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.
36 #include <sys/types.h>
38 #include <sys/param.h>
44 #ifdef HAVE_SYS_FILE_H
61 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
62 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
65 #if defined(__APPLE__) || defined (BSD)
66 # define MDB_USE_POSIX_SEM 1
67 # define MDB_FDATASYNC fsync
68 #elif defined(ANDROID)
69 # define MDB_FDATASYNC fsync
74 #ifdef MDB_USE_POSIX_SEM
75 #include <semaphore.h>
80 #include <valgrind/memcheck.h>
81 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
82 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
83 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
84 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
85 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
87 #define VGMEMP_CREATE(h,r,z)
88 #define VGMEMP_ALLOC(h,a,s)
89 #define VGMEMP_FREE(h,a)
90 #define VGMEMP_DESTROY(h)
91 #define VGMEMP_DEFINED(a,s)
95 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
96 /* Solaris just defines one or the other */
97 # define LITTLE_ENDIAN 1234
98 # define BIG_ENDIAN 4321
99 # ifdef _LITTLE_ENDIAN
100 # define BYTE_ORDER LITTLE_ENDIAN
102 # define BYTE_ORDER BIG_ENDIAN
105 # define BYTE_ORDER __BYTE_ORDER
109 #ifndef LITTLE_ENDIAN
110 #define LITTLE_ENDIAN __LITTLE_ENDIAN
113 #define BIG_ENDIAN __BIG_ENDIAN
116 #if defined(__i386) || defined(__x86_64)
117 #define MISALIGNED_OK 1
123 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
124 # error "Unknown or unsupported endianness (BYTE_ORDER)"
125 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
126 # error "Two's complement, reasonably sized integer types, please"
129 /** @defgroup internal MDB Internals
132 /** @defgroup compat Windows Compatibility Macros
133 * A bunch of macros to minimize the amount of platform-specific ifdefs
134 * needed throughout the rest of the code. When the features this library
135 * needs are similar enough to POSIX to be hidden in a one-or-two line
136 * replacement, this macro approach is used.
140 #define pthread_t DWORD
141 #define pthread_mutex_t HANDLE
142 #define pthread_key_t DWORD
143 #define pthread_self() GetCurrentThreadId()
144 #define pthread_key_create(x,y) \
145 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
146 #define pthread_key_delete(x) TlsFree(x)
147 #define pthread_getspecific(x) TlsGetValue(x)
148 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
149 #define pthread_mutex_unlock(x) ReleaseMutex(x)
150 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
151 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
152 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
153 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
154 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
155 #define getpid() GetCurrentProcessId()
156 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
157 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
158 #define ErrCode() GetLastError()
159 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
160 #define close(fd) CloseHandle(fd)
161 #define munmap(ptr,len) UnmapViewOfFile(ptr)
164 #ifdef MDB_USE_POSIX_SEM
166 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
167 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
168 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
169 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
172 mdb_sem_wait(sem_t *sem)
175 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
180 /** Lock the reader mutex.
182 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
183 /** Unlock the reader mutex.
185 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
187 /** Lock the writer mutex.
188 * Only a single write transaction is allowed at a time. Other writers
189 * will block waiting for this mutex.
191 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
192 /** Unlock the writer mutex.
194 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
195 #endif /* MDB_USE_POSIX_SEM */
197 /** Get the error code for the last failed system function.
199 #define ErrCode() errno
201 /** An abstraction for a file handle.
202 * On POSIX systems file handles are small integers. On Windows
203 * they're opaque pointers.
207 /** A value for an invalid file handle.
208 * Mainly used to initialize file variables and signify that they are
211 #define INVALID_HANDLE_VALUE (-1)
213 /** Get the size of a memory page for the system.
214 * This is the basic size that the platform's memory manager uses, and is
215 * fundamental to the use of memory-mapped files.
217 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
220 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
223 #define MNAME_LEN (sizeof(pthread_mutex_t))
229 /** A flag for opening a file and requesting synchronous data writes.
230 * This is only used when writing a meta page. It's not strictly needed;
231 * we could just do a normal write and then immediately perform a flush.
232 * But if this flag is available it saves us an extra system call.
234 * @note If O_DSYNC is undefined but exists in /usr/include,
235 * preferably set some compiler flag to get the definition.
236 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
239 # define MDB_DSYNC O_DSYNC
243 /** Function for flushing the data of a file. Define this to fsync
244 * if fdatasync() is not supported.
246 #ifndef MDB_FDATASYNC
247 # define MDB_FDATASYNC fdatasync
251 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
262 /** A page number in the database.
263 * Note that 64 bit page numbers are overkill, since pages themselves
264 * already represent 12-13 bits of addressable memory, and the OS will
265 * always limit applications to a maximum of 63 bits of address space.
267 * @note In the #MDB_node structure, we only store 48 bits of this value,
268 * which thus limits us to only 60 bits of addressable data.
270 typedef MDB_ID pgno_t;
272 /** A transaction ID.
273 * See struct MDB_txn.mt_txnid for details.
275 typedef MDB_ID txnid_t;
277 /** @defgroup debug Debug Macros
281 /** Enable debug output.
282 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
283 * read from and written to the database (used for free space management).
288 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
289 # define DPRINTF (void) /* Vararg macros may be unsupported */
291 static int mdb_debug;
292 static txnid_t mdb_debug_start;
294 /** Print a debug message with printf formatting. */
295 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
296 ((void) ((mdb_debug) && \
297 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
299 # define DPRINTF(fmt, ...) ((void) 0)
300 # define MDB_DEBUG_SKIP
302 /** Print a debug string.
303 * The string is printed literally, with no format processing.
305 #define DPUTS(arg) DPRINTF("%s", arg)
308 /** A default memory page size.
309 * The actual size is platform-dependent, but we use this for
310 * boot-strapping. We probably should not be using this any more.
311 * The #GET_PAGESIZE() macro is used to get the actual size.
313 * Note that we don't currently support Huge pages. On Linux,
314 * regular data files cannot use Huge pages, and in general
315 * Huge pages aren't actually pageable. We rely on the OS
316 * demand-pager to read our data and page it out when memory
317 * pressure from other processes is high. So until OSs have
318 * actual paging support for Huge pages, they're not viable.
320 #define MDB_PAGESIZE 4096
322 /** The minimum number of keys required in a database page.
323 * Setting this to a larger value will place a smaller bound on the
324 * maximum size of a data item. Data items larger than this size will
325 * be pushed into overflow pages instead of being stored directly in
326 * the B-tree node. This value used to default to 4. With a page size
327 * of 4096 bytes that meant that any item larger than 1024 bytes would
328 * go into an overflow page. That also meant that on average 2-3KB of
329 * each overflow page was wasted space. The value cannot be lower than
330 * 2 because then there would no longer be a tree structure. With this
331 * value, items larger than 2KB will go into overflow pages, and on
332 * average only 1KB will be wasted.
334 #define MDB_MINKEYS 2
336 /** A stamp that identifies a file as an MDB file.
337 * There's nothing special about this value other than that it is easily
338 * recognizable, and it will reflect any byte order mismatches.
340 #define MDB_MAGIC 0xBEEFC0DE
342 /** The version number for a database's file format. */
343 #define MDB_VERSION 1
345 /** The maximum size of a key in the database.
346 * While data items have essentially unbounded size, we require that
347 * keys all fit onto a regular page. This limit could be raised a bit
348 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
350 #define MAXKEYSIZE 511
355 * This is used for printing a hex dump of a key's contents.
357 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
358 /** Display a key in hex.
360 * Invoke a function to display a key in hex.
362 #define DKEY(x) mdb_dkey(x, kbuf)
364 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
368 /** An invalid page number.
369 * Mainly used to denote an empty tree.
371 #define P_INVALID (~(pgno_t)0)
373 /** Test if a flag \b f is set in a flag word \b w. */
374 #define F_ISSET(w, f) (((w) & (f)) == (f))
376 /** Used for offsets within a single page.
377 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
380 typedef uint16_t indx_t;
382 /** Default size of memory map.
383 * This is certainly too small for any actual applications. Apps should always set
384 * the size explicitly using #mdb_env_set_mapsize().
386 #define DEFAULT_MAPSIZE 1048576
388 /** @defgroup readers Reader Lock Table
389 * Readers don't acquire any locks for their data access. Instead, they
390 * simply record their transaction ID in the reader table. The reader
391 * mutex is needed just to find an empty slot in the reader table. The
392 * slot's address is saved in thread-specific data so that subsequent read
393 * transactions started by the same thread need no further locking to proceed.
395 * Since the database uses multi-version concurrency control, readers don't
396 * actually need any locking. This table is used to keep track of which
397 * readers are using data from which old transactions, so that we'll know
398 * when a particular old transaction is no longer in use. Old transactions
399 * that have discarded any data pages can then have those pages reclaimed
400 * for use by a later write transaction.
402 * The lock table is constructed such that reader slots are aligned with the
403 * processor's cache line size. Any slot is only ever used by one thread.
404 * This alignment guarantees that there will be no contention or cache
405 * thrashing as threads update their own slot info, and also eliminates
406 * any need for locking when accessing a slot.
408 * A writer thread will scan every slot in the table to determine the oldest
409 * outstanding reader transaction. Any freed pages older than this will be
410 * reclaimed by the writer. The writer doesn't use any locks when scanning
411 * this table. This means that there's no guarantee that the writer will
412 * see the most up-to-date reader info, but that's not required for correct
413 * operation - all we need is to know the upper bound on the oldest reader,
414 * we don't care at all about the newest reader. So the only consequence of
415 * reading stale information here is that old pages might hang around a
416 * while longer before being reclaimed. That's actually good anyway, because
417 * the longer we delay reclaiming old pages, the more likely it is that a
418 * string of contiguous pages can be found after coalescing old pages from
419 * many old transactions together.
421 * @todo We don't actually do such coalescing yet, we grab pages from one
422 * old transaction at a time.
425 /** Number of slots in the reader table.
426 * This value was chosen somewhat arbitrarily. 126 readers plus a
427 * couple mutexes fit exactly into 8KB on my development machine.
428 * Applications should set the table size using #mdb_env_set_maxreaders().
430 #define DEFAULT_READERS 126
432 /** The size of a CPU cache line in bytes. We want our lock structures
433 * aligned to this size to avoid false cache line sharing in the
435 * This value works for most CPUs. For Itanium this should be 128.
441 /** The information we store in a single slot of the reader table.
442 * In addition to a transaction ID, we also record the process and
443 * thread ID that owns a slot, so that we can detect stale information,
444 * e.g. threads or processes that went away without cleaning up.
445 * @note We currently don't check for stale records. We simply re-init
446 * the table when we know that we're the only process opening the
449 typedef struct MDB_rxbody {
450 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
451 * Multiple readers that start at the same time will probably have the
452 * same ID here. Again, it's not important to exclude them from
453 * anything; all we need to know is which version of the DB they
454 * started from so we can avoid overwriting any data used in that
455 * particular version.
458 /** The process ID of the process owning this reader txn. */
460 /** The thread ID of the thread owning this txn. */
464 /** The actual reader record, with cacheline padding. */
465 typedef struct MDB_reader {
468 /** shorthand for mrb_txnid */
469 #define mr_txnid mru.mrx.mrb_txnid
470 #define mr_pid mru.mrx.mrb_pid
471 #define mr_tid mru.mrx.mrb_tid
472 /** cache line alignment */
473 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
477 /** The header for the reader table.
478 * The table resides in a memory-mapped file. (This is a different file
479 * than is used for the main database.)
481 * For POSIX the actual mutexes reside in the shared memory of this
482 * mapped file. On Windows, mutexes are named objects allocated by the
483 * kernel; we store the mutex names in this mapped file so that other
484 * processes can grab them. This same approach is also used on
485 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
486 * process-shared POSIX mutexes. For these cases where a named object
487 * is used, the object name is derived from a 64 bit FNV hash of the
488 * environment pathname. As such, naming collisions are extremely
489 * unlikely. If a collision occurs, the results are unpredictable.
491 typedef struct MDB_txbody {
492 /** Stamp identifying this as an MDB file. It must be set
495 /** Version number of this lock file. Must be set to #MDB_VERSION. */
496 uint32_t mtb_version;
497 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
498 char mtb_rmname[MNAME_LEN];
500 /** Mutex protecting access to this table.
501 * This is the reader lock that #LOCK_MUTEX_R acquires.
503 pthread_mutex_t mtb_mutex;
505 /** The ID of the last transaction committed to the database.
506 * This is recorded here only for convenience; the value can always
507 * be determined by reading the main database meta pages.
510 /** The number of slots that have been used in the reader table.
511 * This always records the maximum count, it is not decremented
512 * when readers release their slots.
514 unsigned mtb_numreaders;
517 /** The actual reader table definition. */
518 typedef struct MDB_txninfo {
521 #define mti_magic mt1.mtb.mtb_magic
522 #define mti_version mt1.mtb.mtb_version
523 #define mti_mutex mt1.mtb.mtb_mutex
524 #define mti_rmname mt1.mtb.mtb_rmname
525 #define mti_txnid mt1.mtb.mtb_txnid
526 #define mti_numreaders mt1.mtb.mtb_numreaders
527 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
530 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
531 char mt2_wmname[MNAME_LEN];
532 #define mti_wmname mt2.mt2_wmname
534 pthread_mutex_t mt2_wmutex;
535 #define mti_wmutex mt2.mt2_wmutex
537 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
539 MDB_reader mti_readers[1];
543 /** Common header for all page types.
