!env->me_rmutex) {
3099 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3100 if (!env->me_wmutex) {
3104 #elif defined(MDB_USE_POSIX_SEM)
3105 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3106 if (env->me_rmutex == SEM_FAILED) {
3110 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3111 if (env->me_wmutex == SEM_FAILED) {
3121 env->me_lfd = INVALID_HANDLE_VALUE;
3126 /** The name of the lock file in the DB environment */
3127 #define LOCKNAME "/lock.mdb"
3128 /** The name of the data file in the DB environment */
3129 #define DATANAME "/data.mdb"
3130 /** The suffix of the lock file when no subdir is used */
3131 #define LOCKSUFF "-lock"
3134 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3136 int oflags, rc, len, excl;
3137 char *lpath, *dpath;
3140 if (flags & MDB_NOSUBDIR) {
3141 rc = len + sizeof(LOCKSUFF) + len + 1;
3143 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3148 if (flags & MDB_NOSUBDIR) {
3149 dpath = lpath + len + sizeof(LOCKSUFF);
3150 sprintf(lpath, "%s" LOCKSUFF, path);
3151 strcpy(dpath, path);
3153 dpath = lpath + len + sizeof(LOCKNAME);
3154 sprintf(lpath, "%s" LOCKNAME, path);
3155 sprintf(dpath, "%s" DATANAME, path);
3158 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3163 if (F_ISSET(flags, MDB_RDONLY)) {
3164 oflags = GENERIC_READ;
3165 len = OPEN_EXISTING;
3167 oflags = GENERIC_READ|GENERIC_WRITE;
3170 mode = FILE_ATTRIBUTE_NORMAL;
3171 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3172 NULL, len, mode, NULL);
3174 if (F_ISSET(flags, MDB_RDONLY))
3177 oflags = O_RDWR | O_CREAT;
3179 env->me_fd = open(dpath, oflags, mode);
3181 if (env->me_fd == INVALID_HANDLE_VALUE) {
3186 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
3187 if (flags & (MDB_RDONLY|MDB_NOSYNC|MDB_NOMETASYNC|MDB_WRITEMAP)) {
3188 env->me_mfd = env->me_fd;
3190 /* synchronous fd for meta writes */
3192 env->me_mfd = CreateFile(dpath, oflags,
3193 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3194 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3196 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3198 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3203 env->me_path = strdup(path);
3204 DPRINTF("opened dbenv %p", (void *) env);
3205 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3207 /* Windows TLS callbacks need help finding their TLS info. */
3208 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3209 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3216 rc = mdb_env_share_locks(env);
3221 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3222 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3223 if (!env->me_dbxs || !env->me_dbflags)
3229 if (env->me_fd != INVALID_HANDLE_VALUE) {
3231 env->me_fd = INVALID_HANDLE_VALUE;
3233 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3235 env->me_lfd = INVALID_HANDLE_VALUE;
3243 mdb_env_close(MDB_env *env)
3250 VGMEMP_DESTROY(env);
3251 while (env->me_dpages) {
3252 dp = env->me_dpages;
3253 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3254 env->me_dpages = dp->mp_next;
3258 free(env->me_dbflags);
3262 pthread_key_delete(env->me_txkey);
3264 /* Delete our key from the global list */
3266 for (i=0; i<mdb_tls_nkeys; i++)
3267 if (mdb_tls_keys[i] == env->me_txkey) {
3268 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3276 munmap(env->me_map, env->me_mapsize);
3278 if (env->me_mfd != env->me_fd)
3282 pid_t pid = env->me_pid;
3284 for (i=0; i<env->me_numreaders; i++)
3285 if (env->me_txns->mti_readers[i].mr_pid == pid)
3286 env->me_txns->mti_readers[i].mr_pid = 0;
3288 CloseHandle(env->me_rmutex);
3289 CloseHandle(env->me_wmutex);
3290 /* Windows automatically destroys the mutexes when
3291 * the last handle closes.
3293 #elif defined(MDB_USE_POSIX_SEM)
3294 sem_close(env->me_rmutex);
3295 sem_close(env->me_wmutex);
3297 if (!mdb_env_excl_lock(env, &excl) && excl) {
3298 /* we are the only remaining user of the environment.
3299 clean up semaphores. */
3300 sem_unlink(env->me_txns->mti_rmname);
3301 sem_unlink(env->me_txns->mti_wmname);
3305 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3308 mdb_midl_free(env->me_free_pgs);
3312 /** Compare two items pointing at aligned size_t's */
3314 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3316 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3317 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3320 /** Compare two items pointing at aligned int's */
3322 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3324 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3325 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3328 /** Compare two items pointing at ints of unknown alignment.
3329 * Nodes and keys are guaranteed to be 2-byte aligned.
3332 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3334 #if BYTE_ORDER == LITTLE_ENDIAN
3335 unsigned short *u, *c;
3338 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3339 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3342 } while(!x && u > (unsigned short *)a->mv_data);
3345 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3349 /** Compare two items lexically */
3351 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3358 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3364 diff = memcmp(a->mv_data, b->mv_data, len);
3365 return diff ? diff : len_diff<0 ? -1 : len_diff;
3368 /** Compare two items in reverse byte order */
3370 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3372 const unsigned char *p1, *p2, *p1_lim;
3376 p1_lim = (const unsigned char *)a->mv_data;
3377 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3378 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3380 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3386 while (p1 > p1_lim) {
3387 diff = *--p1 - *--p2;
3391 return len_diff<0 ? -1 : len_diff;
3394 /** Search for key within a page, using binary search.
3395 * Returns the smallest entry larger or equal to the key.
3396 * If exactp is non-null, stores whether the found entry was an exact match
3397 * in *exactp (1 or 0).
3398 * Updates the cursor index with the index of the found entry.
3399 * If no entry larger or equal to the key is found, returns NULL.
3402 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3404 unsigned int i = 0, nkeys;
3407 MDB_page *mp = mc->mc_pg[mc->mc_top];
3408 MDB_node *node = NULL;
3413 nkeys = NUMKEYS(mp);
3418 COPY_PGNO(pgno, mp->mp_pgno);
3419 DPRINTF("searching %u keys in %s %spage %zu",
3420 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3427 low = IS_LEAF(mp) ? 0 : 1;
3429 cmp = mc->mc_dbx->md_cmp;
3431 /* Branch pages have no data, so if using integer keys,
3432 * alignment is guaranteed. Use faster mdb_cmp_int.
3434 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3435 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3442 nodekey.mv_size = mc->mc_db->md_pad;
3443 node = NODEPTR(mp, 0); /* fake */
3444 while (low <= high) {
3445 i = (low + high) >> 1;
3446 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3447 rc = cmp(key, &nodekey);
3448 DPRINTF("found leaf index %u [%s], rc = %i",
3449 i, DKEY(&nodekey), rc);
3458 while (low <= high) {
3459 i = (low + high) >> 1;
3461 node = NODEPTR(mp, i);
3462 nodekey.mv_size = NODEKSZ(node);
3463 nodekey.mv_data = NODEKEY(node);
3465 rc = cmp(key, &nodekey);
3468 DPRINTF("found leaf index %u [%s], rc = %i",
3469 i, DKEY(&nodekey), rc);
3471 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3472 i, DKEY(&nodekey), NODEPGNO(node), rc);
3483 if (rc > 0) { /* Found entry is less than the key. */
3484 i++; /* Skip to get the smallest entry larger than key. */
3486 node = NODEPTR(mp, i);
3489 *exactp = (rc == 0);
3490 /* store the key index */
3491 mc->mc_ki[mc->mc_top] = i;
3493 /* There is no entry larger or equal to the key. */
3496 /* nodeptr is fake for LEAF2 */
3502 mdb_cursor_adjust(MDB_cursor *mc, func)
3506 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3507 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3514 /** Pop a page off the top of the cursor's stack. */
3516 mdb_cursor_pop(MDB_cursor *mc)
3520 MDB_page *top = mc->mc_pg[mc->mc_top];
3526 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3527 mc->mc_dbi, (void *) mc);
3531 /** Push a page onto the top of the cursor's stack. */
3533 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3535 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3536 mc->mc_dbi, (void *) mc);
3538 if (mc->mc_snum >= CURSOR_STACK) {
3539 assert(mc->mc_snum < CURSOR_STACK);
3540 return MDB_CURSOR_FULL;
3543 mc->mc_top = mc->mc_snum++;
3544 mc->mc_pg[mc->mc_top] = mp;
3545 mc->mc_ki[mc->mc_top] = 0;
3550 /** Find the address of the page corresponding to a given page number.
3551 * @param[in] txn the transaction for this access.
3552 * @param[in] pgno the page number for the page to retrieve.
3553 * @param[out] ret address of a pointer where the page's address will be stored.
3554 * @return 0 on success, non-zero on failure.
3557 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3561 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
3562 if (pgno < txn->mt_next_pgno)
3563 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3566 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3568 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3569 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3570 p = txn->mt_u.dirty_list[x].mptr;
3574 if (pgno < txn->mt_next_pgno)
3575 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3580 DPRINTF("page %zu not found", pgno);
3583 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3586 /** Search for the page a given key should be in.
3587 * Pushes parent pages on the cursor stack. This function continues a
3588 * search on a cursor that has already been initialized. (Usually by
3589 * #mdb_page_search() but also by #mdb_node_move().)
3590 * @param[in,out] mc the cursor for this operation.
3591 * @param[in] key the key to search for. If NULL, search for the lowest
3592 * page. (This is used by #mdb_cursor_first().)
3593 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3594 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3595 * @return 0 on success, non-zero on failure.
3598 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3600 MDB_page *mp = mc->mc_pg[mc->mc_top];
3605 while (IS_BRANCH(mp)) {
3609 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3610 assert(NUMKEYS(mp) > 1);
3611 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3613 if (key == NULL) /* Initialize cursor to first page. */
3615 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3616 /* cursor to last page */
3620 node = mdb_node_search(mc, key, &exact);
3622 i = NUMKEYS(mp) - 1;
3624 i = mc->mc_ki[mc->mc_top];
3633 DPRINTF("following index %u for key [%s]",
3635 assert(i < NUMKEYS(mp));
3636 node = NODEPTR(mp, i);
3638 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3641 mc->mc_ki[mc->mc_top] = i;
3642 if ((rc = mdb_cursor_push(mc, mp)))
3646 if ((rc = mdb_page_touch(mc)) != 0)
3648 mp = mc->mc_pg[mc->mc_top];
3653 DPRINTF("internal error, index points to a %02X page!?",
3655 return MDB_CORRUPTED;
3658 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3659 key ? DKEY(key) : NULL);
3664 /** Search for the page a given key should be in.
3665 * Pushes parent pages on the cursor stack. This function just sets up
3666 * the search; it finds the root page for \b mc's database and sets this
3667 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3668 * called to complete the search.
3669 * @param[in,out] mc the cursor for this operation.
3670 * @param[in] key the key to search for. If NULL, search for the lowest
3671 * page. (This is used by #mdb_cursor_first().)
3672 * @param[in] modify If true, visited pages are updated with new page numbers.
3673 * @return 0 on success, non-zero on failure.
3676 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
3681 /* Make sure the txn is still viable, then find the root from
3682 * the txn's db table.
3684 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3685 DPUTS("transaction has failed, must abort");
3688 /* Make sure we're using an up-to-date root */
3689 if (mc->mc_dbi > MAIN_DBI) {
3690 if ((*mc->mc_dbflag & DB_STALE) ||
3691 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
3693 unsigned char dbflag = 0;
3694 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3695 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
3698 if (*mc->mc_dbflag & DB_STALE) {
3701 MDB_node *leaf = mdb_node_search(&mc2,
3702 &mc->mc_dbx->md_name, &exact);
3704 return MDB_NOTFOUND;
3705 mdb_node_read(mc->mc_txn, leaf, &data);
3706 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3708 if (flags & MDB_PS_MODIFY)
3710 *mc->mc_dbflag = dbflag;
3713 root = mc->mc_db->md_root;
3715 if (root == P_INVALID) { /* Tree is empty. */
3716 DPUTS("tree is empty");
3717 return MDB_NOTFOUND;
3722 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
3723 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3729 DPRINTF("db %u root page %zu has flags 0x%X",
3730 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3732 if (flags & MDB_PS_MODIFY) {
3733 if ((rc = mdb_page_touch(mc)))
3737 if (flags & MDB_PS_ROOTONLY)
3740 return mdb_page_search_root(mc, key, flags);
3743 /** Return the data associated with a given node.
3744 * @param[in] txn The transaction for this operation.
3745 * @param[in] leaf The node being read.
3746 * @param[out] data Updated to point to the node's data.
3747 * @return 0 on success, non-zero on failure.
3750 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3752 MDB_page *omp; /* overflow page */
3756 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3757 data->mv_size = NODEDSZ(leaf);
3758 data->mv_data = NODEDATA(leaf);
3762 /* Read overflow data.
3764 data->mv_size = NODEDSZ(leaf);
3765 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3766 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3767 DPRINTF("read overflow page %zu failed", pgno);
3770 data->mv_data = METADATA(omp);
3776 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3777 MDB_val *key, MDB_val *data)
3786 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3788 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3791 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3795 mdb_cursor_init(&mc, txn, dbi, &mx);
3796 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3799 /** Find a sibling for a page.
3800 * Replaces the page at the top of the cursor's stack with the
3801 * specified sibling, if one exists.
3802 * @param[in] mc The cursor for this operation.
3803 * @param[in] move_right Non-zero if the right sibling is requested,
3804 * otherwise the left sibling.
3805 * @return 0 on success, non-zero on failure.
3808 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3814 if (mc->mc_snum < 2) {
3815 return MDB_NOTFOUND; /* root has no siblings */
3819 DPRINTF("parent page is page %zu, index %u",
3820 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3822 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3823 : (mc->mc_ki[mc->mc_top] == 0)) {
3824 DPRINTF("no more keys left, moving to %s sibling",
3825 move_right ? "right" : "left");
3826 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3830 mc->mc_ki[mc->mc_top]++;
3832 mc->mc_ki[mc->mc_top]--;
3833 DPRINTF("just moving to %s index key %u",
3834 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3836 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3838 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3839 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3842 mdb_cursor_push(mc, mp);
3847 /** Move the cursor to the next data item. */
3849 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3855 if (mc->mc_flags & C_EOF) {
3856 return MDB_NOTFOUND;
3859 assert(mc->mc_flags & C_INITIALIZED);
3861 mp = mc->mc_pg[mc->mc_top];
3863 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3864 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3865 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3866 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3867 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3868 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3872 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3873 if (op == MDB_NEXT_DUP)
3874 return MDB_NOTFOUND;
3878 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3880 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3881 DPUTS("=====> move to next sibling page");
3882 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3883 mc->mc_flags |= C_EOF;
3884 mc->mc_flags &= ~C_INITIALIZED;
3885 return MDB_NOTFOUND;
3887 mp = mc->mc_pg[mc->mc_top];
3888 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3890 mc->mc_ki[mc->mc_top]++;
3892 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3893 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3896 key->mv_size = mc->mc_db->md_pad;
3897 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3901 assert(IS_LEAF(mp));
3902 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3904 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3905 mdb_xcursor_init1(mc, leaf);
3908 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3911 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3912 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3913 if (rc != MDB_SUCCESS)
3918 MDB_GET_KEY(leaf, key);
3922 /** Move the cursor to the previous data item. */
3924 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3930 assert(mc->mc_flags & C_INITIALIZED);
3932 mp = mc->mc_pg[mc->mc_top];
3934 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3935 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3936 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3937 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3938 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3939 if (op != MDB_PREV || rc == MDB_SUCCESS)
3942 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3943 if (op == MDB_PREV_DUP)
3944 return MDB_NOTFOUND;
3949 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3951 if (mc->mc_ki[mc->mc_top] == 0) {
3952 DPUTS("=====> move to prev sibling page");
3953 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3954 mc->mc_flags &= ~C_INITIALIZED;
3955 return MDB_NOTFOUND;
3957 mp = mc->mc_pg[mc->mc_top];
3958 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3959 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3961 mc->mc_ki[mc->mc_top]--;
3963 mc->mc_flags &= ~C_EOF;
3965 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3966 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3969 key->mv_size = mc->mc_db->md_pad;
3970 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3974 assert(IS_LEAF(mp));
3975 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3977 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3978 mdb_xcursor_init1(mc, leaf);
3981 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3984 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3985 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3986 if (rc != MDB_SUCCESS)
3991 MDB_GET_KEY(leaf, key);
3995 /** Set the cursor on a specific data item. */
3997 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3998 MDB_cursor_op op, int *exactp)
4002 MDB_node *leaf = NULL;
4007 assert(key->mv_size > 0);
4009 /* See if we're already on the right page */
4010 if (mc->mc_flags & C_INITIALIZED) {
4013 mp = mc->mc_pg[mc->mc_top];
4015 mc->mc_ki[mc->mc_top] = 0;
4016 return MDB_NOTFOUND;
4018 if (mp->mp_flags & P_LEAF2) {
4019 nodekey.mv_size = mc->mc_db->md_pad;
4020 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4022 leaf = NODEPTR(mp, 0);
4023 MDB_GET_KEY(leaf, &nodekey);
4025 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4027 /* Probably happens rarely, but first node on the page
4028 * was the one we wanted.
4030 mc->mc_ki[mc->mc_top] = 0;
4037 unsigned int nkeys = NUMKEYS(mp);
4039 if (mp->mp_flags & P_LEAF2) {
4040 nodekey.mv_data = LEAF2KEY(mp,
4041 nkeys-1, nodekey.mv_size);
4043 leaf = NODEPTR(mp, nkeys-1);
4044 MDB_GET_KEY(leaf, &nodekey);
4046 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4048 /* last node was the one we wanted */
4049 mc->mc_ki[mc->mc_top] = nkeys-1;
4055 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4056 /* This is definitely the right page, skip search_page */
4057 if (mp->mp_flags & P_LEAF2) {
4058 nodekey.mv_data = LEAF2KEY(mp,
4059 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4061 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4062 MDB_GET_KEY(leaf, &nodekey);
4064 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4066 /* current node was the one we wanted */
4076 /* If any parents have right-sibs, search.
4077 * Otherwise, there's nothing further.