544 * Overflow records occupy a number of contiguous pages with no
545 * headers on any page after the first.
547 typedef struct MDB_page {
548 #define mp_pgno mp_p.p_pgno
549 #define mp_next mp_p.p_next
551 pgno_t p_pgno; /**< page number */
552 void * p_next; /**< for in-memory list of freed structs */
555 /** @defgroup mdb_page Page Flags
557 * Flags for the page headers.
560 #define P_BRANCH 0x01 /**< branch page */
561 #define P_LEAF 0x02 /**< leaf page */
562 #define P_OVERFLOW 0x04 /**< overflow page */
563 #define P_META 0x08 /**< meta page */
564 #define P_DIRTY 0x10 /**< dirty page */
565 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
566 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
568 uint16_t mp_flags; /**< @ref mdb_page */
569 #define mp_lower mp_pb.pb.pb_lower
570 #define mp_upper mp_pb.pb.pb_upper
571 #define mp_pages mp_pb.pb_pages
574 indx_t pb_lower; /**< lower bound of free space */
575 indx_t pb_upper; /**< upper bound of free space */
577 uint32_t pb_pages; /**< number of overflow pages */
579 indx_t mp_ptrs[1]; /**< dynamic size */
582 /** Size of the page header, excluding dynamic data at the end */
583 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
585 /** Address of first usable data byte in a page, after the header */
586 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
588 /** Number of nodes on a page */
589 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
591 /** The amount of space remaining in the page */
592 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
594 /** The percentage of space used in the page, in tenths of a percent. */
595 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
596 ((env)->me_psize - PAGEHDRSZ))
597 /** The minimum page fill factor, in tenths of a percent.
598 * Pages emptier than this are candidates for merging.
600 #define FILL_THRESHOLD 250
602 /** Test if a page is a leaf page */
603 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
604 /** Test if a page is a LEAF2 page */
605 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
606 /** Test if a page is a branch page */
607 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
608 /** Test if a page is an overflow page */
609 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
610 /** Test if a page is a sub page */
611 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
613 /** The number of overflow pages needed to store the given size. */
614 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
616 /** Header for a single key/data pair within a page.
617 * We guarantee 2-byte alignment for nodes.
619 typedef struct MDB_node {
620 /** lo and hi are used for data size on leaf nodes and for
621 * child pgno on branch nodes. On 64 bit platforms, flags
622 * is also used for pgno. (Branch nodes have no flags).
623 * They are in host byte order in case that lets some
624 * accesses be optimized into a 32-bit word access.
626 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
627 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
628 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
629 /** @defgroup mdb_node Node Flags
631 * Flags for node headers.
634 #define F_BIGDATA 0x01 /**< data put on overflow page */
635 #define F_SUBDATA 0x02 /**< data is a sub-database */
636 #define F_DUPDATA 0x04 /**< data has duplicates */
638 /** valid flags for #mdb_node_add() */
639 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
642 unsigned short mn_flags; /**< @ref mdb_node */
643 unsigned short mn_ksize; /**< key size */
644 char mn_data[1]; /**< key and data are appended here */
647 /** Size of the node header, excluding dynamic data at the end */
648 #define NODESIZE offsetof(MDB_node, mn_data)
650 /** Bit position of top word in page number, for shifting mn_flags */
651 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
653 /** Size of a node in a branch page with a given key.
654 * This is just the node header plus the key, there is no data.
656 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
658 /** Size of a node in a leaf page with a given key and data.
659 * This is node header plus key plus data size.
661 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
663 /** Address of node \b i in page \b p */
664 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
666 /** Address of the key for the node */
667 #define NODEKEY(node) (void *)((node)->mn_data)
669 /** Address of the data for a node */
670 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
672 /** Get the page number pointed to by a branch node */
673 #define NODEPGNO(node) \
674 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
675 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
676 /** Set the page number in a branch node */
677 #define SETPGNO(node,pgno) do { \
678 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
679 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
681 /** Get the size of the data in a leaf node */
682 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
683 /** Set the size of the data for a leaf node */
684 #define SETDSZ(node,size) do { \
685 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
686 /** The size of a key in a node */
687 #define NODEKSZ(node) ((node)->mn_ksize)
689 /** Copy a page number from src to dst */
691 #define COPY_PGNO(dst,src) dst = src
693 #if SIZE_MAX > 4294967295UL
694 #define COPY_PGNO(dst,src) do { \
695 unsigned short *s, *d; \
696 s = (unsigned short *)&(src); \
697 d = (unsigned short *)&(dst); \
704 #define COPY_PGNO(dst,src) do { \
705 unsigned short *s, *d; \
706 s = (unsigned short *)&(src); \
707 d = (unsigned short *)&(dst); \
713 /** The address of a key in a LEAF2 page.
714 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
715 * There are no node headers, keys are stored contiguously.
717 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
719 /** Set the \b node's key into \b key, if requested. */
720 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
721 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
723 /** Information about a single database in the environment. */
724 typedef struct MDB_db {
725 uint32_t md_pad; /**< also ksize for LEAF2 pages */
726 uint16_t md_flags; /**< @ref mdb_dbi_open */
727 uint16_t md_depth; /**< depth of this tree */
728 pgno_t md_branch_pages; /**< number of internal pages */
729 pgno_t md_leaf_pages; /**< number of leaf pages */
730 pgno_t md_overflow_pages; /**< number of overflow pages */
731 size_t md_entries; /**< number of data items */
732 pgno_t md_root; /**< the root page of this tree */
735 /** Handle for the DB used to track free pages. */
737 /** Handle for the default DB. */
740 /** Meta page content. */
741 typedef struct MDB_meta {
742 /** Stamp identifying this as an MDB file. It must be set
745 /** Version number of this lock file. Must be set to #MDB_VERSION. */
747 void *mm_address; /**< address for fixed mapping */
748 size_t mm_mapsize; /**< size of mmap region */
749 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
750 /** The size of pages used in this DB */
751 #define mm_psize mm_dbs[0].md_pad
752 /** Any persistent environment flags. @ref mdb_env */
753 #define mm_flags mm_dbs[0].md_flags
754 pgno_t mm_last_pg; /**< last used page in file */
755 txnid_t mm_txnid; /**< txnid that committed this page */
758 /** Buffer for a stack-allocated dirty page.
759 * The members define size and alignment, and silence type
760 * aliasing warnings. They are not used directly; that could
761 * mean incorrectly using several union members in parallel.
763 typedef union MDB_pagebuf {
764 char mb_raw[MDB_PAGESIZE];
767 char mm_pad[PAGEHDRSZ];
772 /** Auxiliary DB info.
773 * The information here is mostly static/read-only. There is
774 * only a single copy of this record in the environment.
776 typedef struct MDB_dbx {
777 MDB_val md_name; /**< name of the database */
778 MDB_cmp_func *md_cmp; /**< function for comparing keys */
779 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
780 MDB_rel_func *md_rel; /**< user relocate function */
781 void *md_relctx; /**< user-provided context for md_rel */
784 /** A database transaction.
785 * Every operation requires a transaction handle.
788 MDB_txn *mt_parent; /**< parent of a nested txn */
789 MDB_txn *mt_child; /**< nested txn under this txn */
790 pgno_t mt_next_pgno; /**< next unallocated page */
791 /** The ID of this transaction. IDs are integers incrementing from 1.
792 * Only committed write transactions increment the ID. If a transaction
793 * aborts, the ID may be re-used by the next writer.
796 MDB_env *mt_env; /**< the DB environment */
797 /** The list of pages that became unused during this transaction.
801 MDB_ID2L dirty_list; /**< modified pages */
802 MDB_reader *reader; /**< this thread's slot in the reader table */
804 /** Array of records for each DB known in the environment. */
806 /** Array of MDB_db records for each known DB */
808 /** @defgroup mt_dbflag Transaction DB Flags
812 #define DB_DIRTY 0x01 /**< DB was written in this txn */
813 #define DB_STALE 0x02 /**< DB record is older than txnID */
815 /** Array of cursors for each DB */
816 MDB_cursor **mt_cursors;
817 /** Array of flags for each DB */
818 unsigned char *mt_dbflags;
819 /** Number of DB records in use. This number only ever increments;
820 * we don't decrement it when individual DB handles are closed.
824 /** @defgroup mdb_txn Transaction Flags
828 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
829 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
830 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
832 unsigned int mt_flags; /**< @ref mdb_txn */
833 /** Tracks which of the two meta pages was used at the start
834 * of this transaction.
836 unsigned int mt_toggle;
839 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
840 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
841 * raise this on a 64 bit machine.
843 #define CURSOR_STACK 32
847 /** Cursors are used for all DB operations */
849 /** Next cursor on this DB in this txn */
851 /** Original cursor if this is a shadow */
853 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
854 struct MDB_xcursor *mc_xcursor;
855 /** The transaction that owns this cursor */
857 /** The database handle this cursor operates on */
859 /** The database record for this cursor */
861 /** The database auxiliary record for this cursor */
863 /** The @ref mt_dbflag for this database */
864 unsigned char *mc_dbflag;
865 unsigned short mc_snum; /**< number of pushed pages */
866 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
867 /** @defgroup mdb_cursor Cursor Flags
869 * Cursor state flags.
872 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
873 #define C_EOF 0x02 /**< No more data */
874 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
875 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
876 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
877 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
879 unsigned int mc_flags; /**< @ref mdb_cursor */
880 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
881 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
884 /** Context for sorted-dup records.
885 * We could have gone to a fully recursive design, with arbitrarily
886 * deep nesting of sub-databases. But for now we only handle these
887 * levels - main DB, optional sub-DB, sorted-duplicate DB.
889 typedef struct MDB_xcursor {
890 /** A sub-cursor for traversing the Dup DB */
891 MDB_cursor mx_cursor;
892 /** The database record for this Dup DB */
894 /** The auxiliary DB record for this Dup DB */
896 /** The @ref mt_dbflag for this Dup DB */
897 unsigned char mx_dbflag;
900 /** A set of pages freed by an earlier transaction. */
901 typedef struct MDB_oldpages {
902 /** Usually we only read one record from the FREEDB at a time, but
903 * in case we read more, this will chain them together.
905 struct MDB_oldpages *mo_next;
906 /** The ID of the transaction in which these pages were freed. */
908 /** An #MDB_IDL of the pages */
909 pgno_t mo_pages[1]; /* dynamic */
912 /** The database environment. */
914 HANDLE me_fd; /**< The main data file */
915 HANDLE me_lfd; /**< The lock file */
916 HANDLE me_mfd; /**< just for writing the meta pages */
917 /** Failed to update the meta page. Probably an I/O error. */
918 #define MDB_FATAL_ERROR 0x80000000U
919 /** Read-only Filesystem. Allow read access, no locking. */
920 #define MDB_ROFS 0x40000000U
921 /** Some fields are initialized. */
922 #define MDB_ENV_ACTIVE 0x20000000U
923 uint32_t me_flags; /**< @ref mdb_env */
924 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
925 unsigned int me_maxreaders; /**< size of the reader table */
926 unsigned int me_numreaders; /**< max numreaders set by this env */
927 MDB_dbi me_numdbs; /**< number of DBs opened */
928 MDB_dbi me_maxdbs; /**< size of the DB table */
929 pid_t me_pid; /**< process ID of this env */
930 char *me_path; /**< path to the DB files */
931 char *me_map; /**< the memory map of the data file */
932 MDB_txninfo *me_txns; /**< the memory map of the lock file */
933 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
934 MDB_txn *me_txn; /**< current write transaction */
935 size_t me_mapsize; /**< size of the data memory map */
936 off_t me_size; /**< current file size */
937 pgno_t me_maxpg; /**< me_mapsize / me_psize */
938 txnid_t me_pgfirst; /**< ID of first old page record we used */
939 txnid_t me_pglast; /**< ID of last old page record we used */
940 MDB_dbx *me_dbxs; /**< array of static DB info */
941 uint16_t *me_dbflags; /**< array of DB flags */
942 MDB_oldpages *me_pghead; /**< list of old page records */
943 MDB_oldpages *me_pgfree; /**< list of page records to free */
944 pthread_key_t me_txkey; /**< thread-key for readers */
945 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
946 /** IDL of pages that became unused in a write txn */
948 /** ID2L of pages that were written during a write txn */
949 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
951 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
953 #elif defined(MDB_USE_POSIX_SEM)
954 sem_t *me_rmutex; /* Shared mutexes are not supported */
958 /** max number of pages to commit in one writev() call */
959 #define MDB_COMMIT_PAGES 64
960 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
961 #undef MDB_COMMIT_PAGES
962 #define MDB_COMMIT_PAGES IOV_MAX
965 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
966 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
967 static int mdb_page_touch(MDB_cursor *mc);
969 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
970 static int mdb_page_search_root(MDB_cursor *mc,
971 MDB_val *key, int modify);
972 #define MDB_PS_MODIFY 1
973 #define MDB_PS_ROOTONLY 2
974 static int mdb_page_search(MDB_cursor *mc,
975 MDB_val *key, int flags);
976 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
978 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
979 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
980 pgno_t newpgno, unsigned int nflags);
982 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
983 static int mdb_env_pick_meta(const MDB_env *env);
984 static int mdb_env_write_meta(MDB_txn *txn);
985 static void mdb_env_close0(MDB_env *env, int excl);
987 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
988 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
989 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
990 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
991 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
992 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
993 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
994 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
995 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
997 static int mdb_rebalance(MDB_cursor *mc);
998 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
1000 static void mdb_cursor_pop(MDB_cursor *mc);
1001 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1003 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1004 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1005 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1006 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1007 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1009 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1010 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1012 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1013 static void mdb_xcursor_init0(MDB_cursor *mc);
1014 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1016 static int mdb_drop0(MDB_cursor *mc, int subs);
1017 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1020 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1024 static SECURITY_DESCRIPTOR mdb_null_sd;
1025 static SECURITY_ATTRIBUTES mdb_all_sa;
1026 static int mdb_sec_inited;
1029 /** Return the library version info. */
1031 mdb_version(int *major, int *minor, int *patch)
1033 if (major) *major = MDB_VERSION_MAJOR;
1034 if (minor) *minor = MDB_VERSION_MINOR;
1035 if (patch) *patch = MDB_VERSION_PATCH;
1036 return MDB_VERSION_STRING;
1039 /** Table of descriptions for MDB @ref errors */
1040 static char *const mdb_errstr[] = {
1041 "MDB_KEYEXIST: Key/data pair already exists",
1042 "MDB_NOTFOUND: No matching key/data pair found",
1043 "MDB_PAGE_NOTFOUND: Requested page not found",
1044 "MDB_CORRUPTED: Located page was wrong type",
1045 "MDB_PANIC: Update of meta page failed",
1046 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1047 "MDB_INVALID: File is not an MDB file",
1048 "MDB_MAP_FULL: Environment mapsize limit reached",
1049 "MDB_DBS_FULL: Environment maxdbs limit reached",
1050 "MDB_READERS_FULL: Environment maxreaders limit reached",
1051 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1052 "MDB_TXN_FULL: Nested transaction has too many dirty pages - transaction too big",
1053 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1054 "MDB_PAGE_FULL: Internal error - page has no more space"
1058 mdb_strerror(int err)
1062 return ("Successful return: 0");
1064 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1065 i = err - MDB_KEYEXIST;
1066 return mdb_errstr[i];
1069 return strerror(err);
1073 /** Display a key in hexadecimal and return the address of the result.