4079 for (i=0; i<mc->mc_top; i++)
4081 NUMKEYS(mc->mc_pg[i])-1)
4083 if (i == mc->mc_top) {
4084 /* There are no other pages */
4085 mc->mc_ki[mc->mc_top] = nkeys;
4086 return MDB_NOTFOUND;
4090 /* There are no other pages */
4091 mc->mc_ki[mc->mc_top] = 0;
4092 return MDB_NOTFOUND;
4096 rc = mdb_page_search(mc, key, 0);
4097 if (rc != MDB_SUCCESS)
4100 mp = mc->mc_pg[mc->mc_top];
4101 assert(IS_LEAF(mp));
4104 leaf = mdb_node_search(mc, key, exactp);
4105 if (exactp != NULL && !*exactp) {
4106 /* MDB_SET specified and not an exact match. */
4107 return MDB_NOTFOUND;
4111 DPUTS("===> inexact leaf not found, goto sibling");
4112 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4113 return rc; /* no entries matched */
4114 mp = mc->mc_pg[mc->mc_top];
4115 assert(IS_LEAF(mp));
4116 leaf = NODEPTR(mp, 0);
4120 mc->mc_flags |= C_INITIALIZED;
4121 mc->mc_flags &= ~C_EOF;
4124 key->mv_size = mc->mc_db->md_pad;
4125 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4129 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4130 mdb_xcursor_init1(mc, leaf);
4133 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4134 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4135 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4138 if (op == MDB_GET_BOTH) {
4144 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4145 if (rc != MDB_SUCCESS)
4148 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4150 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4152 rc = mc->mc_dbx->md_dcmp(data, &d2);
4154 if (op == MDB_GET_BOTH || rc > 0)
4155 return MDB_NOTFOUND;
4160 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4161 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4166 /* The key already matches in all other cases */
4167 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4168 MDB_GET_KEY(leaf, key);
4169 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4174 /** Move the cursor to the first item in the database. */
4176 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4181 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4182 rc = mdb_page_search(mc, NULL, 0);
4183 if (rc != MDB_SUCCESS)
4186 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4188 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4189 mc->mc_flags |= C_INITIALIZED;
4190 mc->mc_flags &= ~C_EOF;
4192 mc->mc_ki[mc->mc_top] = 0;
4194 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4195 key->mv_size = mc->mc_db->md_pad;
4196 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4201 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4202 mdb_xcursor_init1(mc, leaf);
4203 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4208 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4209 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4213 MDB_GET_KEY(leaf, key);
4217 /** Move the cursor to the last item in the database. */
4219 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4224 if (!(mc->mc_flags & C_EOF)) {
4226 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4229 lkey.mv_size = MAXKEYSIZE+1;
4230 lkey.mv_data = NULL;
4231 rc = mdb_page_search(mc, &lkey, 0);
4232 if (rc != MDB_SUCCESS)
4235 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4237 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4238 mc->mc_flags |= C_INITIALIZED|C_EOF;
4240 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4242 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4243 key->mv_size = mc->mc_db->md_pad;
4244 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4249 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4250 mdb_xcursor_init1(mc, leaf);
4251 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4256 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4257 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4262 MDB_GET_KEY(leaf, key);
4267 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4277 case MDB_GET_BOTH_RANGE:
4278 if (data == NULL || mc->mc_xcursor == NULL) {
4286 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4288 } else if (op == MDB_SET_RANGE)
4289 rc = mdb_cursor_set(mc, key, data, op, NULL);
4291 rc = mdb_cursor_set(mc, key, data, op, &exact);
4293 case MDB_GET_MULTIPLE:
4295 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4296 !(mc->mc_flags & C_INITIALIZED)) {
4301 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4302 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4305 case MDB_NEXT_MULTIPLE:
4307 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4311 if (!(mc->mc_flags & C_INITIALIZED))
4312 rc = mdb_cursor_first(mc, key, data);
4314 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4315 if (rc == MDB_SUCCESS) {
4316 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4319 mx = &mc->mc_xcursor->mx_cursor;
4320 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4322 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4323 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4331 case MDB_NEXT_NODUP:
4332 if (!(mc->mc_flags & C_INITIALIZED))
4333 rc = mdb_cursor_first(mc, key, data);
4335 rc = mdb_cursor_next(mc, key, data, op);
4339 case MDB_PREV_NODUP:
4340 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4341 rc = mdb_cursor_last(mc, key, data);
4342 mc->mc_flags &= ~C_EOF;
4344 rc = mdb_cursor_prev(mc, key, data, op);
4347 rc = mdb_cursor_first(mc, key, data);
4351 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4352 !(mc->mc_flags & C_INITIALIZED) ||
4353 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4357 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4360 rc = mdb_cursor_last(mc, key, data);
4364 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4365 !(mc->mc_flags & C_INITIALIZED) ||
4366 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4370 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4373 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4381 /** Touch all the pages in the cursor stack.
4382 * Makes sure all the pages are writable, before attempting a write operation.
4383 * @param[in] mc The cursor to operate on.
4386 mdb_cursor_touch(MDB_cursor *mc)
4390 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4392 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4393 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4396 *mc->mc_dbflag = DB_DIRTY;
4398 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4399 rc = mdb_page_touch(mc);
4403 mc->mc_top = mc->mc_snum-1;
4408 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4411 MDB_node *leaf = NULL;
4412 MDB_val xdata, *rdata, dkey;
4415 int do_sub = 0, insert = 0;
4416 unsigned int mcount = 0;
4420 char dbuf[MAXKEYSIZE+1];
4421 unsigned int nflags;
4424 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4427 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4428 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4432 if (flags == MDB_CURRENT) {
4433 if (!(mc->mc_flags & C_INITIALIZED))
4436 } else if (mc->mc_db->md_root == P_INVALID) {
4438 /* new database, write a root leaf page */
4439 DPUTS("allocating new root leaf page");
4440 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4444 mdb_cursor_push(mc, np);
4445 mc->mc_db->md_root = np->mp_pgno;
4446 mc->mc_db->md_depth++;
4447 *mc->mc_dbflag = DB_DIRTY;
4448 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4450 np->mp_flags |= P_LEAF2;
4451 mc->mc_flags |= C_INITIALIZED;
4457 if (flags & MDB_APPEND) {
4459 rc = mdb_cursor_last(mc, &k2, &d2);
4461 rc = mc->mc_dbx->md_cmp(key, &k2);
4464 mc->mc_ki[mc->mc_top]++;
4470 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4472 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4473 DPRINTF("duplicate key [%s]", DKEY(key));
4475 return MDB_KEYEXIST;
4477 if (rc && rc != MDB_NOTFOUND)
4481 /* Cursor is positioned, now make sure all pages are writable */
4482 rc2 = mdb_cursor_touch(mc);
4487 /* The key already exists */
4488 if (rc == MDB_SUCCESS) {
4489 /* there's only a key anyway, so this is a no-op */
4490 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4491 unsigned int ksize = mc->mc_db->md_pad;
4492 if (key->mv_size != ksize)
4494 if (flags == MDB_CURRENT) {
4495 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4496 memcpy(ptr, key->mv_data, ksize);
4501 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4504 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4505 /* Was a single item before, must convert now */
4507 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4508 /* Just overwrite the current item */
4509 if (flags == MDB_CURRENT)
4512 dkey.mv_size = NODEDSZ(leaf);
4513 dkey.mv_data = NODEDATA(leaf);
4514 #if UINT_MAX < SIZE_MAX
4515 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4516 #ifdef MISALIGNED_OK
4517 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4519 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4522 /* if data matches, ignore it */
4523 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4524 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4526 /* create a fake page for the dup items */
4527 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4528 dkey.mv_data = dbuf;
4529 fp = (MDB_page *)&pbuf;
4530 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4531 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4532 fp->mp_lower = PAGEHDRSZ;
4533 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4534 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4535 fp->mp_flags |= P_LEAF2;
4536 fp->mp_pad = data->mv_size;
4537 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4539 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4540 (dkey.mv_size & 1) + (data->mv_size & 1);
4542 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4545 xdata.mv_size = fp->mp_upper;
4550 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4551 /* See if we need to convert from fake page to subDB */
4553 unsigned int offset;
4556 fp = NODEDATA(leaf);
4557 if (flags == MDB_CURRENT) {
4559 fp->mp_flags |= P_DIRTY;
4560 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4561 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4565 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4566 offset = fp->mp_pad;
4567 if (SIZELEFT(fp) >= offset)
4569 offset *= 4; /* space for 4 more */
4571 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4573 offset += offset & 1;
4574 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4575 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4577 /* yes, convert it */
4579 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4580 dummy.md_pad = fp->mp_pad;
4581 dummy.md_flags = MDB_DUPFIXED;
4582 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4583 dummy.md_flags |= MDB_INTEGERKEY;
4586 dummy.md_branch_pages = 0;
4587 dummy.md_leaf_pages = 1;
4588 dummy.md_overflow_pages = 0;
4589 dummy.md_entries = NUMKEYS(fp);
4591 xdata.mv_size = sizeof(MDB_db);
4592 xdata.mv_data = &dummy;
4593 if ((rc = mdb_page_alloc(mc, 1, &mp)))
4595 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4596 flags |= F_DUPDATA|F_SUBDATA;
4597 dummy.md_root = mp->mp_pgno;
4599 /* no, just grow it */
4601 xdata.mv_size = NODEDSZ(leaf) + offset;
4602 xdata.mv_data = &pbuf;
4603 mp = (MDB_page *)&pbuf;
4604 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4607 mp->mp_flags = fp->mp_flags | P_DIRTY;
4608 mp->mp_pad = fp->mp_pad;
4609 mp->mp_lower = fp->mp_lower;
4610 mp->mp_upper = fp->mp_upper + offset;
4612 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4614 nsize = NODEDSZ(leaf) - fp->mp_upper;
4615 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4616 for (i=0; i<NUMKEYS(fp); i++)
4617 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4619 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4623 /* data is on sub-DB, just store it */
4624 flags |= F_DUPDATA|F_SUBDATA;
4628 /* overflow page overwrites need special handling */
4629 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4632 int ovpages, dpages;
4634 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4635 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4636 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4637 mdb_page_get(mc->mc_txn, pg, &omp);
4638 /* Is the ov page writable and large enough? */
4639 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4640 /* yes, overwrite it. Note in this case we don't
4641 * bother to try shrinking the node if the new data
4642 * is smaller than the overflow threshold.
4644 if (F_ISSET(flags, MDB_RESERVE))
4645 data->mv_data = METADATA(omp);
4647 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4650 /* no, free ovpages */
4652 mc->mc_db->md_overflow_pages -= ovpages;
4653 for (i=0; i<ovpages; i++) {
4654 DPRINTF("freed ov page %zu", pg);
4655 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4659 } else if (NODEDSZ(leaf) == data->mv_size) {
4660 /* same size, just replace it. Note that we could
4661 * also reuse this node if the new data is smaller,
4662 * but instead we opt to shrink the node in that case.
4664 if (F_ISSET(flags, MDB_RESERVE))
4665 data->mv_data = NODEDATA(leaf);
4667 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4670 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4671 mc->mc_db->md_entries--;
4673 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4680 nflags = flags & NODE_ADD_FLAGS;
4681 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4682 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4683 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4684 nflags &= ~MDB_APPEND;
4686 nflags |= MDB_SPLIT_REPLACE;
4687 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4689 /* There is room already in this leaf page. */
4690 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4691 if (rc == 0 && !do_sub && insert) {
4692 /* Adjust other cursors pointing to mp */
4693 MDB_cursor *m2, *m3;
4694 MDB_dbi dbi = mc->mc_dbi;
4695 unsigned i = mc->mc_top;
4696 MDB_page *mp = mc->mc_pg[i];
4698 if (mc->mc_flags & C_SUB)
4701 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4702 if (mc->mc_flags & C_SUB)
4703 m3 = &m2->mc_xcursor->mx_cursor;
4706 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4707 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4714 if (rc != MDB_SUCCESS)
4715 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4717 /* Now store the actual data in the child DB. Note that we're
4718 * storing the user data in the keys field, so there are strict
4719 * size limits on dupdata. The actual data fields of the child
4720 * DB are all zero size.
4727 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4728 if (flags & MDB_CURRENT) {
4729 xflags = MDB_CURRENT;
4731 mdb_xcursor_init1(mc, leaf);
4732 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4734 /* converted, write the original data first */
4736 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4740 /* Adjust other cursors pointing to mp */
4742 unsigned i = mc->mc_top;
4743 MDB_page *mp = mc->mc_pg[i];
4745 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4746 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4747 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4748 mdb_xcursor_init1(m2, leaf);
4753 if (flags & MDB_APPENDDUP)
4754 xflags |= MDB_APPEND;
4755 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4756 if (flags & F_SUBDATA) {
4757 void *db = NODEDATA(leaf);
4758 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4761 /* sub-writes might have failed so check rc again.
4762 * Don't increment count if we just replaced an existing item.
4764 if (!rc && !(flags & MDB_CURRENT))
4765 mc->mc_db->md_entries++;
4766 if (flags & MDB_MULTIPLE) {
4768 if (mcount < data[1].mv_size) {
4769 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4770 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4780 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4785 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4788 if (!mc->mc_flags & C_INITIALIZED)
4791 rc = mdb_cursor_touch(mc);
4795 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4797 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4798 if (flags != MDB_NODUPDATA) {
4799 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4800 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4802 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4803 /* If sub-DB still has entries, we're done */
4804 if (mc->mc_xcursor->mx_db.md_entries) {
4805 if (leaf->mn_flags & F_SUBDATA) {
4806 /* update subDB info */
4807 void *db = NODEDATA(leaf);
4808 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4810 /* shrink fake page */
4811 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4813 mc->mc_db->md_entries--;
4816 /* otherwise fall thru and delete the sub-DB */
4819 if (leaf->mn_flags & F_SUBDATA) {
4820 /* add all the child DB's pages to the free list */
4821 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4822 if (rc == MDB_SUCCESS) {
4823 mc->mc_db->md_entries -=
4824 mc->mc_xcursor->mx_db.md_entries;
4829 return mdb_cursor_del0(mc, leaf);
4832 /** Allocate and initialize new pages for a database.
4833 * @param[in] mc a cursor on the database being added to.
4834 * @param[in] flags flags defining what type of page is being allocated.
4835 * @param[in] num the number of pages to allocate. This is usually 1,
4836 * unless allocating overflow pages for a large record.
4837 * @param[out] mp Address of a page, or NULL on failure.
4838 * @return 0 on success, non-zero on failure.
4841 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
4846 if ((rc = mdb_page_alloc(mc, num, &np)))
4848 DPRINTF("allocated new mpage %zu, page size %u",
4849 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4850 np->mp_flags = flags | P_DIRTY;
4851 np->mp_lower = PAGEHDRSZ;
4852 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4855 mc->mc_db->md_branch_pages++;
4856 else if (IS_LEAF(np))
4857 mc->mc_db->md_leaf_pages++;
4858 else if (IS_OVERFLOW(np)) {
4859 mc->mc_db->md_overflow_pages += num;
4867 /** Calculate the size of a leaf node.
4868 * The size depends on the environment's page size; if a data item
4869 * is too large it will be put onto an overflow page and the node
4870 * size will only include the key and not the data. Sizes are always
4871 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4872 * of the #MDB_node headers.
4873 * @param[in] env The environment handle.
4874 * @param[in] key The key for the node.
4875 * @param[in] data The data for the node.
4876 * @return The number of bytes needed to store the node.
4879 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4883 sz = LEAFSIZE(key, data);
4884 if (sz >= env->me_psize / MDB_MINKEYS) {
4885 /* put on overflow page */
4886 sz -= data->mv_size - sizeof(pgno_t);
4890 return sz + sizeof(indx_t);
4893 /** Calculate the size of a branch node.
4894 * The size should depend on the environment's page size but since
4895 * we currently don't support spilling large keys onto overflow
4896 * pages, it's simply the size of the #MDB_node header plus the
4897 * size of the key. Sizes are always rounded up to an even number
4898 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4899 * @param[in] env The environment handle.
4900 * @param[in] key The key for the node.
4901 * @return The number of bytes needed to store the node.
4904 mdb_branch_size(MDB_env *env, MDB_val *key)
4909 if (sz >= env->me_psize / MDB_MINKEYS) {
4910 /* put on overflow page */
4911 /* not implemented */
4912 /* sz -= key->size - sizeof(pgno_t); */
4915 return sz + sizeof(indx_t);
4918 /** Add a node to the page pointed to by the cursor.
4919 * @param[in] mc The cursor for this operation.
4920 * @param[in] indx The index on the page where the new node should be added.
4921 * @param[in] key The key for the new node.
4922 * @param[in] data The data for the new node, if any.
4923 * @param[in] pgno The page number, if adding a branch node.
4924 * @param[in] flags Flags for the node.
4925 * @return 0 on success, non-zero on failure. Possible errors are:
4927 * <li>ENOMEM - failed to allocate overflow pages for the node.
4928 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
4929 * should never happen since all callers already calculate the
4930 * page's free space before calling this function.
4934 mdb_node_add(MDB_cursor *mc, indx_t indx,
4935 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4938 size_t node_size = NODESIZE;
4941 MDB_page *mp = mc->mc_pg[mc->mc_top];
4942 MDB_page *ofp = NULL; /* overflow page */
4945 assert(mp->mp_upper >= mp->mp_lower);
4947 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4948 IS_LEAF(mp) ? "leaf" : "branch",
4949 IS_SUBP(mp) ? "sub-" : "",
4950 mp->mp_pgno, indx, data ? data->mv_size : 0,
4951 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4954 /* Move higher keys up one slot. */
4955 int ksize = mc->mc_db->md_pad, dif;
4956 char *ptr = LEAF2KEY(mp, indx, ksize);
4957 dif = NUMKEYS(mp) - indx;
4959 memmove(ptr+ksize, ptr, dif*ksize);
4960 /* insert new key */
4961 memcpy(ptr, key->mv_data, ksize);
4963 /* Just using these for counting */
4964 mp->mp_lower += sizeof(indx_t);
4965 mp->mp_upper -= ksize - sizeof(indx_t);
4970 node_size += key->mv_size;
4974 if (F_ISSET(flags, F_BIGDATA)) {
4975 /* Data already on overflow page. */
4976 node_size += sizeof(pgno_t);
4977 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4978 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4980 /* Put data on overflow page. */
4981 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4982 data->mv_size, node_size+data->mv_size);
4983 node_size += sizeof(pgno_t);
4984 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
4986 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4989 node_size += data->mv_size;
4992 node_size += node_size & 1;
4994 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4995 DPRINTF("not enough room in page %zu, got %u ptrs",
4996 mp->mp_pgno, NUMKEYS(mp));
4997 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4998 mp->mp_upper - mp->mp_lower);
4999 DPRINTF("node size = %zu", node_size);
5000 return MDB_PAGE_FULL;
5003 /* Move higher pointers up one slot. */
5004 for (i = NUMKEYS(mp); i > indx; i--)
5005 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5007 /* Adjust free space offsets. */
5008 ofs = mp->mp_upper - node_size;
5009 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5010 mp->mp_ptrs[indx] = ofs;
5012 mp->mp_lower += sizeof(indx_t);
5014 /* Write the node data. */
5015 node = NODEPTR(mp, indx);
5016 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5017 node->mn_flags = flags;
5019 SETDSZ(node,data->mv_size);
5024 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5029 if (F_ISSET(flags, F_BIGDATA))
5030 memcpy(node->mn_data + key->mv_size, data->mv_data,
5032 else if (F_ISSET(flags, MDB_RESERVE))
5033 data->mv_data = node->mn_data + key->mv_size;
5035 memcpy(node->mn_data + key->mv_size, data->mv_data,
5038 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5040 if (F_ISSET(flags, MDB_RESERVE))
5041 data->mv_data = METADATA(ofp);
5043 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5050 /** Delete the specified node from a page.