1074 * @param[in] key the key to display
1075 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1076 * @return The key in hexadecimal form.
1079 mdb_dkey(MDB_val *key, char *buf)
1082 unsigned char *c = key->mv_data;
1084 if (key->mv_size > MAXKEYSIZE)
1085 return "MAXKEYSIZE";
1086 /* may want to make this a dynamic check: if the key is mostly
1087 * printable characters, print it as-is instead of converting to hex.
1091 for (i=0; i<key->mv_size; i++)
1092 ptr += sprintf(ptr, "%02x", *c++);
1094 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1099 /** Display all the keys in the page. */
1101 mdb_page_list(MDB_page *mp)
1104 unsigned int i, nkeys, nsize;
1108 nkeys = NUMKEYS(mp);
1109 fprintf(stderr, "numkeys %d\n", nkeys);
1110 for (i=0; i<nkeys; i++) {
1111 node = NODEPTR(mp, i);
1112 key.mv_size = node->mn_ksize;
1113 key.mv_data = node->mn_data;
1114 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1115 if (F_ISSET(node->mn_flags, F_BIGDATA))
1116 nsize += sizeof(pgno_t);
1118 nsize += NODEDSZ(node);
1119 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1124 mdb_cursor_chk(MDB_cursor *mc)
1130 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1131 for (i=0; i<mc->mc_top; i++) {
1133 node = NODEPTR(mp, mc->mc_ki[i]);
1134 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1137 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1143 /** Count all the pages in each DB and in the freelist
1144 * and make sure it matches the actual number of pages
1147 static void mdb_audit(MDB_txn *txn)
1151 MDB_ID freecount, count;
1156 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1157 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1158 freecount += *(MDB_ID *)data.mv_data;
1161 for (i = 0; i<txn->mt_numdbs; i++) {
1162 MDB_xcursor mx, *mxp;
1163 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1164 mdb_cursor_init(&mc, txn, i, mxp);
1165 if (txn->mt_dbs[i].md_root == P_INVALID)
1167 count += txn->mt_dbs[i].md_branch_pages +
1168 txn->mt_dbs[i].md_leaf_pages +
1169 txn->mt_dbs[i].md_overflow_pages;
1170 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1171 mdb_page_search(&mc, NULL, 0);
1175 mp = mc.mc_pg[mc.mc_top];
1176 for (j=0; j<NUMKEYS(mp); j++) {
1177 MDB_node *leaf = NODEPTR(mp, j);
1178 if (leaf->mn_flags & F_SUBDATA) {
1180 memcpy(&db, NODEDATA(leaf), sizeof(db));
1181 count += db.md_branch_pages + db.md_leaf_pages +
1182 db.md_overflow_pages;
1186 while (mdb_cursor_sibling(&mc, 1) == 0);
1189 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1190 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1191 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1197 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1199 return txn->mt_dbxs[dbi].md_cmp(a, b);
1203 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1205 if (txn->mt_dbxs[dbi].md_dcmp)
1206 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1208 return EINVAL; /* too bad you can't distinguish this from a valid result */
1211 /** Allocate a single page.
1212 * Re-use old malloc'd pages first, otherwise just malloc.
1215 mdb_page_malloc(MDB_cursor *mc) {
1217 size_t sz = mc->mc_txn->mt_env->me_psize;
1218 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1219 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1220 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1221 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1222 } else if ((ret = malloc(sz)) != NULL) {
1223 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1228 /** Allocate pages for writing.
1229 * If there are free pages available from older transactions, they
1230 * will be re-used first. Otherwise a new page will be allocated.
1231 * @param[in] mc cursor A cursor handle identifying the transaction and
1232 * database for which we are allocating.
1233 * @param[in] num the number of pages to allocate.
1234 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1235 * will always be satisfied by a single contiguous chunk of memory.
1236 * @return 0 on success, non-zero on failure.
1239 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1241 MDB_txn *txn = mc->mc_txn;
1243 pgno_t pgno = P_INVALID;
1245 txnid_t oldest = 0, last;
1249 /* The free list won't have any content at all until txn 2 has
1250 * committed. The pages freed by txn 2 will be unreferenced
1251 * after txn 3 commits, and so will be safe to re-use in txn 4.
1253 if (txn->mt_txnid > 3) {
1255 if (!txn->mt_env->me_pghead &&
1256 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1257 /* See if there's anything in the free DB */
1264 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1265 if (!txn->mt_env->me_pgfirst) {
1266 mdb_page_search(&m2, NULL, 0);
1267 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1268 kptr = (txnid_t *)NODEKEY(leaf);
1275 last = txn->mt_env->me_pglast + 1;
1277 key.mv_data = &last;
1278 key.mv_size = sizeof(last);
1279 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1282 last = *(txnid_t *)key.mv_data;
1288 oldest = txn->mt_txnid - 1;
1289 nr = txn->mt_env->me_txns->mti_numreaders;
1290 r = txn->mt_env->me_txns->mti_readers;
1291 for (i=0; i<nr; i++) {
1292 if (!r[i].mr_pid) continue;
1299 if (oldest > last) {
1300 /* It's usable, grab it.
1305 if (!txn->mt_env->me_pgfirst) {
1306 mdb_node_read(txn, leaf, &data);
1308 txn->mt_env->me_pglast = last;
1309 if (!txn->mt_env->me_pgfirst)
1310 txn->mt_env->me_pgfirst = last;
1311 idl = (MDB_ID *) data.mv_data;
1312 /* We might have a zero-length IDL due to freelist growth
1313 * during a prior commit
1315 if (!idl[0]) goto again;
1316 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1319 mop->mo_next = txn->mt_env->me_pghead;
1320 mop->mo_txnid = last;
1321 txn->mt_env->me_pghead = mop;
1322 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1327 DPRINTF("IDL read txn %zu root %zu num %zu",
1328 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1329 for (i=0; i<idl[0]; i++) {
1330 DPRINTF("IDL %zu", idl[i+1]);
1337 if (txn->mt_env->me_pghead) {
1338 MDB_oldpages *mop = txn->mt_env->me_pghead;
1341 int retry = 60, readit = 0, n2 = num-1;
1342 unsigned int i, j, k;
1344 /* If current list is too short, must fetch more and coalesce */
1345 if (mop->mo_pages[0] < (unsigned)num)
1348 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1350 /* bail out if we're operating on the freelist.
1351 * TODO: get all of this working. Many circular dependencies...
1353 if (mc->mc_dbi == FREE_DBI)
1361 last = mop->mo_txnid + 1;
1363 /* We haven't hit the readers list yet? */
1369 oldest = txn->mt_txnid - 1;
1370 nr = txn->mt_env->me_txns->mti_numreaders;
1371 r = txn->mt_env->me_txns->mti_readers;
1372 for (i=0; i<nr; i++) {
1373 if (!r[i].mr_pid) continue;
1380 /* There's nothing we can use on the freelist */
1381 if (oldest - last < 1)
1385 key.mv_data = &last;
1386 key.mv_size = sizeof(last);
1387 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1390 idl = (MDB_ID *) data.mv_data;
1391 mop2 = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - 2*sizeof(pgno_t) + MDB_IDL_SIZEOF(mop->mo_pages));
1394 /* merge in sorted order */
1395 i = idl[0]; j = mop->mo_pages[0]; mop2->mo_pages[0] = k = i+j;
1396 mop->mo_pages[0] = P_INVALID;
1397 while (i>0 || j>0) {
1398 if (i && idl[i] < mop->mo_pages[j])
1399 mop2->mo_pages[k--] = idl[i--];
1401 mop2->mo_pages[k--] = mop->mo_pages[j--];
1403 txn->mt_env->me_pglast = last;
1404 mop2->mo_txnid = last;
1405 mop2->mo_next = mop->mo_next;
1406 txn->mt_env->me_pghead = mop2;
1409 /* Keep trying to read until we have enough */
1410 if (mop->mo_pages[0] < (unsigned)num) {
1415 /* current list has enough pages, but are they contiguous? */
1416 for (i=mop->mo_pages[0]; i>=(unsigned)num; i--) {
1417 if (mop->mo_pages[i-n2] == mop->mo_pages[i] + n2) {
1418 pgno = mop->mo_pages[i];
1420 /* move any stragglers down */
1421 for (j=i+num; j<=mop->mo_pages[0]; j++)
1422 mop->mo_pages[i++] = mop->mo_pages[j];
1423 mop->mo_pages[0] -= num;
1428 /* Stop if we succeeded, or no more retries */
1429 if (!retry || pgno != P_INVALID)
1436 /* peel pages off tail, so we only have to truncate the list */
1437 pgno = MDB_IDL_LAST(mop->mo_pages);
1440 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1441 txn->mt_env->me_pghead = mop->mo_next;
1442 if (mc->mc_dbi == FREE_DBI) {
1443 mop->mo_next = txn->mt_env->me_pgfree;
1444 txn->mt_env->me_pgfree = mop;
1452 if (pgno == P_INVALID) {
1453 /* DB size is maxed out */
1454 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1455 DPUTS("DB size maxed out");
1456 return MDB_MAP_FULL;
1459 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1460 if (pgno == P_INVALID) {
1461 pgno = txn->mt_next_pgno;
1462 txn->mt_next_pgno += num;
1464 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1467 if (txn->mt_env->me_dpages && num == 1) {
1468 np = txn->mt_env->me_dpages;
1469 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1470 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1471 txn->mt_env->me_dpages = np->mp_next;
1473 size_t sz = txn->mt_env->me_psize * num;
1474 if ((np = malloc(sz)) == NULL)
1476 VGMEMP_ALLOC(txn->mt_env, np, sz);
1478 if (pgno == P_INVALID) {
1479 np->mp_pgno = txn->mt_next_pgno;
1480 txn->mt_next_pgno += num;
1485 mid.mid = np->mp_pgno;
1487 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1488 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1490 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1497 /** Copy a page: avoid copying unused portions of the page.
1498 * @param[in] dst page to copy into
1499 * @param[in] src page to copy from
1502 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1504 dst->mp_flags = src->mp_flags | P_DIRTY;
1505 dst->mp_pages = src->mp_pages;
1507 if (IS_LEAF2(src)) {
1508 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1510 unsigned int i, nkeys = NUMKEYS(src);
1511 for (i=0; i<nkeys; i++)
1512 dst->mp_ptrs[i] = src->mp_ptrs[i];
1513 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1514 psize - src->mp_upper);
1518 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1519 * @param[in] mc cursor pointing to the page to be touched
1520 * @return 0 on success, non-zero on failure.
1523 mdb_page_touch(MDB_cursor *mc)
1525 MDB_page *mp = mc->mc_pg[mc->mc_top];
1529 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1531 if ((rc = mdb_page_alloc(mc, 1, &np)))
1533 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1534 assert(mp->mp_pgno != np->mp_pgno);
1535 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1537 /* If page isn't full, just copy the used portion */
1538 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1541 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1543 np->mp_flags |= P_DIRTY;
1548 /* Adjust other cursors pointing to mp */
1549 if (mc->mc_flags & C_SUB) {
1550 MDB_cursor *m2, *m3;
1551 MDB_dbi dbi = mc->mc_dbi-1;
1553 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1554 if (m2 == mc) continue;
1555 m3 = &m2->mc_xcursor->mx_cursor;
1556 if (m3->mc_snum < mc->mc_snum) continue;
1557 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1558 m3->mc_pg[mc->mc_top] = mp;
1564 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1565 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1566 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1567 m2->mc_pg[mc->mc_top] = mp;
1571 mc->mc_pg[mc->mc_top] = mp;
1572 /** If this page has a parent, update the parent to point to
1576 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1578 mc->mc_db->md_root = mp->mp_pgno;
1579 } else if (mc->mc_txn->mt_parent) {
1582 /* If txn has a parent, make sure the page is in our
1585 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1586 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1587 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1588 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1589 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1590 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1591 mc->mc_pg[mc->mc_top] = mp;
1597 np = mdb_page_malloc(mc);
1600 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1601 mid.mid = np->mp_pgno;
1603 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1611 mdb_env_sync(MDB_env *env, int force)
1614 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1615 if (env->me_flags & MDB_WRITEMAP) {
1616 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1617 ? MS_ASYNC : MS_SYNC;
1618 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1621 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1625 if (MDB_FDATASYNC(env->me_fd))
1632 /** Make shadow copies of all of parent txn's cursors */
1634 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1636 MDB_cursor *mc, *m2;
1637 unsigned int i, j, size;
1639 for (i=0;i<src->mt_numdbs; i++) {
1640 if (src->mt_cursors[i]) {
1641 size = sizeof(MDB_cursor);
1642 if (src->mt_cursors[i]->mc_xcursor)
1643 size += sizeof(MDB_xcursor);
1644 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1651 mc->mc_db = &dst->mt_dbs[i];
1652 mc->mc_dbx = m2->mc_dbx;
1653 mc->mc_dbflag = &dst->mt_dbflags[i];
1654 mc->mc_snum = m2->mc_snum;
1655 mc->mc_top = m2->mc_top;
1656 mc->mc_flags = m2->mc_flags | C_SHADOW;
1657 for (j=0; j<mc->mc_snum; j++) {
1658 mc->mc_pg[j] = m2->mc_pg[j];
1659 mc->mc_ki[j] = m2->mc_ki[j];
1661 if (m2->mc_xcursor) {
1662 MDB_xcursor *mx, *mx2;
1663 mx = (MDB_xcursor *)(mc+1);
1664 mc->mc_xcursor = mx;
1665 mx2 = m2->mc_xcursor;
1666 mx->mx_db = mx2->mx_db;
1667 mx->mx_dbx = mx2->mx_dbx;
1668 mx->mx_dbflag = mx2->mx_dbflag;
1669 mx->mx_cursor.mc_txn = dst;
1670 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1671 mx->mx_cursor.mc_db = &mx->mx_db;
1672 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1673 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1674 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1675 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1676 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1677 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1678 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1679 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1682 mc->mc_xcursor = NULL;
1684 mc->mc_next = dst->mt_cursors[i];
1685 dst->mt_cursors[i] = mc;
1692 /** Merge shadow cursors back into parent's */
1694 mdb_cursor_merge(MDB_txn *txn)
1697 for (i=0; i<txn->mt_numdbs; i++) {
1698 if (txn->mt_cursors[i]) {
1700 while ((mc = txn->mt_cursors[i])) {
1701 txn->mt_cursors[i] = mc->mc_next;
1702 if (mc->mc_flags & C_SHADOW) {
1703 MDB_cursor *m2 = mc->mc_orig;
1705 m2->mc_snum = mc->mc_snum;
1706 m2->mc_top = mc->mc_top;
1707 for (j=0; j<mc->mc_snum; j++) {
1708 m2->mc_pg[j] = mc->mc_pg[j];
1709 m2->mc_ki[j] = mc->mc_ki[j];
1712 if (mc->mc_flags & C_ALLOCD)
1720 mdb_txn_reset0(MDB_txn *txn);
1722 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1723 * @param[in] txn the transaction handle to initialize
1724 * @return 0 on success, non-zero on failure. This can only
1725 * fail for read-only transactions, and then only if the
1726 * reader table is full.
1729 mdb_txn_renew0(MDB_txn *txn)
1731 MDB_env *env = txn->mt_env;
1736 txn->mt_numdbs = env->me_numdbs;
1737 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1739 if (txn->mt_flags & MDB_TXN_RDONLY) {
1740 if (env->me_flags & MDB_ROFS) {
1741 i = mdb_env_pick_meta(env);
1742 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1743 txn->mt_u.reader = NULL;
1745 MDB_reader *r = pthread_getspecific(env->me_txkey);
1747 pid_t pid = env->me_pid;
1748 pthread_t tid = pthread_self();
1751 for (i=0; i<env->me_txns->mti_numreaders; i++)
1752 if (env->me_txns->mti_readers[i].mr_pid == 0)
1754 if (i == env->me_maxreaders) {
1755 UNLOCK_MUTEX_R(env);
1756 return MDB_READERS_FULL;
1758 env->me_txns->mti_readers[i].mr_pid = pid;
1759 env->me_txns->mti_readers[i].mr_tid = tid;
1760 if (i >= env->me_txns->mti_numreaders)
1761 env->me_txns->mti_numreaders = i+1;
1762 /* Save numreaders for un-mutexed mdb_env_close() */
1763 env->me_numreaders = env->me_txns->mti_numreaders;
1764 UNLOCK_MUTEX_R(env);
1765 r = &env->me_txns->mti_readers[i];
1766 if ((rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1767 env->me_txns->mti_readers[i].mr_pid = 0;
1771 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1772 txn->mt_u.reader = r;
1774 txn->mt_toggle = txn->mt_txnid & 1;
1775 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1779 txn->mt_txnid = env->me_txns->mti_txnid;
1780 txn->mt_toggle = txn->mt_txnid & 1;
1781 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1784 if (txn->mt_txnid == mdb_debug_start)
1787 txn->mt_u.dirty_list = env->me_dirty_list;
1788 txn->mt_u.dirty_list[0].mid = 0;
1789 txn->mt_free_pgs = env->me_free_pgs;
1790 txn->mt_free_pgs[0] = 0;
1794 /* Copy the DB info and flags */
1795 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1796 for (i=2; i<txn->mt_numdbs; i++)
1797 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1798 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1799 if (txn->mt_numdbs > 2)
1800 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1806 mdb_txn_renew(MDB_txn *txn)
1810 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1813 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1814 DPUTS("environment had fatal error, must shutdown!");
1818 rc = mdb_txn_renew0(txn);
1819 if (rc == MDB_SUCCESS) {
1820 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1821 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1822 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1828 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1833 if (env->me_flags & MDB_FATAL_ERROR) {
1834 DPUTS("environment had fatal error, must shutdown!");
1837 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1840 /* Nested transactions: Max 1 child, write txns only, no writemap */
1841 if (parent->mt_child ||
1842 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1843 (env->me_flags & MDB_WRITEMAP))
1848 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1849 if (!(flags & MDB_RDONLY))
1850 size += env->me_maxdbs * sizeof(MDB_cursor *);
1852 if ((txn = calloc(1, size)) == NULL) {
1853 DPRINTF("calloc: %s", strerror(ErrCode()));
1856 txn->mt_dbs = (MDB_db *)(txn+1);
1857 if (flags & MDB_RDONLY) {
1858 txn->mt_flags |= MDB_TXN_RDONLY;
1859 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1861 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1862 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1867 txn->mt_free_pgs = mdb_midl_alloc();
1868 if (!txn->mt_free_pgs) {
1872 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1873 if (!txn->mt_u.dirty_list) {
1874 free(txn->mt_free_pgs);
1878 txn->mt_txnid = parent->mt_txnid;
1879 txn->mt_toggle = parent->mt_toggle;
1880 txn->mt_u.dirty_list[0].mid = 0;
1881 txn->mt_free_pgs[0] = 0;
1882 txn->mt_next_pgno = parent->mt_next_pgno;
1883 parent->mt_child = txn;
1884 txn->mt_parent = parent;
1885 txn->mt_numdbs = parent->mt_numdbs;
1886 txn->mt_dbxs = parent->mt_dbxs;
1887 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1888 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1889 mdb_cursor_shadow(parent, txn);
1892 rc = mdb_txn_renew0(txn);
1898 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1899 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1900 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1906 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1907 * @param[in] txn the transaction handle to reset
1910 mdb_txn_reset0(MDB_txn *txn)
1912 MDB_env *env = txn->mt_env;
1914 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1915 if (!(env->me_flags & MDB_ROFS))
1916 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1922 /* close(free) all cursors */
1923 for (i=0; i<txn->mt_numdbs; i++) {
1924 if (txn->mt_cursors[i]) {
1926 while ((mc = txn->mt_cursors[i])) {
1927 txn->mt_cursors[i] = mc->mc_next;
1928 if (mc->mc_flags & C_ALLOCD)
1934 if (!(env->me_flags & MDB_WRITEMAP)) {
1935 /* return all dirty pages to dpage list */
1936 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1937 dp = txn->mt_u.dirty_list[i].mptr;
1938 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1939 dp->mp_next = txn->mt_env->me_dpages;
1940 VGMEMP_FREE(txn->mt_env, dp);
1941 txn->mt_env->me_dpages = dp;
1943 /* large pages just get freed directly */
1944 VGMEMP_FREE(txn->mt_env, dp);
1950 if (txn->mt_parent) {
1951 txn->mt_parent->mt_child = NULL;
1952 mdb_midl_free(txn->mt_free_pgs);
1953 free(txn->mt_u.dirty_list);
1956 if (mdb_midl_shrink(&txn->mt_free_pgs))
1957 env->me_free_pgs = txn->mt_free_pgs;
1960 while ((mop = txn->mt_env->me_pghead)) {
1961 txn->mt_env->me_pghead = mop->mo_next;
1964 txn->mt_env->me_pgfirst = 0;
1965 txn->mt_env->me_pglast = 0;
1968 /* The writer mutex was locked in mdb_txn_begin. */
1969 UNLOCK_MUTEX_W(env);
1974 mdb_txn_reset(MDB_txn *txn)
1979 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1980 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1981 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1983 mdb_txn_reset0(txn);
1987 mdb_txn_abort(MDB_txn *txn)
1992 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1993 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1994 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1997 mdb_txn_abort(txn->mt_child);
1999 mdb_txn_reset0(txn);
2004 mdb_txn_commit(MDB_txn *txn)
2012 pgno_t next, freecnt;
2015 assert(txn != NULL);
2016 assert(txn->mt_env != NULL);
2018 if (txn->mt_child) {
2019 mdb_txn_commit(txn->mt_child);
2020 txn->mt_child = NULL;
2025 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2026 if (txn->mt_numdbs > env->me_numdbs) {
2027 /* update the DB flags */
2029 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2030 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2037 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2038 DPUTS("error flag is set, can't commit");
2040 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2045 if (txn->mt_parent) {
2051 /* Merge (and close) our cursors with parent's */
2052 mdb_cursor_merge(txn);
2054 /* Update parent's DB table */
2055 ip = &txn->mt_parent->mt_dbs[2];
2056 jp = &txn->mt_dbs[2];
2057 for (i = 2; i < txn->mt_numdbs; i++) {
2058 if (ip->md_root != jp->md_root)
2062 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2064 /* Append our free list to parent's */
2065 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
2067 mdb_midl_free(txn->mt_free_pgs);
2069 /* Merge our dirty list with parent's */
2070 dst = txn->mt_parent->mt_u.dirty_list;
2071 src = txn->mt_u.dirty_list;
2072 x = mdb_mid2l_search(dst, src[1].mid);
2073 for (y=1; y<=src[0].mid; y++) {
2074 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
2078 dst[x].mptr = src[y].mptr;
2081 for (; y<=src[0].mid; y++) {
2082 if (++x >= MDB_IDL_UM_MAX) {
2084 return MDB_TXN_FULL;
2089 free(txn->mt_u.dirty_list);
2090 txn->mt_parent->mt_child = NULL;
2095 if (txn != env->me_txn) {
2096 DPUTS("attempt to commit unknown transaction");
2101 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2104 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2105 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2107 /* Update DB root pointers. Their pages have already been
2108 * touched so this is all in-place and cannot fail.