5051 * @param[in] mp The page to operate on.
5052 * @param[in] indx The index of the node to delete.
5053 * @param[in] ksize The size of a node. Only used if the page is
5054 * part of a #MDB_DUPFIXED database.
5057 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5060 indx_t i, j, numkeys, ptr;
5067 COPY_PGNO(pgno, mp->mp_pgno);
5068 DPRINTF("delete node %u on %s page %zu", indx,
5069 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5072 assert(indx < NUMKEYS(mp));
5075 int x = NUMKEYS(mp) - 1 - indx;
5076 base = LEAF2KEY(mp, indx, ksize);
5078 memmove(base, base + ksize, x * ksize);
5079 mp->mp_lower -= sizeof(indx_t);
5080 mp->mp_upper += ksize - sizeof(indx_t);
5084 node = NODEPTR(mp, indx);
5085 sz = NODESIZE + node->mn_ksize;
5087 if (F_ISSET(node->mn_flags, F_BIGDATA))
5088 sz += sizeof(pgno_t);
5090 sz += NODEDSZ(node);
5094 ptr = mp->mp_ptrs[indx];
5095 numkeys = NUMKEYS(mp);
5096 for (i = j = 0; i < numkeys; i++) {
5098 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5099 if (mp->mp_ptrs[i] < ptr)
5100 mp->mp_ptrs[j] += sz;
5105 base = (char *)mp + mp->mp_upper;
5106 memmove(base + sz, base, ptr - mp->mp_upper);
5108 mp->mp_lower -= sizeof(indx_t);
5112 /** Compact the main page after deleting a node on a subpage.
5113 * @param[in] mp The main page to operate on.
5114 * @param[in] indx The index of the subpage on the main page.
5117 mdb_node_shrink(MDB_page *mp, indx_t indx)
5124 indx_t i, numkeys, ptr;
5126 node = NODEPTR(mp, indx);
5127 sp = (MDB_page *)NODEDATA(node);
5128 osize = NODEDSZ(node);
5130 delta = sp->mp_upper - sp->mp_lower;
5131 SETDSZ(node, osize - delta);
5132 xp = (MDB_page *)((char *)sp + delta);
5134 /* shift subpage upward */
5136 nsize = NUMKEYS(sp) * sp->mp_pad;
5137 memmove(METADATA(xp), METADATA(sp), nsize);
5140 nsize = osize - sp->mp_upper;
5141 numkeys = NUMKEYS(sp);
5142 for (i=numkeys-1; i>=0; i--)
5143 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5145 xp->mp_upper = sp->mp_lower;
5146 xp->mp_lower = sp->mp_lower;
5147 xp->mp_flags = sp->mp_flags;
5148 xp->mp_pad = sp->mp_pad;
5149 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5151 /* shift lower nodes upward */
5152 ptr = mp->mp_ptrs[indx];
5153 numkeys = NUMKEYS(mp);
5154 for (i = 0; i < numkeys; i++) {
5155 if (mp->mp_ptrs[i] <= ptr)
5156 mp->mp_ptrs[i] += delta;
5159 base = (char *)mp + mp->mp_upper;
5160 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5161 mp->mp_upper += delta;
5164 /** Initial setup of a sorted-dups cursor.
5165 * Sorted duplicates are implemented as a sub-database for the given key.
5166 * The duplicate data items are actually keys of the sub-database.
5167 * Operations on the duplicate data items are performed using a sub-cursor
5168 * initialized when the sub-database is first accessed. This function does
5169 * the preliminary setup of the sub-cursor, filling in the fields that
5170 * depend only on the parent DB.
5171 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5174 mdb_xcursor_init0(MDB_cursor *mc)
5176 MDB_xcursor *mx = mc->mc_xcursor;
5178 mx->mx_cursor.mc_xcursor = NULL;
5179 mx->mx_cursor.mc_txn = mc->mc_txn;
5180 mx->mx_cursor.mc_db = &mx->mx_db;
5181 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5182 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5183 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5184 mx->mx_cursor.mc_snum = 0;
5185 mx->mx_cursor.mc_top = 0;
5186 mx->mx_cursor.mc_flags = C_SUB;
5187 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5188 mx->mx_dbx.md_dcmp = NULL;
5189 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5192 /** Final setup of a sorted-dups cursor.
5193 * Sets up the fields that depend on the data from the main cursor.
5194 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5195 * @param[in] node The data containing the #MDB_db record for the
5196 * sorted-dup database.
5199 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5201 MDB_xcursor *mx = mc->mc_xcursor;
5203 if (node->mn_flags & F_SUBDATA) {
5204 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5205 mx->mx_cursor.mc_pg[0] = 0;
5206 mx->mx_cursor.mc_snum = 0;
5207 mx->mx_cursor.mc_flags = C_SUB;
5209 MDB_page *fp = NODEDATA(node);
5210 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5211 mx->mx_db.md_flags = 0;
5212 mx->mx_db.md_depth = 1;
5213 mx->mx_db.md_branch_pages = 0;
5214 mx->mx_db.md_leaf_pages = 1;
5215 mx->mx_db.md_overflow_pages = 0;
5216 mx->mx_db.md_entries = NUMKEYS(fp);
5217 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5218 mx->mx_cursor.mc_snum = 1;
5219 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5220 mx->mx_cursor.mc_top = 0;
5221 mx->mx_cursor.mc_pg[0] = fp;
5222 mx->mx_cursor.mc_ki[0] = 0;
5223 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5224 mx->mx_db.md_flags = MDB_DUPFIXED;
5225 mx->mx_db.md_pad = fp->mp_pad;
5226 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5227 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5230 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5232 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5234 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5235 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5236 #if UINT_MAX < SIZE_MAX
5237 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5238 #ifdef MISALIGNED_OK
5239 mx->mx_dbx.md_cmp = mdb_cmp_long;
5241 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5246 /** Initialize a cursor for a given transaction and database. */
5248 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5253 mc->mc_db = &txn->mt_dbs[dbi];
5254 mc->mc_dbx = &txn->mt_dbxs[dbi];
5255 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5260 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5262 mc->mc_xcursor = mx;
5263 mdb_xcursor_init0(mc);
5265 mc->mc_xcursor = NULL;
5267 if (*mc->mc_dbflag & DB_STALE) {
5268 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5273 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5276 MDB_xcursor *mx = NULL;
5277 size_t size = sizeof(MDB_cursor);
5279 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5282 /* Allow read access to the freelist */
5283 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5286 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5287 size += sizeof(MDB_xcursor);
5289 if ((mc = malloc(size)) != NULL) {
5290 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5291 mx = (MDB_xcursor *)(mc + 1);
5293 mdb_cursor_init(mc, txn, dbi, mx);
5294 if (txn->mt_cursors) {
5295 mc->mc_next = txn->mt_cursors[dbi];
5296 txn->mt_cursors[dbi] = mc;
5298 mc->mc_flags |= C_ALLOCD;
5309 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5311 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5314 if (txn->mt_cursors)
5317 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5321 /* Return the count of duplicate data items for the current key */
5323 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5327 if (mc == NULL || countp == NULL)
5330 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5333 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5334 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5337 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5340 *countp = mc->mc_xcursor->mx_db.md_entries;
5346 mdb_cursor_close(MDB_cursor *mc)
5349 /* remove from txn, if tracked */
5350 if (mc->mc_txn->mt_cursors) {
5351 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5352 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5354 *prev = mc->mc_next;
5356 if (mc->mc_flags & C_ALLOCD)
5362 mdb_cursor_txn(MDB_cursor *mc)
5364 if (!mc) return NULL;
5369 mdb_cursor_dbi(MDB_cursor *mc)
5375 /** Replace the key for a node with a new key.
5376 * @param[in] mp The page containing the node to operate on.
5377 * @param[in] indx The index of the node to operate on.
5378 * @param[in] key The new key to use.
5379 * @return 0 on success, non-zero on failure.