2110 if (txn->mt_numdbs > 2) {
2113 data.mv_size = sizeof(MDB_db);
2115 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2116 for (i = 2; i < txn->mt_numdbs; i++) {
2117 if (txn->mt_dbflags[i] & DB_DIRTY) {
2118 data.mv_data = &txn->mt_dbs[i];
2119 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2124 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2126 /* should only be one record now */
2127 if (env->me_pghead) {
2128 /* make sure first page of freeDB is touched and on freelist */
2129 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2137 /* Delete IDLs we used from the free list */
2138 if (env->me_pgfirst) {
2143 key.mv_size = sizeof(cur);
2144 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
2147 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
2148 rc = mdb_cursor_del(&mc, 0);
2152 env->me_pgfirst = 0;
2156 /* save to free list */
2158 freecnt = txn->mt_free_pgs[0];
2159 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2162 /* make sure last page of freeDB is touched and on freelist */
2163 key.mv_size = MAXKEYSIZE+1;
2165 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2172 MDB_IDL idl = txn->mt_free_pgs;
2173 mdb_midl_sort(txn->mt_free_pgs);
2174 DPRINTF("IDL write txn %zu root %zu num %zu",
2175 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2176 for (i=1; i<=idl[0]; i++) {
2177 DPRINTF("IDL %zu", idl[i]);
2181 /* write to last page of freeDB */
2182 key.mv_size = sizeof(pgno_t);
2183 key.mv_data = &txn->mt_txnid;
2184 data.mv_data = txn->mt_free_pgs;
2185 /* The free list can still grow during this call,
2186 * despite the pre-emptive touches above. So check
2187 * and make sure the entire thing got written.
2190 freecnt = txn->mt_free_pgs[0];
2191 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2192 mdb_midl_sort(txn->mt_free_pgs);
2193 rc = mdb_cursor_put(&mc, &key, &data, 0);
2196 } while (freecnt != txn->mt_free_pgs[0]);
2198 /* should only be one record now */
2200 if (env->me_pghead) {
2206 mop = env->me_pghead;
2208 key.mv_size = sizeof(id);
2210 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2211 data.mv_data = mop->mo_pages;
2212 orig = mop->mo_pages[0];
2213 /* These steps may grow the freelist again
2214 * due to freed overflow pages...
2216 rc = mdb_cursor_put(&mc, &key, &data, 0);
2219 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2220 /* could have been used again here */
2221 if (mop->mo_pages[0] != orig) {
2222 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2223 data.mv_data = mop->mo_pages;
2225 rc = mdb_cursor_put(&mc, &key, &data, 0);
2230 /* was completely used up */
2231 rc = mdb_cursor_del(&mc, 0);
2237 env->me_pgfirst = 0;
2241 /* Check for growth of freelist again */
2242 if (freecnt != txn->mt_free_pgs[0])
2245 if (env->me_pghead) {
2246 free(env->me_pghead);
2247 env->me_pghead = NULL;
2250 while (env->me_pgfree) {
2251 MDB_oldpages *mop = env->me_pgfree;
2252 env->me_pgfree = mop->mo_next;
2256 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2257 if (mdb_midl_shrink(&txn->mt_free_pgs))
2258 env->me_free_pgs = txn->mt_free_pgs;
2265 if (env->me_flags & MDB_WRITEMAP) {
2266 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2267 dp = txn->mt_u.dirty_list[i].mptr;
2268 /* clear dirty flag */
2269 dp->mp_flags &= ~P_DIRTY;
2270 txn->mt_u.dirty_list[i].mid = 0;
2272 txn->mt_u.dirty_list[0].mid = 0;
2276 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2282 /* Windows actually supports scatter/gather I/O, but only on
2283 * unbuffered file handles. Since we're relying on the OS page
2284 * cache for all our data, that's self-defeating. So we just
2285 * write pages one at a time. We use the ov structure to set
2286 * the write offset, to at least save the overhead of a Seek
2290 memset(&ov, 0, sizeof(ov));
2291 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2293 dp = txn->mt_u.dirty_list[i].mptr;
2294 DPRINTF("committing page %zu", dp->mp_pgno);
2295 size = dp->mp_pgno * env->me_psize;
2296 ov.Offset = size & 0xffffffff;
2297 ov.OffsetHigh = size >> 16;
2298 ov.OffsetHigh >>= 16;
2299 /* clear dirty flag */
2300 dp->mp_flags &= ~P_DIRTY;
2301 wsize = env->me_psize;
2302 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2303 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2306 DPRINTF("WriteFile: %d", n);
2313 struct iovec iov[MDB_COMMIT_PAGES];
2317 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2318 dp = txn->mt_u.dirty_list[i].mptr;
2319 if (dp->mp_pgno != next) {
2321 rc = writev(env->me_fd, iov, n);
2325 DPUTS("short write, filesystem full?");
2327 DPRINTF("writev: %s", strerror(n));
2334 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2337 DPRINTF("committing page %zu", dp->mp_pgno);
2338 iov[n].iov_len = env->me_psize;
2339 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2340 iov[n].iov_base = (char *)dp;
2341 size += iov[n].iov_len;
2342 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2343 /* clear dirty flag */
2344 dp->mp_flags &= ~P_DIRTY;
2345 if (++n >= MDB_COMMIT_PAGES) {
2355 rc = writev(env->me_fd, iov, n);
2359 DPUTS("short write, filesystem full?");
2361 DPRINTF("writev: %s", strerror(n));
2368 /* Drop the dirty pages.
2370 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2371 dp = txn->mt_u.dirty_list[i].mptr;
2372 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2373 dp->mp_next = txn->mt_env->me_dpages;
2374 VGMEMP_FREE(txn->mt_env, dp);
2375 txn->mt_env->me_dpages = dp;
2377 VGMEMP_FREE(txn->mt_env, dp);
2380 txn->mt_u.dirty_list[i].mid = 0;
2382 txn->mt_u.dirty_list[0].mid = 0;
2385 if ((n = mdb_env_sync(env, 0)) != 0 ||
2386 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2393 if (txn->mt_numdbs > env->me_numdbs) {
2394 /* update the DB flags */
2396 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2397 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2401 UNLOCK_MUTEX_W(env);
2407 /** Read the environment parameters of a DB environment before
2408 * mapping it into memory.
2409 * @param[in] env the environment handle
2410 * @param[out] meta address of where to store the meta information
2411 * @return 0 on success, non-zero on failure.
2414 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2421 /* We don't know the page size yet, so use a minimum value.
2425 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2427 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2432 else if (rc != MDB_PAGESIZE) {
2436 DPRINTF("read: %s", strerror(err));
2440 p = (MDB_page *)&pbuf;
2442 if (!F_ISSET(p->mp_flags, P_META)) {
2443 DPRINTF("page %zu not a meta page", p->mp_pgno);
2448 if (m->mm_magic != MDB_MAGIC) {
2449 DPUTS("meta has invalid magic");
2453 if (m->mm_version != MDB_VERSION) {
2454 DPRINTF("database is version %u, expected version %u",
2455 m->mm_version, MDB_VERSION);
2456 return MDB_VERSION_MISMATCH;
2459 memcpy(meta, m, sizeof(*m));
2463 /** Write the environment parameters of a freshly created DB environment.
2464 * @param[in] env the environment handle
2465 * @param[out] meta address of where to store the meta information
2466 * @return 0 on success, non-zero on failure.
2469 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2476 DPUTS("writing new meta page");
2478 GET_PAGESIZE(psize);
2480 meta->mm_magic = MDB_MAGIC;
2481 meta->mm_version = MDB_VERSION;
2482 meta->mm_psize = psize;
2483 meta->mm_last_pg = 1;
2484 meta->mm_flags = env->me_flags & 0xffff;
2485 meta->mm_flags |= MDB_INTEGERKEY;
2486 meta->mm_dbs[0].md_root = P_INVALID;
2487 meta->mm_dbs[1].md_root = P_INVALID;
2489 p = calloc(2, psize);
2491 p->mp_flags = P_META;
2494 memcpy(m, meta, sizeof(*meta));
2496 q = (MDB_page *)((char *)p + psize);
2499 q->mp_flags = P_META;
2502 memcpy(m, meta, sizeof(*meta));
2507 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2508 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2511 rc = write(env->me_fd, p, psize * 2);
2512 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2518 /** Update the environment info to commit a transaction.
2519 * @param[in] txn the transaction that's being committed
2520 * @return 0 on success, non-zero on failure.
2523 mdb_env_write_meta(MDB_txn *txn)
2526 MDB_meta meta, metab, *mp;
2528 int rc, len, toggle;
2535 assert(txn != NULL);
2536 assert(txn->mt_env != NULL);
2538 toggle = !txn->mt_toggle;
2539 DPRINTF("writing meta page %d for root page %zu",
2540 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2543 mp = env->me_metas[toggle];
2545 if (env->me_flags & MDB_WRITEMAP) {
2546 /* Persist any increases of mapsize config */
2547 if (env->me_mapsize > mp->mm_mapsize)
2548 mp->mm_mapsize = env->me_mapsize;
2549 mp->mm_dbs[0] = txn->mt_dbs[0];
2550 mp->mm_dbs[1] = txn->mt_dbs[1];
2551 mp->mm_last_pg = txn->mt_next_pgno - 1;
2552 mp->mm_txnid = txn->mt_txnid;
2553 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2554 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2557 ptr += env->me_psize;
2558 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2565 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2566 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2568 ptr = (char *)&meta;
2569 if (env->me_mapsize > mp->mm_mapsize) {
2570 /* Persist any increases of mapsize config */
2571 meta.mm_mapsize = env->me_mapsize;
2572 off = offsetof(MDB_meta, mm_mapsize);
2574 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2576 len = sizeof(MDB_meta) - off;
2579 meta.mm_dbs[0] = txn->mt_dbs[0];
2580 meta.mm_dbs[1] = txn->mt_dbs[1];
2581 meta.mm_last_pg = txn->mt_next_pgno - 1;
2582 meta.mm_txnid = txn->mt_txnid;
2585 off += env->me_psize;
2588 /* Write to the SYNC fd */
2589 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2590 env->me_fd : env->me_mfd;
2593 memset(&ov, 0, sizeof(ov));
2595 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2598 rc = pwrite(mfd, ptr, len, off);
2603 DPUTS("write failed, disk error?");
2604 /* On a failure, the pagecache still contains the new data.
2605 * Write some old data back, to prevent it from being used.
2606 * Use the non-SYNC fd; we know it will fail anyway.
2608 meta.mm_last_pg = metab.mm_last_pg;
2609 meta.mm_txnid = metab.mm_txnid;
2611 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2613 r2 = pwrite(env->me_fd, ptr, len, off);
2616 env->me_flags |= MDB_FATAL_ERROR;
2620 /* Memory ordering issues are irrelevant; since the entire writer
2621 * is wrapped by wmutex, all of these changes will become visible
2622 * after the wmutex is unlocked. Since the DB is multi-version,
2623 * readers will get consistent data regardless of how fresh or
2624 * how stale their view of these values is.
2626 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2631 /** Check both meta pages to see which one is newer.
2632 * @param[in] env the environment handle
2633 * @return meta toggle (0 or 1).
2636 mdb_env_pick_meta(const MDB_env *env)
2638 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2642 mdb_env_create(MDB_env **env)
2646 e = calloc(1, sizeof(MDB_env));
2650 e->me_free_pgs = mdb_midl_alloc();
2651 if (!e->me_free_pgs) {
2655 e->me_maxreaders = DEFAULT_READERS;
2657 e->me_fd = INVALID_HANDLE_VALUE;
2658 e->me_lfd = INVALID_HANDLE_VALUE;
2659 e->me_mfd = INVALID_HANDLE_VALUE;
2660 #ifdef MDB_USE_POSIX_SEM
2661 e->me_rmutex = SEM_FAILED;
2662 e->me_wmutex = SEM_FAILED;
2664 e->me_pid = getpid();
2665 VGMEMP_CREATE(e,0,0);
2671 mdb_env_set_mapsize(MDB_env *env, size_t size)
2675 env->me_mapsize = size;
2677 env->me_maxpg = env->me_mapsize / env->me_psize;
2682 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2686 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2691 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2693 if (env->me_map || readers < 1)
2695 env->me_maxreaders = readers;
2700 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2702 if (!env || !readers)
2704 *readers = env->me_maxreaders;
2708 /** Further setup required for opening an MDB environment
2711 mdb_env_open2(MDB_env *env)
2713 unsigned int flags = env->me_flags;
2714 int i, newenv = 0, prot;
2718 memset(&meta, 0, sizeof(meta));
2720 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2723 DPUTS("new mdbenv");
2725 meta.mm_mapsize = env->me_mapsize > DEFAULT_MAPSIZE ? env->me_mapsize : DEFAULT_MAPSIZE;
2728 if (env->me_mapsize < meta.mm_mapsize)
2729 env->me_mapsize = meta.mm_mapsize;
2734 LONG sizelo, sizehi;
2735 sizelo = env->me_mapsize & 0xffffffff;
2736 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2738 /* Windows won't create mappings for zero length files.
2739 * Just allocate the maxsize right now.
2742 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2743 if (!SetEndOfFile(env->me_fd))
2745 SetFilePointer(env->me_fd, 0, NULL, 0);
2747 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2748 PAGE_READWRITE : PAGE_READONLY,
2749 sizehi, sizelo, NULL);
2752 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2753 FILE_MAP_WRITE : FILE_MAP_READ,
2754 0, 0, env->me_mapsize, meta.mm_address);
2762 if (flags & MDB_WRITEMAP) {
2764 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2767 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2769 if (env->me_map == MAP_FAILED) {
2776 if (flags & MDB_FIXEDMAP)
2777 meta.mm_address = env->me_map;
2778 i = mdb_env_init_meta(env, &meta);
2779 if (i != MDB_SUCCESS) {
2782 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2783 /* Can happen because the address argument to mmap() is just a
2784 * hint. mmap() can pick another, e.g. if the range is in use.
2785 * The MAP_FIXED flag would prevent that, but then mmap could
2786 * instead unmap existing pages to make room for the new map.
2788 return EBUSY; /* TODO: Make a new MDB_* error code? */
2790 env->me_psize = meta.mm_psize;
2792 env->me_maxpg = env->me_mapsize / env->me_psize;
2794 p = (MDB_page *)env->me_map;
2795 env->me_metas[0] = METADATA(p);
2796 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2800 int toggle = mdb_env_pick_meta(env);
2801 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2803 DPRINTF("opened database version %u, pagesize %u",
2804 env->me_metas[0]->mm_version, env->me_psize);
2805 DPRINTF("using meta page %d", toggle);
2806 DPRINTF("depth: %u", db->md_depth);
2807 DPRINTF("entries: %zu", db->md_entries);
2808 DPRINTF("branch pages: %zu", db->md_branch_pages);
2809 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2810 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2811 DPRINTF("root: %zu", db->md_root);
2819 /** Release a reader thread's slot in the reader lock table.
2820 * This function is called automatically when a thread exits.
2821 * @param[in] ptr This points to the slot in the reader lock table.
2824 mdb_env_reader_dest(void *ptr)
2826 MDB_reader *reader = ptr;
2832 /** Junk for arranging thread-specific callbacks on Windows. This is
2833 * necessarily platform and compiler-specific. Windows supports up
2834 * to 1088 keys. Let's assume nobody opens more than 64 environments
2835 * in a single process, for now. They can override this if needed.
2837 #ifndef MAX_TLS_KEYS
2838 #define MAX_TLS_KEYS 64
2840 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2841 static int mdb_tls_nkeys;
2843 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2847 case DLL_PROCESS_ATTACH: break;
2848 case DLL_THREAD_ATTACH: break;
2849 case DLL_THREAD_DETACH:
2850 for (i=0; i<mdb_tls_nkeys; i++) {
2851 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2852 mdb_env_reader_dest(r);
2855 case DLL_PROCESS_DETACH: break;
2860 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2862 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2866 /* Force some symbol references.
2867 * _tls_used forces the linker to create the TLS directory if not already done
2868 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2870 #pragma comment(linker, "/INCLUDE:_tls_used")
2871 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2872 #pragma const_seg(".CRT$XLB")
2873 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2874 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2877 #pragma comment(linker, "/INCLUDE:__tls_used")
2878 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2879 #pragma data_seg(".CRT$XLB")
2880 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2882 #endif /* WIN 32/64 */
2883 #endif /* !__GNUC__ */
2886 /** Downgrade the exclusive lock on the region back to shared */
2888 mdb_env_share_locks(MDB_env *env, int *excl)
2890 int rc = 0, toggle = mdb_env_pick_meta(env);
2892 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2897 /* First acquire a shared lock. The Unlock will
2898 * then release the existing exclusive lock.
2900 memset(&ov, 0, sizeof(ov));
2901 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2904 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2910 struct flock lock_info;
2911 /* The shared lock replaces the existing lock */
2912 memset((void *)&lock_info, 0, sizeof(lock_info));
2913 lock_info.l_type = F_RDLCK;
2914 lock_info.l_whence = SEEK_SET;
2915 lock_info.l_start = 0;
2916 lock_info.l_len = 1;
2917 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2918 (rc = ErrCode()) == EINTR) ;
2919 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
2926 /** Try to get exlusive lock, otherwise shared.
2927 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
2930 mdb_env_excl_lock(MDB_env *env, int *excl)
2934 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2938 memset(&ov, 0, sizeof(ov));
2939 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2946 struct flock lock_info;
2947 memset((void *)&lock_info, 0, sizeof(lock_info));
2948 lock_info.l_type = F_WRLCK;
2949 lock_info.l_whence = SEEK_SET;
2950 lock_info.l_start = 0;
2951 lock_info.l_len = 1;
2952 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2953 (rc = ErrCode()) == EINTR) ;
2957 # ifdef MDB_USE_POSIX_SEM
2958 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
2961 lock_info.l_type = F_RDLCK;
2962 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
2963 (rc = ErrCode()) == EINTR) ;
2971 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
2973 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2975 * @(#) $Revision: 5.1 $
2976 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2977 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2979 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2983 * Please do not copyright this code. This code is in the public domain.
2985 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2986 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2987 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2988 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2989 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2990 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2991 * PERFORMANCE OF THIS SOFTWARE.
2994 * chongo <Landon Curt Noll> /\oo/\
2995 * http://www.isthe.com/chongo/
2997 * Share and Enjoy! :-)
3000 typedef unsigned long long mdb_hash_t;
3001 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3003 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3004 * @param[in] str string to hash
3005 * @param[in] hval initial value for hash
3006 * @return 64 bit hash
3008 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3009 * hval arg on the first call.
3012 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3014 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3015 unsigned char *end = s + val->mv_size;
3017 * FNV-1a hash each octet of the string
3020 /* xor the bottom with the current octet */
3021 hval ^= (mdb_hash_t)*s++;
3023 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3024 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3025 (hval << 7) + (hval << 8) + (hval << 40);
3027 /* return our new hash value */
3031 /** Hash the string and output the hash in hex.
3032 * @param[in] str string to hash
3033 * @param[out] hexbuf an array of 17 chars to hold the hash
3036 mdb_hash_hex(MDB_val *val, char *hexbuf)
3039 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3040 for (i=0; i<8; i++) {
3041 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3047 /** Open and/or initialize the lock region for the environment.
3048 * @param[in] env The MDB environment.
3049 * @param[in] lpath The pathname of the file used for the lock region.
3050 * @param[in] mode The Unix permissions for the file, if we create it.
3051 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3052 * @return 0 on success, non-zero on failure.
3055 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3063 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3064 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3065 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
3067 if (rc == ERROR_WRITE_PROTECT && (env->me_flags & MDB_RDONLY)) {
3068 env->me_flags |= MDB_ROFS;
3073 /* Try to get exclusive lock. If we succeed, then
3074 * nobody is using the lock region and we should initialize it.
3076 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3077 size = GetFileSize(env->me_lfd, NULL);
3083 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
3085 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3086 env->me_flags |= MDB_ROFS;
3091 /* Lose record locks when exec*() */
3092 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3093 fcntl(env->me_lfd, F_SETFD, fdflags);
3095 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
3096 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1) {
3098 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3099 env->me_flags |= MDB_ROFS;
3106 /* Try to get exclusive lock. If we succeed, then
3107 * nobody is using the lock region and we should initialize it.
3109 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3111 size = lseek(env->me_lfd, 0, SEEK_END);
3113 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3114 if (size < rsize && *excl > 0) {
3116 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3117 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3119 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3123 size = rsize - sizeof(MDB_txninfo);
3124 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3129 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3131 if (!mh) goto fail_errno;
3132 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3134 if (!env->me_txns) goto fail_errno;
3136 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3138 if (m == MAP_FAILED) goto fail_errno;
3144 BY_HANDLE_FILE_INFORMATION stbuf;
3153 if (!mdb_sec_inited) {
3154 InitializeSecurityDescriptor(&mdb_null_sd,
3155 SECURITY_DESCRIPTOR_REVISION);
3156 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3157 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3158 mdb_all_sa.bInheritHandle = FALSE;
3159 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3162 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3163 idbuf.volume = stbuf.dwVolumeSerialNumber;
3164 idbuf.nhigh = stbuf.nFileIndexHigh;
3165 idbuf.nlow = stbuf.nFileIndexLow;
3166 val.mv_data = &idbuf;
3167 val.mv_size = sizeof(idbuf);
3168 mdb_hash_hex(&val, hexbuf);
3169 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3170 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3171 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3172 if (!env->me_rmutex) goto fail_errno;
3173 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3174 if (!env->me_wmutex) goto fail_errno;
3175 #elif defined(MDB_USE_POSIX_SEM)
3184 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3185 idbuf.dev = stbuf.st_dev;
3186 idbuf.ino = stbuf.st_ino;
3187 val.mv_data = &idbuf;
3188 val.mv_size = sizeof(idbuf);
3189 mdb_hash_hex(&val, hexbuf);
3190 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3191 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3192 /* Clean up after a previous run, if needed: Try to
3193 * remove both semaphores before doing anything else.
3195 sem_unlink(env->me_txns->mti_rmname);
3196 sem_unlink(env->me_txns->mti_wmname);
3197 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3198 O_CREAT|O_EXCL, mode, 1);
3199 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3200 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3201 O_CREAT|O_EXCL, mode, 1);
3202 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3203 #else /* MDB_USE_POSIX_SEM */
3204 pthread_mutexattr_t mattr;
3206 if ((rc = pthread_mutexattr_init(&mattr))
3207 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3208 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3209 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3211 pthread_mutexattr_destroy(&mattr);
3212 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3214 env->me_txns->mti_version = MDB_VERSION;
3215 env->me_txns->mti_magic = MDB_MAGIC;
3216 env->me_txns->mti_txnid = 0;
3217 env->me_txns->mti_numreaders = 0;
3220 if (env->me_txns->mti_magic != MDB_MAGIC) {
3221 DPUTS("lock region has invalid magic");
3225 if (env->me_txns->mti_version != MDB_VERSION) {
3226 DPRINTF("lock region is version %u, expected version %u",
3227 env->me_txns->mti_version, MDB_VERSION);
3228 rc = MDB_VERSION_MISMATCH;
3232 if (rc != EACCES && rc != EAGAIN) {
3236 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3237 if (!env->me_rmutex) goto fail_errno;
3238 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3239 if (!env->me_wmutex) goto fail_errno;
3240 #elif defined(MDB_USE_POSIX_SEM)
3241 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3242 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3243 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3244 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3255 /** The name of the lock file in the DB environment */
3256 #define LOCKNAME "/lock.mdb"
3257 /** The name of the data file in the DB environment */
3258 #define DATANAME "/data.mdb"
3259 /** The suffix of the lock file when no subdir is used */
3260 #define LOCKSUFF "-lock"
3261 /** Only a subset of the @ref mdb_env flags can be changed
3262 * at runtime. Changing other flags requires closing the
3263 * environment and re-opening it with the new flags.
3265 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3266 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP)
3269 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3271 int oflags, rc, len, excl;
3272 char *lpath, *dpath;
3274 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3278 if (flags & MDB_NOSUBDIR) {
3279 rc = len + sizeof(LOCKSUFF) + len + 1;
3281 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3286 if (flags & MDB_NOSUBDIR) {
3287 dpath = lpath + len + sizeof(LOCKSUFF);
3288 sprintf(lpath, "%s" LOCKSUFF, path);
3289 strcpy(dpath, path);
3291 dpath = lpath + len + sizeof(LOCKNAME);
3292 sprintf(lpath, "%s" LOCKNAME, path);
3293 sprintf(dpath, "%s" DATANAME, path);
3296 flags |= env->me_flags;
3297 /* silently ignore WRITEMAP if we're only getting read access */
3298 if (F_ISSET(flags, MDB_RDONLY|MDB_WRITEMAP))
3299 flags ^= MDB_WRITEMAP;
3300 env->me_flags = flags |= MDB_ENV_ACTIVE;
3302 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3307 if (F_ISSET(flags, MDB_RDONLY)) {
3308 oflags = GENERIC_READ;
3309 len = OPEN_EXISTING;
3311 oflags = GENERIC_READ|GENERIC_WRITE;
3314 mode = FILE_ATTRIBUTE_NORMAL;
3315 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3316 NULL, len, mode, NULL);
3318 if (F_ISSET(flags, MDB_RDONLY))
3321 oflags = O_RDWR | O_CREAT;
3323 env->me_fd = open(dpath, oflags, mode);
3325 if (env->me_fd == INVALID_HANDLE_VALUE) {
3330 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3331 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3332 env->me_mfd = env->me_fd;
3334 /* Synchronous fd for meta writes. Needed even with
3335 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3338 env->me_mfd = CreateFile(dpath, oflags,
3339 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3340 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3342 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3344 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3349 DPRINTF("opened dbenv %p", (void *) env);
3350 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3353 env->me_numdbs = 2; /* this notes that me_txkey was set */
3355 /* Windows TLS callbacks need help finding their TLS info. */
3356 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3357 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3364 rc = mdb_env_share_locks(env, &excl);
3368 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3369 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3370 env->me_path = strdup(path);
3371 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3377 mdb_env_close0(env, excl);
3383 /** Destroy resources from mdb_env_open() and clear our readers */
3385 mdb_env_close0(MDB_env *env, int excl)
3389 if (!(env->me_flags & MDB_ENV_ACTIVE))
3392 free(env->me_dbflags);
3396 if (env->me_numdbs) {
3397 pthread_key_delete(env->me_txkey);
3399 /* Delete our key from the global list */
3400 for (i=0; i<mdb_tls_nkeys; i++)
3401 if (mdb_tls_keys[i] == env->me_txkey) {
3402 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3410 munmap(env->me_map, env->me_mapsize);
3412 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3414 if (env->me_fd != INVALID_HANDLE_VALUE)
3417 pid_t pid = env->me_pid;
3418 /* Clearing readers is done in this function because
3419 * me_txkey with its destructor must be disabled first.