5382 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5388 indx_t ptr, i, numkeys;
5391 node = NODEPTR(mp, indx);
5392 ptr = mp->mp_ptrs[indx];
5396 char kbuf2[(MAXKEYSIZE*2+1)];
5397 k2.mv_data = NODEKEY(node);
5398 k2.mv_size = node->mn_ksize;
5399 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5401 mdb_dkey(&k2, kbuf2),
5407 delta0 = delta = key->mv_size - node->mn_ksize;
5409 /* Must be 2-byte aligned. If new key is
5410 * shorter by 1, the shift will be skipped.
5412 delta += (delta & 1);
5414 if (delta > 0 && SIZELEFT(mp) < delta) {
5415 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5416 return MDB_PAGE_FULL;
5419 numkeys = NUMKEYS(mp);
5420 for (i = 0; i < numkeys; i++) {
5421 if (mp->mp_ptrs[i] <= ptr)
5422 mp->mp_ptrs[i] -= delta;
5425 base = (char *)mp + mp->mp_upper;
5426 len = ptr - mp->mp_upper + NODESIZE;
5427 memmove(base - delta, base, len);
5428 mp->mp_upper -= delta;
5430 node = NODEPTR(mp, indx);
5433 /* But even if no shift was needed, update ksize */
5435 node->mn_ksize = key->mv_size;
5438 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5443 /** Move a node from csrc to cdst.
5446 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5452 unsigned short flags;
5456 /* Mark src and dst as dirty. */
5457 if ((rc = mdb_page_touch(csrc)) ||
5458 (rc = mdb_page_touch(cdst)))
5461 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5462 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5463 key.mv_size = csrc->mc_db->md_pad;
5464 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5466 data.mv_data = NULL;
5470 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5471 assert(!((long)srcnode&1));
5472 srcpg = NODEPGNO(srcnode);
5473 flags = srcnode->mn_flags;
5474 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5475 unsigned int snum = csrc->mc_snum;
5477 /* must find the lowest key below src */
5478 mdb_page_search_root(csrc, NULL, 0);
5479 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5480 key.mv_size = csrc->mc_db->md_pad;
5481 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5483 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5484 key.mv_size = NODEKSZ(s2);
5485 key.mv_data = NODEKEY(s2);
5487 csrc->mc_snum = snum--;
5488 csrc->mc_top = snum;
5490 key.mv_size = NODEKSZ(srcnode);
5491 key.mv_data = NODEKEY(srcnode);
5493 data.mv_size = NODEDSZ(srcnode);
5494 data.mv_data = NODEDATA(srcnode);
5496 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5497 unsigned int snum = cdst->mc_snum;
5500 /* must find the lowest key below dst */
5501 mdb_page_search_root(cdst, NULL, 0);
5502 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5503 bkey.mv_size = cdst->mc_db->md_pad;
5504 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5506 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5507 bkey.mv_size = NODEKSZ(s2);
5508 bkey.mv_data = NODEKEY(s2);
5510 cdst->mc_snum = snum--;
5511 cdst->mc_top = snum;
5512 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5515 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5516 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5517 csrc->mc_ki[csrc->mc_top],
5519 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5520 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5522 /* Add the node to the destination page.
5524 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5525 if (rc != MDB_SUCCESS)
5528 /* Delete the node from the source page.
5530 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5533 /* Adjust other cursors pointing to mp */
5534 MDB_cursor *m2, *m3;
5535 MDB_dbi dbi = csrc->mc_dbi;
5536 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5538 if (csrc->mc_flags & C_SUB)
5541 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5542 if (m2 == csrc) continue;
5543 if (csrc->mc_flags & C_SUB)
5544 m3 = &m2->mc_xcursor->mx_cursor;
5547 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5548 csrc->mc_ki[csrc->mc_top]) {
5549 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5550 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5555 /* Update the parent separators.
5557 if (csrc->mc_ki[csrc->mc_top] == 0) {
5558 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5559 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5560 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5562 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5563 key.mv_size = NODEKSZ(srcnode);
5564 key.mv_data = NODEKEY(srcnode);
5566 DPRINTF("update separator for source page %zu to [%s]",
5567 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5568 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5569 &key)) != MDB_SUCCESS)
5572 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5574 nullkey.mv_size = 0;
5575 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5576 assert(rc == MDB_SUCCESS);
5580 if (cdst->mc_ki[cdst->mc_top] == 0) {
5581 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5582 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5583 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5585 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5586 key.mv_size = NODEKSZ(srcnode);
5587 key.mv_data = NODEKEY(srcnode);
5589 DPRINTF("update separator for destination page %zu to [%s]",
5590 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5591 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5592 &key)) != MDB_SUCCESS)
5595 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5597 nullkey.mv_size = 0;
5598 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5599 assert(rc == MDB_SUCCESS);
5606 /** Merge one page into another.
5607 * The nodes from the page pointed to by \b csrc will
5608 * be copied to the page pointed to by \b cdst and then
5609 * the \b csrc page will be freed.
5610 * @param[in] csrc Cursor pointing to the source page.
5611 * @param[in] cdst Cursor pointing to the destination page.
5614 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5622 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5623 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5625 assert(csrc->mc_snum > 1); /* can't merge root page */
5626 assert(cdst->mc_snum > 1);
5628 /* Mark dst as dirty. */
5629 if ((rc = mdb_page_touch(cdst)))
5632 /* Move all nodes from src to dst.
5634 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5635 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5636 key.mv_size = csrc->mc_db->md_pad;
5637 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5638 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5639 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5640 if (rc != MDB_SUCCESS)
5642 key.mv_data = (char *)key.mv_data + key.mv_size;
5645 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5646 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5647 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5648 unsigned int snum = csrc->mc_snum;
5650 /* must find the lowest key below src */
5651 mdb_page_search_root(csrc, NULL, 0);
5652 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5653 key.mv_size = csrc->mc_db->md_pad;
5654 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5656 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5657 key.mv_size = NODEKSZ(s2);
5658 key.mv_data = NODEKEY(s2);
5660 csrc->mc_snum = snum--;
5661 csrc->mc_top = snum;
5663 key.mv_size = srcnode->mn_ksize;
5664 key.mv_data = NODEKEY(srcnode);
5667 data.mv_size = NODEDSZ(srcnode);
5668 data.mv_data = NODEDATA(srcnode);
5669 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5670 if (rc != MDB_SUCCESS)
5675 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5676 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);
5678 /* Unlink the src page from parent and add to free list.
5680 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5681 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5683 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5687 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5688 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5689 csrc->mc_db->md_leaf_pages--;
5691 csrc->mc_db->md_branch_pages--;
5693 /* Adjust other cursors pointing to mp */
5694 MDB_cursor *m2, *m3;
5695 MDB_dbi dbi = csrc->mc_dbi;
5696 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5698 if (csrc->mc_flags & C_SUB)
5701 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5702 if (csrc->mc_flags & C_SUB)
5703 m3 = &m2->mc_xcursor->mx_cursor;
5706 if (m3 == csrc) continue;
5707 if (m3->mc_snum < csrc->mc_snum) continue;
5708 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5709 m3->mc_pg[csrc->mc_top] = mp;
5710 m3->mc_ki[csrc->mc_top] += nkeys;
5714 mdb_cursor_pop(csrc);
5716 return mdb_rebalance(csrc);
5719 /** Copy the contents of a cursor.
5720 * @param[in] csrc The cursor to copy from.
5721 * @param[out] cdst The cursor to copy to.
5724 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5728 cdst->mc_txn = csrc->mc_txn;
5729 cdst->mc_dbi = csrc->mc_dbi;
5730 cdst->mc_db = csrc->mc_db;
5731 cdst->mc_dbx = csrc->mc_dbx;
5732 cdst->mc_snum = csrc->mc_snum;
5733 cdst->mc_top = csrc->mc_top;
5734 cdst->mc_flags = csrc->mc_flags;
5736 for (i=0; i<csrc->mc_snum; i++) {
5737 cdst->mc_pg[i] = csrc->mc_pg[i];
5738 cdst->mc_ki[i] = csrc->mc_ki[i];
5742 /** Rebalance the tree after a delete operation.
5743 * @param[in] mc Cursor pointing to the page where rebalancing
5745 * @return 0 on success, non-zero on failure.
5748 mdb_rebalance(MDB_cursor *mc)
5758 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5759 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5760 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5761 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5765 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5768 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5769 DPRINTF("no need to rebalance page %zu, above fill threshold",
5775 if (mc->mc_snum < 2) {
5776 MDB_page *mp = mc->mc_pg[0];
5777 if (NUMKEYS(mp) == 0) {
5778 DPUTS("tree is completely empty");
5779 mc->mc_db->md_root = P_INVALID;
5780 mc->mc_db->md_depth = 0;
5781 mc->mc_db->md_leaf_pages = 0;
5782 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5786 /* Adjust other cursors pointing to mp */
5787 MDB_cursor *m2, *m3;
5788 MDB_dbi dbi = mc->mc_dbi;
5790 if (mc->mc_flags & C_SUB)
5793 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5794 if (m2 == mc) continue;
5795 if (mc->mc_flags & C_SUB)
5796 m3 = &m2->mc_xcursor->mx_cursor;
5799 if (m3->mc_snum < mc->mc_snum) continue;
5800 if (m3->mc_pg[0] == mp) {
5806 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5807 DPUTS("collapsing root page!");
5808 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5809 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5810 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5813 mc->mc_db->md_depth--;
5814 mc->mc_db->md_branch_pages--;
5816 /* Adjust other cursors pointing to mp */
5817 MDB_cursor *m2, *m3;
5818 MDB_dbi dbi = mc->mc_dbi;
5820 if (mc->mc_flags & C_SUB)
5823 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5824 if (m2 == mc) continue;
5825 if (mc->mc_flags & C_SUB)
5826 m3 = &m2->mc_xcursor->mx_cursor;
5829 if (m3->mc_snum < mc->mc_snum) continue;
5830 if (m3->mc_pg[0] == mp) {
5831 m3->mc_pg[0] = mc->mc_pg[0];
5836 DPUTS("root page doesn't need rebalancing");
5840 /* The parent (branch page) must have at least 2 pointers,
5841 * otherwise the tree is invalid.
5843 ptop = mc->mc_top-1;
5844 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5846 /* Leaf page fill factor is below the threshold.
5847 * Try to move keys from left or right neighbor, or
5848 * merge with a neighbor page.
5853 mdb_cursor_copy(mc, &mn);
5854 mn.mc_xcursor = NULL;
5856 if (mc->mc_ki[ptop] == 0) {
5857 /* We're the leftmost leaf in our parent.
5859 DPUTS("reading right neighbor");
5861 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5862 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5864 mn.mc_ki[mn.mc_top] = 0;
5865 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5867 /* There is at least one neighbor to the left.
5869 DPUTS("reading left neighbor");
5871 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5872 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5874 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5875 mc->mc_ki[mc->mc_top] = 0;
5878 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5879 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);
5881 /* If the neighbor page is above threshold and has at least two
5882 * keys, move one key from it.
5884 * Otherwise we should try to merge them.
5886 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5887 return mdb_node_move(&mn, mc);
5888 else { /* FIXME: if (has_enough_room()) */
5889 mc->mc_flags &= ~C_INITIALIZED;
5890 if (mc->mc_ki[ptop] == 0)
5891 return mdb_page_merge(&mn, mc);
5893 return mdb_page_merge(mc, &mn);
5897 /** Complete a delete operation started by #mdb_cursor_del(). */
5899 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5903 /* add overflow pages to free list */
5904 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5908 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5909 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5910 mc->mc_db->md_overflow_pages -= ovpages;
5911 for (i=0; i<ovpages; i++) {
5912 DPRINTF("freed ov page %zu", pg);
5913 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5917 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5918 mc->mc_db->md_entries--;
5919 rc = mdb_rebalance(mc);
5920 if (rc != MDB_SUCCESS)
5921 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5927 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5928 MDB_val *key, MDB_val *data)
5933 MDB_val rdata, *xdata;
5937 assert(key != NULL);
5939 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5941 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5944 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5948 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5952 mdb_cursor_init(&mc, txn, dbi, &mx);
5963 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5965 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5969 /** Split a page and insert a new node.
5970 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5971 * The cursor will be updated to point to the actual page and index where
5972 * the node got inserted after the split.
5973 * @param[in] newkey The key for the newly inserted node.
5974 * @param[in] newdata The data for the newly inserted node.
5975 * @param[in] newpgno The page number, if the new node is a branch node.
5976 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
5977 * @return 0 on success, non-zero on failure.
5980 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5981 unsigned int nflags)
5984 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
5987 unsigned int i, j, split_indx, nkeys, pmax;
5989 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5991 MDB_page *mp, *rp, *pp;
5996 mp = mc->mc_pg[mc->mc_top];
5997 newindx = mc->mc_ki[mc->mc_top];
5999 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6000 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6001 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6003 /* Create a right sibling. */
6004 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6006 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6008 if (mc->mc_snum < 2) {
6009 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6011 /* shift current top to make room for new parent */
6012 mc->mc_pg[1] = mc->mc_pg[0];
6013 mc->mc_ki[1] = mc->mc_ki[0];
6016 mc->mc_db->md_root = pp->mp_pgno;
6017 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6018 mc->mc_db->md_depth++;
6021 /* Add left (implicit) pointer. */
6022 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6023 /* undo the pre-push */
6024 mc->mc_pg[0] = mc->mc_pg[1];
6025 mc->mc_ki[0] = mc->mc_ki[1];
6026 mc->mc_db->md_root = mp->mp_pgno;
6027 mc->mc_db->md_depth--;
6034 ptop = mc->mc_top-1;
6035 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6038 mc->mc_flags |= C_SPLITTING;
6039 mdb_cursor_copy(mc, &mn);
6040 mn.mc_pg[mn.mc_top] = rp;
6041 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6043 if (nflags & MDB_APPEND) {
6044 mn.mc_ki[mn.mc_top] = 0;
6046 split_indx = newindx;
6051 nkeys = NUMKEYS(mp);
6052 split_indx = nkeys / 2;
6053 if (newindx < split_indx)
6059 unsigned int lsize, rsize, ksize;
6060 /* Move half of the keys to the right sibling */
6062 x = mc->mc_ki[mc->mc_top] - split_indx;
6063 ksize = mc->mc_db->md_pad;
6064 split = LEAF2KEY(mp, split_indx, ksize);
6065 rsize = (nkeys - split_indx) * ksize;
6066 lsize = (nkeys - split_indx) * sizeof(indx_t);
6067 mp->mp_lower -= lsize;
6068 rp->mp_lower += lsize;
6069 mp->mp_upper += rsize - lsize;
6070 rp->mp_upper -= rsize - lsize;
6071 sepkey.mv_size = ksize;
6072 if (newindx == split_indx) {
6073 sepkey.mv_data = newkey->mv_data;
6075 sepkey.mv_data = split;
6078 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6079 memcpy(rp->mp_ptrs, split, rsize);
6080 sepkey.mv_data = rp->mp_ptrs;
6081 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6082 memcpy(ins, newkey->mv_data, ksize);
6083 mp->mp_lower += sizeof(indx_t);
6084 mp->mp_upper -= ksize - sizeof(indx_t);
6087 memcpy(rp->mp_ptrs, split, x * ksize);
6088 ins = LEAF2KEY(rp, x, ksize);
6089 memcpy(ins, newkey->mv_data, ksize);
6090 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6091 rp->mp_lower += sizeof(indx_t);
6092 rp->mp_upper -= ksize - sizeof(indx_t);
6093 mc->mc_ki[mc->mc_top] = x;
6094 mc->mc_pg[mc->mc_top] = rp;
6099 /* For leaf pages, check the split point based on what
6100 * fits where, since otherwise mdb_node_add can fail.
6102 * This check is only needed when the data items are
6103 * relatively large, such that being off by one will
6104 * make the difference between success or failure.
6105 * When the size of the data items is much smaller than
6106 * one-half of a page, this check is irrelevant.
6109 unsigned int psize, nsize;
6110 /* Maximum free space in an empty page */
6111 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6112 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6113 if ((nkeys < 20) || (nsize > pmax/4)) {
6114 if (newindx <= split_indx) {
6117 for (i=0; i<split_indx; i++) {
6118 node = NODEPTR(mp, i);
6119 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6120 if (F_ISSET(node->mn_flags, F_BIGDATA))
6121 psize += sizeof(pgno_t);
6123 psize += NODEDSZ(node);
6127 split_indx = newindx;
6138 for (i=nkeys-1; i>=split_indx; i--) {
6139 node = NODEPTR(mp, i);
6140 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6141 if (F_ISSET(node->mn_flags, F_BIGDATA))
6142 psize += sizeof(pgno_t);
6144 psize += NODEDSZ(node);
6148 split_indx = newindx;
6158 /* First find the separating key between the split pages.
6159 * The case where newindx == split_indx is ambiguous; the
6160 * new item could go to the new page or stay on the original
6161 * page. If newpos == 1 it goes to the new page.
6163 if (newindx == split_indx && newpos) {
6164 sepkey.mv_size = newkey->mv_size;
6165 sepkey.mv_data = newkey->mv_data;
6167 node = NODEPTR(mp, split_indx);
6168 sepkey.mv_size = node->mn_ksize;
6169 sepkey.mv_data = NODEKEY(node);
6173 DPRINTF("separator is [%s]", DKEY(&sepkey));
6175 /* Copy separator key to the parent.
6177 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6181 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6184 if (mn.mc_snum == mc->mc_snum) {
6185 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6186 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6187 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6188 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6193 /* Right page might now have changed parent.
6194 * Check if left page also changed parent.