3421 for (i = env->me_numreaders; --i >= 0; )
3422 if (env->me_txns->mti_readers[i].mr_pid == pid)
3423 env->me_txns->mti_readers[i].mr_pid = 0;
3425 if (env->me_rmutex) {
3426 CloseHandle(env->me_rmutex);
3427 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3429 /* Windows automatically destroys the mutexes when
3430 * the last handle closes.
3432 #elif defined(MDB_USE_POSIX_SEM)
3433 if (env->me_rmutex != SEM_FAILED) {
3434 sem_close(env->me_rmutex);
3435 if (env->me_wmutex != SEM_FAILED)
3436 sem_close(env->me_wmutex);
3437 /* If we have the filelock: If we are the
3438 * only remaining user, clean up semaphores.
3441 mdb_env_excl_lock(env, &excl);
3443 sem_unlink(env->me_txns->mti_rmname);
3444 sem_unlink(env->me_txns->mti_wmname);
3448 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3450 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3453 /* Unlock the lockfile. Windows would have unlocked it
3454 * after closing anyway, but not necessarily at once.
3456 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3462 env->me_flags &= ~MDB_ENV_ACTIVE;
3466 mdb_env_copy(MDB_env *env, const char *path)
3468 MDB_txn *txn = NULL;
3472 HANDLE newfd = INVALID_HANDLE_VALUE;
3474 if (env->me_flags & MDB_NOSUBDIR) {
3475 lpath = (char *)path;
3478 len += sizeof(DATANAME);
3479 lpath = malloc(len);
3482 sprintf(lpath, "%s" DATANAME, path);
3485 /* The destination path must exist, but the destination file must not.
3486 * We don't want the OS to cache the writes, since the source data is
3487 * already in the OS cache.
3490 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3491 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3493 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3499 if (!(env->me_flags & MDB_NOSUBDIR))
3501 if (newfd == INVALID_HANDLE_VALUE) {
3506 #ifdef F_NOCACHE /* __APPLE__ */
3507 rc = fcntl(newfd, F_NOCACHE, 1);
3514 /* Do the lock/unlock of the reader mutex before starting the
3515 * write txn. Otherwise other read txns could block writers.
3517 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3521 if (!(env->me_flags & MDB_ROFS)) {
3522 /* We must start the actual read txn after blocking writers */
3523 mdb_txn_reset0(txn);
3525 /* Temporarily block writers until we snapshot the meta pages */
3528 rc = mdb_txn_renew0(txn);
3530 UNLOCK_MUTEX_W(env);
3535 wsize = env->me_psize * 2;
3539 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3540 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3543 rc = write(newfd, env->me_map, wsize);
3544 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3546 if (! (env->me_flags & MDB_ROFS))
3547 UNLOCK_MUTEX_W(env);
3552 ptr = env->me_map + wsize;
3553 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3554 #define MAX_WRITE 2147483648U
3558 if (wsize > MAX_WRITE)
3562 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3563 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3572 if (wsize > MAX_WRITE)
3576 wres = write(newfd, ptr, w2);
3577 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3586 if (newfd != INVALID_HANDLE_VALUE)
3593 mdb_env_close(MDB_env *env)
3600 VGMEMP_DESTROY(env);
3601 while ((dp = env->me_dpages) != NULL) {
3602 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3603 env->me_dpages = dp->mp_next;
3607 mdb_env_close0(env, 0);
3608 mdb_midl_free(env->me_free_pgs);
3612 /** Compare two items pointing at aligned size_t's */
3614 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3616 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3617 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3620 /** Compare two items pointing at aligned int's */
3622 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3624 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3625 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3628 /** Compare two items pointing at ints of unknown alignment.
3629 * Nodes and keys are guaranteed to be 2-byte aligned.
3632 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3634 #if BYTE_ORDER == LITTLE_ENDIAN
3635 unsigned short *u, *c;
3638 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3639 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3642 } while(!x && u > (unsigned short *)a->mv_data);
3645 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3649 /** Compare two items lexically */
3651 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3658 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3664 diff = memcmp(a->mv_data, b->mv_data, len);
3665 return diff ? diff : len_diff<0 ? -1 : len_diff;
3668 /** Compare two items in reverse byte order */
3670 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3672 const unsigned char *p1, *p2, *p1_lim;
3676 p1_lim = (const unsigned char *)a->mv_data;
3677 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3678 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3680 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3686 while (p1 > p1_lim) {
3687 diff = *--p1 - *--p2;
3691 return len_diff<0 ? -1 : len_diff;
3694 /** Search for key within a page, using binary search.
3695 * Returns the smallest entry larger or equal to the key.
3696 * If exactp is non-null, stores whether the found entry was an exact match
3697 * in *exactp (1 or 0).
3698 * Updates the cursor index with the index of the found entry.
3699 * If no entry larger or equal to the key is found, returns NULL.
3702 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3704 unsigned int i = 0, nkeys;
3707 MDB_page *mp = mc->mc_pg[mc->mc_top];
3708 MDB_node *node = NULL;
3713 nkeys = NUMKEYS(mp);
3718 COPY_PGNO(pgno, mp->mp_pgno);
3719 DPRINTF("searching %u keys in %s %spage %zu",
3720 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3727 low = IS_LEAF(mp) ? 0 : 1;
3729 cmp = mc->mc_dbx->md_cmp;
3731 /* Branch pages have no data, so if using integer keys,
3732 * alignment is guaranteed. Use faster mdb_cmp_int.
3734 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3735 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3742 nodekey.mv_size = mc->mc_db->md_pad;
3743 node = NODEPTR(mp, 0); /* fake */
3744 while (low <= high) {
3745 i = (low + high) >> 1;
3746 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3747 rc = cmp(key, &nodekey);
3748 DPRINTF("found leaf index %u [%s], rc = %i",
3749 i, DKEY(&nodekey), rc);
3758 while (low <= high) {
3759 i = (low + high) >> 1;
3761 node = NODEPTR(mp, i);
3762 nodekey.mv_size = NODEKSZ(node);
3763 nodekey.mv_data = NODEKEY(node);
3765 rc = cmp(key, &nodekey);
3768 DPRINTF("found leaf index %u [%s], rc = %i",
3769 i, DKEY(&nodekey), rc);
3771 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3772 i, DKEY(&nodekey), NODEPGNO(node), rc);
3783 if (rc > 0) { /* Found entry is less than the key. */
3784 i++; /* Skip to get the smallest entry larger than key. */
3786 node = NODEPTR(mp, i);
3789 *exactp = (rc == 0);
3790 /* store the key index */
3791 mc->mc_ki[mc->mc_top] = i;
3793 /* There is no entry larger or equal to the key. */
3796 /* nodeptr is fake for LEAF2 */
3802 mdb_cursor_adjust(MDB_cursor *mc, func)
3806 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3807 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3814 /** Pop a page off the top of the cursor's stack. */
3816 mdb_cursor_pop(MDB_cursor *mc)
3819 #ifndef MDB_DEBUG_SKIP
3820 MDB_page *top = mc->mc_pg[mc->mc_top];
3826 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3827 mc->mc_dbi, (void *) mc);
3831 /** Push a page onto the top of the cursor's stack. */
3833 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3835 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3836 mc->mc_dbi, (void *) mc);
3838 if (mc->mc_snum >= CURSOR_STACK) {
3839 assert(mc->mc_snum < CURSOR_STACK);
3840 return MDB_CURSOR_FULL;
3843 mc->mc_top = mc->mc_snum++;
3844 mc->mc_pg[mc->mc_top] = mp;
3845 mc->mc_ki[mc->mc_top] = 0;
3850 /** Find the address of the page corresponding to a given page number.
3851 * @param[in] txn the transaction for this access.
3852 * @param[in] pgno the page number for the page to retrieve.
3853 * @param[out] ret address of a pointer where the page's address will be stored.
3854 * @return 0 on success, non-zero on failure.
3857 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3861 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
3862 if (pgno < txn->mt_next_pgno)
3863 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3866 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3868 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3869 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3870 p = txn->mt_u.dirty_list[x].mptr;
3874 if (pgno < txn->mt_next_pgno)
3875 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3880 DPRINTF("page %zu not found", pgno);
3883 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3886 /** Search for the page a given key should be in.
3887 * Pushes parent pages on the cursor stack. This function continues a
3888 * search on a cursor that has already been initialized. (Usually by
3889 * #mdb_page_search() but also by #mdb_node_move().)
3890 * @param[in,out] mc the cursor for this operation.
3891 * @param[in] key the key to search for. If NULL, search for the lowest
3892 * page. (This is used by #mdb_cursor_first().)
3893 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3894 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3895 * @return 0 on success, non-zero on failure.
3898 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3900 MDB_page *mp = mc->mc_pg[mc->mc_top];
3905 while (IS_BRANCH(mp)) {
3909 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3910 assert(NUMKEYS(mp) > 1);
3911 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3913 if (key == NULL) /* Initialize cursor to first page. */
3915 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3916 /* cursor to last page */
3920 node = mdb_node_search(mc, key, &exact);
3922 i = NUMKEYS(mp) - 1;
3924 i = mc->mc_ki[mc->mc_top];
3933 DPRINTF("following index %u for key [%s]",
3935 assert(i < NUMKEYS(mp));
3936 node = NODEPTR(mp, i);
3938 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3941 mc->mc_ki[mc->mc_top] = i;
3942 if ((rc = mdb_cursor_push(mc, mp)))
3946 if ((rc = mdb_page_touch(mc)) != 0)
3948 mp = mc->mc_pg[mc->mc_top];
3953 DPRINTF("internal error, index points to a %02X page!?",
3955 return MDB_CORRUPTED;
3958 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3959 key ? DKEY(key) : NULL);
3964 /** Search for the page a given key should be in.
3965 * Pushes parent pages on the cursor stack. This function just sets up
3966 * the search; it finds the root page for \b mc's database and sets this
3967 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3968 * called to complete the search.
3969 * @param[in,out] mc the cursor for this operation.
3970 * @param[in] key the key to search for. If NULL, search for the lowest
3971 * page. (This is used by #mdb_cursor_first().)
3972 * @param[in] modify If true, visited pages are updated with new page numbers.
3973 * @return 0 on success, non-zero on failure.
3976 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
3981 /* Make sure the txn is still viable, then find the root from
3982 * the txn's db table.
3984 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3985 DPUTS("transaction has failed, must abort");
3988 /* Make sure we're using an up-to-date root */
3989 if (mc->mc_dbi > MAIN_DBI) {
3990 if ((*mc->mc_dbflag & DB_STALE) ||
3991 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
3993 unsigned char dbflag = 0;
3994 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3995 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
3998 if (*mc->mc_dbflag & DB_STALE) {
4001 MDB_node *leaf = mdb_node_search(&mc2,
4002 &mc->mc_dbx->md_name, &exact);
4004 return MDB_NOTFOUND;
4005 mdb_node_read(mc->mc_txn, leaf, &data);
4006 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4008 if (flags & MDB_PS_MODIFY)
4010 *mc->mc_dbflag = dbflag;
4013 root = mc->mc_db->md_root;
4015 if (root == P_INVALID) { /* Tree is empty. */
4016 DPUTS("tree is empty");
4017 return MDB_NOTFOUND;
4022 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4023 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
4029 DPRINTF("db %u root page %zu has flags 0x%X",
4030 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4032 if (flags & MDB_PS_MODIFY) {
4033 if ((rc = mdb_page_touch(mc)))
4037 if (flags & MDB_PS_ROOTONLY)
4040 return mdb_page_search_root(mc, key, flags);
4043 /** Return the data associated with a given node.
4044 * @param[in] txn The transaction for this operation.
4045 * @param[in] leaf The node being read.
4046 * @param[out] data Updated to point to the node's data.
4047 * @return 0 on success, non-zero on failure.
4050 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4052 MDB_page *omp; /* overflow page */
4056 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4057 data->mv_size = NODEDSZ(leaf);
4058 data->mv_data = NODEDATA(leaf);
4062 /* Read overflow data.
4064 data->mv_size = NODEDSZ(leaf);
4065 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4066 if ((rc = mdb_page_get(txn, pgno, &omp))) {
4067 DPRINTF("read overflow page %zu failed", pgno);
4070 data->mv_data = METADATA(omp);
4076 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4077 MDB_val *key, MDB_val *data)
4086 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4088 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4091 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4095 mdb_cursor_init(&mc, txn, dbi, &mx);
4096 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4099 /** Find a sibling for a page.
4100 * Replaces the page at the top of the cursor's stack with the
4101 * specified sibling, if one exists.
4102 * @param[in] mc The cursor for this operation.
4103 * @param[in] move_right Non-zero if the right sibling is requested,
4104 * otherwise the left sibling.
4105 * @return 0 on success, non-zero on failure.