6196 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6197 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6198 for (i=0; i<ptop; i++) {
6199 mc->mc_pg[i] = mn.mc_pg[i];
6200 mc->mc_ki[i] = mn.mc_ki[i];
6202 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6203 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6207 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6210 mc->mc_flags ^= C_SPLITTING;
6211 if (rc != MDB_SUCCESS) {
6214 if (nflags & MDB_APPEND) {
6215 mc->mc_pg[mc->mc_top] = rp;
6216 mc->mc_ki[mc->mc_top] = 0;
6217 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6220 for (i=0; i<mc->mc_top; i++)
6221 mc->mc_ki[i] = mn.mc_ki[i];
6228 /* Move half of the keys to the right sibling. */
6230 /* grab a page to hold a temporary copy */
6231 copy = mdb_page_malloc(mc);
6235 copy->mp_pgno = mp->mp_pgno;
6236 copy->mp_flags = mp->mp_flags;
6237 copy->mp_lower = PAGEHDRSZ;
6238 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6239 mc->mc_pg[mc->mc_top] = copy;
6240 for (i = j = 0; i <= nkeys; j++) {
6241 if (i == split_indx) {
6242 /* Insert in right sibling. */
6243 /* Reset insert index for right sibling. */
6244 if (i != newindx || (newpos ^ ins_new)) {
6246 mc->mc_pg[mc->mc_top] = rp;
6250 if (i == newindx && !ins_new) {
6251 /* Insert the original entry that caused the split. */
6252 rkey.mv_data = newkey->mv_data;
6253 rkey.mv_size = newkey->mv_size;
6262 /* Update index for the new key. */
6263 mc->mc_ki[mc->mc_top] = j;
6264 } else if (i == nkeys) {
6267 node = NODEPTR(mp, i);
6268 rkey.mv_data = NODEKEY(node);
6269 rkey.mv_size = node->mn_ksize;
6271 xdata.mv_data = NODEDATA(node);
6272 xdata.mv_size = NODEDSZ(node);
6275 pgno = NODEPGNO(node);
6276 flags = node->mn_flags;
6281 if (!IS_LEAF(mp) && j == 0) {
6282 /* First branch index doesn't need key data. */
6286 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6290 nkeys = NUMKEYS(copy);
6291 for (i=0; i<nkeys; i++)
6292 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6293 mp->mp_lower = copy->mp_lower;
6294 mp->mp_upper = copy->mp_upper;
6295 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6296 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6298 /* reset back to original page */
6299 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6300 mc->mc_pg[mc->mc_top] = mp;
6301 if (nflags & MDB_RESERVE) {
6302 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6303 if (!(node->mn_flags & F_BIGDATA))
6304 newdata->mv_data = NODEDATA(node);
6310 /* return tmp page to freelist */
6311 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
6312 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
6313 mc->mc_txn->mt_env->me_dpages = copy;
6316 /* Adjust other cursors pointing to mp */
6317 MDB_cursor *m2, *m3;
6318 MDB_dbi dbi = mc->mc_dbi;
6319 int fixup = NUMKEYS(mp);
6321 if (mc->mc_flags & C_SUB)
6324 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6325 if (m2 == mc) continue;
6326 if (mc->mc_flags & C_SUB)
6327 m3 = &m2->mc_xcursor->mx_cursor;
6330 if (!(m3->mc_flags & C_INITIALIZED))
6332 if (m3->mc_flags & C_SPLITTING)
6337 for (k=m3->mc_top; k>=0; k--) {
6338 m3->mc_ki[k+1] = m3->mc_ki[k];
6339 m3->mc_pg[k+1] = m3->mc_pg[k];
6341 if (m3->mc_ki[0] >= split_indx) {
6346 m3->mc_pg[0] = mc->mc_pg[0];
6350 if (m3->mc_pg[mc->mc_top] == mp) {
6351 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6352 m3->mc_ki[mc->mc_top]++;
6353 if (m3->mc_ki[mc->mc_top] >= fixup) {
6354 m3->mc_pg[mc->mc_top] = rp;
6355 m3->mc_ki[mc->mc_top] -= fixup;
6356 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6358 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6359 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6368 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6369 MDB_val *key, MDB_val *data, unsigned int flags)
6374 assert(key != NULL);
6375 assert(data != NULL);
6377 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6380 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6384 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6388 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6391 mdb_cursor_init(&mc, txn, dbi, &mx);
6392 return mdb_cursor_put(&mc, key, data, flags);
6395 /** Only a subset of the @ref mdb_env flags can be changed
6396 * at runtime. Changing other flags requires closing the environment
6397 * and re-opening it with the new flags.
6399 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
6401 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6403 if ((flag & CHANGEABLE) != flag)
6406 env->me_flags |= flag;
6408 env->me_flags &= ~flag;
6413 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6418 *arg = env->me_flags;
6423 mdb_env_get_path(MDB_env *env, const char **arg)
6428 *arg = env->me_path;
6432 /** Common code for #mdb_stat() and #mdb_env_stat().
6433 * @param[in] env the environment to operate in.
6434 * @param[in] db the #MDB_db record containing the stats to return.
6435 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6436 * @return 0, this function always succeeds.
6439 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6441 arg->ms_psize = env->me_psize;
6442 arg->ms_depth = db->md_depth;
6443 arg->ms_branch_pages = db->md_branch_pages;
6444 arg->ms_leaf_pages = db->md_leaf_pages;
6445 arg->ms_overflow_pages = db->md_overflow_pages;
6446 arg->ms_entries = db->md_entries;
6451 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6455 if (env == NULL || arg == NULL)
6458 toggle = mdb_env_pick_meta(env);
6460 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6463 /** Set the default comparison functions for a database.
6464 * Called immediately after a database is opened to set the defaults.
6465 * The user can then override them with #mdb_set_compare() or
6466 * #mdb_set_dupsort().
6467 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6468 * @param[in] dbi A database handle returned by #mdb_open()
6471 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6473 uint16_t f = txn->mt_dbs[dbi].md_flags;
6475 txn->mt_dbxs[dbi].md_cmp =
6476 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6477 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6479 txn->mt_dbxs[dbi].md_dcmp =
6480 !(f & MDB_DUPSORT) ? 0 :
6481 ((f & MDB_INTEGERDUP)
6482 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6483 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6486 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6491 int rc, dbflag, exact;
6492 unsigned int unused = 0;
6495 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6496 mdb_default_cmp(txn, FREE_DBI);
6502 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
6503 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
6504 mdb_default_cmp(txn, MAIN_DBI);
6508 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6509 mdb_default_cmp(txn, MAIN_DBI);
6512 /* Is the DB already open? */
6514 for (i=2; i<txn->mt_numdbs; i++) {
6515 if (!txn->mt_dbxs[i].md_name.mv_size) {
6516 /* Remember this free slot */
6517 if (!unused) unused = i;
6520 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6521 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6527 /* If no free slot and max hit, fail */
6528 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
6529 return MDB_DBS_FULL;
6531 /* Find the DB info */
6535 key.mv_data = (void *)name;
6536 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6537 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6538 if (rc == MDB_SUCCESS) {
6539 /* make sure this is actually a DB */
6540 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6541 if (!(node->mn_flags & F_SUBDATA))
6543 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6544 /* Create if requested */
6546 data.mv_size = sizeof(MDB_db);
6547 data.mv_data = &dummy;
6548 memset(&dummy, 0, sizeof(dummy));
6549 dummy.md_root = P_INVALID;
6550 dummy.md_flags = flags & 0xffff;
6551 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6555 /* OK, got info, add to table */
6556 if (rc == MDB_SUCCESS) {
6557 unsigned int slot = unused ? unused : txn->mt_numdbs;
6558 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
6559 txn->mt_dbxs[slot].md_name.mv_size = len;
6560 txn->mt_dbxs[slot].md_rel = NULL;
6561 txn->mt_dbflags[slot] = dbflag;
6562 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
6564 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
6565 mdb_default_cmp(txn, slot);
6568 txn->mt_env->me_numdbs++;
6575 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6577 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6580 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6583 void mdb_close(MDB_env *env, MDB_dbi dbi)
6586 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6588 ptr = env->me_dbxs[dbi].md_name.mv_data;
6589 env->me_dbxs[dbi].md_name.mv_data = NULL;
6590 env->me_dbxs[dbi].md_name.mv_size = 0;
6594 /** Add all the DB's pages to the free list.
6595 * @param[in] mc Cursor on the DB to free.
6596 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6597 * @return 0 on success, non-zero on failure.
6600 mdb_drop0(MDB_cursor *mc, int subs)
6604 rc = mdb_page_search(mc, NULL, 0);
6605 if (rc == MDB_SUCCESS) {
6610 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6611 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6614 mdb_cursor_copy(mc, &mx);
6615 while (mc->mc_snum > 0) {
6616 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6617 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6618 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6619 if (ni->mn_flags & F_SUBDATA) {
6620 mdb_xcursor_init1(mc, ni);
6621 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6627 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6629 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6632 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6637 rc = mdb_cursor_sibling(mc, 1);
6639 /* no more siblings, go back to beginning
6640 * of previous level. (stack was already popped
6641 * by mdb_cursor_sibling)
6643 for (i=1; i<mc->mc_top; i++)
6644 mc->mc_pg[i] = mx.mc_pg[i];
6648 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6649 mc->mc_db->md_root);
6654 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6659 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6662 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6665 rc = mdb_cursor_open(txn, dbi, &mc);
6669 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6673 /* Can't delete the main DB */
6674 if (del && dbi > MAIN_DBI) {
6675 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6677 mdb_close(txn->mt_env, dbi);
6679 txn->mt_dbflags[dbi] |= DB_DIRTY;
6680 txn->mt_dbs[dbi].md_depth = 0;
6681 txn->mt_dbs[dbi].md_branch_pages = 0;
6682 txn->mt_dbs[dbi].md_leaf_pages = 0;
6683 txn->mt_dbs[dbi].md_overflow_pages = 0;
6684 txn->mt_dbs[dbi].md_entries = 0;
6685 txn->mt_dbs[dbi].md_root = P_INVALID;
6688 mdb_cursor_close(mc);
6692 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6694 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6697 txn->mt_dbxs[dbi].md_cmp = cmp;
6701 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6703 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6706 txn->mt_dbxs[dbi].md_dcmp = cmp;
6710 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6712 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6715 txn->mt_dbxs[dbi].md_rel = rel;
6719 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6721 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6724 txn->mt_dbxs[dbi].md_relctx = ctx;