4108 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4114 if (mc->mc_snum < 2) {
4115 return MDB_NOTFOUND; /* root has no siblings */
4119 DPRINTF("parent page is page %zu, index %u",
4120 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4122 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4123 : (mc->mc_ki[mc->mc_top] == 0)) {
4124 DPRINTF("no more keys left, moving to %s sibling",
4125 move_right ? "right" : "left");
4126 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4127 /* undo cursor_pop before returning */
4134 mc->mc_ki[mc->mc_top]++;
4136 mc->mc_ki[mc->mc_top]--;
4137 DPRINTF("just moving to %s index key %u",
4138 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4140 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4142 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4143 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4146 mdb_cursor_push(mc, mp);
4151 /** Move the cursor to the next data item. */
4153 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4159 if (mc->mc_flags & C_EOF) {
4160 return MDB_NOTFOUND;
4163 assert(mc->mc_flags & C_INITIALIZED);
4165 mp = mc->mc_pg[mc->mc_top];
4167 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4168 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4169 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4170 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4171 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4172 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4176 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4177 if (op == MDB_NEXT_DUP)
4178 return MDB_NOTFOUND;
4182 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4184 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4185 DPUTS("=====> move to next sibling page");
4186 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4187 mc->mc_flags |= C_EOF;
4188 mc->mc_flags &= ~C_INITIALIZED;
4189 return MDB_NOTFOUND;
4191 mp = mc->mc_pg[mc->mc_top];
4192 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4194 mc->mc_ki[mc->mc_top]++;
4196 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4197 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4200 key->mv_size = mc->mc_db->md_pad;
4201 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4205 assert(IS_LEAF(mp));
4206 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4208 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4209 mdb_xcursor_init1(mc, leaf);
4212 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4215 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4216 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4217 if (rc != MDB_SUCCESS)
4222 MDB_GET_KEY(leaf, key);
4226 /** Move the cursor to the previous data item. */
4228 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4234 assert(mc->mc_flags & C_INITIALIZED);
4236 mp = mc->mc_pg[mc->mc_top];
4238 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4239 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4240 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4241 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4242 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4243 if (op != MDB_PREV || rc == MDB_SUCCESS)
4246 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4247 if (op == MDB_PREV_DUP)
4248 return MDB_NOTFOUND;
4253 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4255 if (mc->mc_ki[mc->mc_top] == 0) {
4256 DPUTS("=====> move to prev sibling page");
4257 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4258 mc->mc_flags &= ~C_INITIALIZED;
4259 return MDB_NOTFOUND;
4261 mp = mc->mc_pg[mc->mc_top];
4262 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4263 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4265 mc->mc_ki[mc->mc_top]--;
4267 mc->mc_flags &= ~C_EOF;
4269 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4270 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4273 key->mv_size = mc->mc_db->md_pad;
4274 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4278 assert(IS_LEAF(mp));
4279 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4281 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4282 mdb_xcursor_init1(mc, leaf);
4285 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4288 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4289 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4290 if (rc != MDB_SUCCESS)
4295 MDB_GET_KEY(leaf, key);
4299 /** Set the cursor on a specific data item. */
4301 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4302 MDB_cursor_op op, int *exactp)
4306 MDB_node *leaf = NULL;
4311 assert(key->mv_size > 0);
4313 /* See if we're already on the right page */
4314 if (mc->mc_flags & C_INITIALIZED) {
4317 mp = mc->mc_pg[mc->mc_top];
4319 mc->mc_ki[mc->mc_top] = 0;
4320 return MDB_NOTFOUND;
4322 if (mp->mp_flags & P_LEAF2) {
4323 nodekey.mv_size = mc->mc_db->md_pad;
4324 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4326 leaf = NODEPTR(mp, 0);
4327 MDB_GET_KEY(leaf, &nodekey);
4329 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4331 /* Probably happens rarely, but first node on the page
4332 * was the one we wanted.
4334 mc->mc_ki[mc->mc_top] = 0;
4341 unsigned int nkeys = NUMKEYS(mp);
4343 if (mp->mp_flags & P_LEAF2) {
4344 nodekey.mv_data = LEAF2KEY(mp,
4345 nkeys-1, nodekey.mv_size);
4347 leaf = NODEPTR(mp, nkeys-1);
4348 MDB_GET_KEY(leaf, &nodekey);
4350 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4352 /* last node was the one we wanted */
4353 mc->mc_ki[mc->mc_top] = nkeys-1;
4359 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4360 /* This is definitely the right page, skip search_page */
4361 if (mp->mp_flags & P_LEAF2) {
4362 nodekey.mv_data = LEAF2KEY(mp,
4363 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4365 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4366 MDB_GET_KEY(leaf, &nodekey);
4368 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4370 /* current node was the one we wanted */
4380 /* If any parents have right-sibs, search.
4381 * Otherwise, there's nothing further.
4383 for (i=0; i<mc->mc_top; i++)
4385 NUMKEYS(mc->mc_pg[i])-1)
4387 if (i == mc->mc_top) {
4388 /* There are no other pages */
4389 mc->mc_ki[mc->mc_top] = nkeys;
4390 return MDB_NOTFOUND;
4394 /* There are no other pages */
4395 mc->mc_ki[mc->mc_top] = 0;
4396 return MDB_NOTFOUND;
4400 rc = mdb_page_search(mc, key, 0);
4401 if (rc != MDB_SUCCESS)
4404 mp = mc->mc_pg[mc->mc_top];
4405 assert(IS_LEAF(mp));
4408 leaf = mdb_node_search(mc, key, exactp);
4409 if (exactp != NULL && !*exactp) {
4410 /* MDB_SET specified and not an exact match. */
4411 return MDB_NOTFOUND;
4415 DPUTS("===> inexact leaf not found, goto sibling");
4416 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4417 return rc; /* no entries matched */
4418 mp = mc->mc_pg[mc->mc_top];
4419 assert(IS_LEAF(mp));
4420 leaf = NODEPTR(mp, 0);
4424 mc->mc_flags |= C_INITIALIZED;
4425 mc->mc_flags &= ~C_EOF;
4428 key->mv_size = mc->mc_db->md_pad;
4429 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4433 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4434 mdb_xcursor_init1(mc, leaf);
4437 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4438 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4439 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4442 if (op == MDB_GET_BOTH) {
4448 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4449 if (rc != MDB_SUCCESS)
4452 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4454 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4456 rc = mc->mc_dbx->md_dcmp(data, &d2);
4458 if (op == MDB_GET_BOTH || rc > 0)
4459 return MDB_NOTFOUND;
4464 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4465 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4470 /* The key already matches in all other cases */
4471 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4472 MDB_GET_KEY(leaf, key);
4473 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4478 /** Move the cursor to the first item in the database. */
4480 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4485 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4486 rc = mdb_page_search(mc, NULL, 0);
4487 if (rc != MDB_SUCCESS)
4490 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4492 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4493 mc->mc_flags |= C_INITIALIZED;
4494 mc->mc_flags &= ~C_EOF;
4496 mc->mc_ki[mc->mc_top] = 0;
4498 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4499 key->mv_size = mc->mc_db->md_pad;
4500 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4505 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4506 mdb_xcursor_init1(mc, leaf);
4507 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4512 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4513 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4517 MDB_GET_KEY(leaf, key);
4521 /** Move the cursor to the last item in the database. */
4523 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4528 if (!(mc->mc_flags & C_EOF)) {
4530 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4533 lkey.mv_size = MAXKEYSIZE+1;
4534 lkey.mv_data = NULL;
4535 rc = mdb_page_search(mc, &lkey, 0);
4536 if (rc != MDB_SUCCESS)
4539 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4541 mc->mc_flags |= C_INITIALIZED|C_EOF;
4542 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4544 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4546 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4547 key->mv_size = mc->mc_db->md_pad;
4548 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4553 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4554 mdb_xcursor_init1(mc, leaf);
4555 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4560 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4561 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4566 MDB_GET_KEY(leaf, key);
4571 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4580 case MDB_GET_CURRENT:
4581 if (!mc->mc_flags & C_INITIALIZED) {
4584 MDB_page *mp = mc->mc_pg[mc->mc_top];
4586 mc->mc_ki[mc->mc_top] = 0;
4592 key->mv_size = mc->mc_db->md_pad;
4593 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4595 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4596 MDB_GET_KEY(leaf, key);
4598 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4599 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4601 rc = mdb_node_read(mc->mc_txn, leaf, data);
4608 case MDB_GET_BOTH_RANGE:
4609 if (data == NULL || mc->mc_xcursor == NULL) {
4617 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4619 } else if (op == MDB_SET_RANGE)
4620 rc = mdb_cursor_set(mc, key, data, op, NULL);
4622 rc = mdb_cursor_set(mc, key, data, op, &exact);
4624 case MDB_GET_MULTIPLE:
4626 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4627 !(mc->mc_flags & C_INITIALIZED)) {
4632 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4633 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4636 case MDB_NEXT_MULTIPLE:
4638 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4642 if (!(mc->mc_flags & C_INITIALIZED))
4643 rc = mdb_cursor_first(mc, key, data);
4645 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4646 if (rc == MDB_SUCCESS) {
4647 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4650 mx = &mc->mc_xcursor->mx_cursor;
4651 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4653 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4654 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4662 case MDB_NEXT_NODUP:
4663 if (!(mc->mc_flags & C_INITIALIZED))
4664 rc = mdb_cursor_first(mc, key, data);
4666 rc = mdb_cursor_next(mc, key, data, op);
4670 case MDB_PREV_NODUP:
4671 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4672 rc = mdb_cursor_last(mc, key, data);
4673 mc->mc_flags |= C_INITIALIZED;
4674 mc->mc_ki[mc->mc_top]++;
4676 rc = mdb_cursor_prev(mc, key, data, op);
4679 rc = mdb_cursor_first(mc, key, data);
4683 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4684 !(mc->mc_flags & C_INITIALIZED) ||
4685 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4689 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4692 rc = mdb_cursor_last(mc, key, data);
4696 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4697 !(mc->mc_flags & C_INITIALIZED) ||
4698 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4702 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4705 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4713 /** Touch all the pages in the cursor stack.
4714 * Makes sure all the pages are writable, before attempting a write operation.
4715 * @param[in] mc The cursor to operate on.
4718 mdb_cursor_touch(MDB_cursor *mc)
4722 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4725 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4726 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4727 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4730 *mc->mc_dbflag = DB_DIRTY;
4732 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4733 rc = mdb_page_touch(mc);
4737 mc->mc_top = mc->mc_snum-1;
4742 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4745 MDB_node *leaf = NULL;
4746 MDB_val xdata, *rdata, dkey;
4749 int do_sub = 0, insert = 0;
4750 unsigned int mcount = 0;
4754 char dbuf[MAXKEYSIZE+1];
4755 unsigned int nflags;
4758 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4761 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4762 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4766 if (flags == MDB_CURRENT) {
4767 if (!(mc->mc_flags & C_INITIALIZED))
4770 } else if (mc->mc_db->md_root == P_INVALID) {
4772 /* new database, write a root leaf page */
4773 DPUTS("allocating new root leaf page");
4774 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4778 mdb_cursor_push(mc, np);
4779 mc->mc_db->md_root = np->mp_pgno;
4780 mc->mc_db->md_depth++;
4781 *mc->mc_dbflag = DB_DIRTY;
4782 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4784 np->mp_flags |= P_LEAF2;
4785 mc->mc_flags |= C_INITIALIZED;
4791 if (flags & MDB_APPEND) {
4793 rc = mdb_cursor_last(mc, &k2, &d2);
4795 rc = mc->mc_dbx->md_cmp(key, &k2);
4798 mc->mc_ki[mc->mc_top]++;
4804 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4806 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4807 DPRINTF("duplicate key [%s]", DKEY(key));
4809 return MDB_KEYEXIST;
4811 if (rc && rc != MDB_NOTFOUND)
4815 /* Cursor is positioned, now make sure all pages are writable */
4816 rc2 = mdb_cursor_touch(mc);
4821 /* The key already exists */
4822 if (rc == MDB_SUCCESS) {
4823 /* there's only a key anyway, so this is a no-op */
4824 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4825 unsigned int ksize = mc->mc_db->md_pad;
4826 if (key->mv_size != ksize)
4828 if (flags == MDB_CURRENT) {
4829 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4830 memcpy(ptr, key->mv_data, ksize);
4835 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4838 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4839 /* Was a single item before, must convert now */
4841 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4842 /* Just overwrite the current item */
4843 if (flags == MDB_CURRENT)
4846 dkey.mv_size = NODEDSZ(leaf);
4847 dkey.mv_data = NODEDATA(leaf);
4848 #if UINT_MAX < SIZE_MAX
4849 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4850 #ifdef MISALIGNED_OK
4851 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4853 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4856 /* if data matches, ignore it */
4857 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4858 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4860 /* create a fake page for the dup items */
4861 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4862 dkey.mv_data = dbuf;
4863 fp = (MDB_page *)&pbuf;
4864 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4865 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4866 fp->mp_lower = PAGEHDRSZ;
4867 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4868 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4869 fp->mp_flags |= P_LEAF2;
4870 fp->mp_pad = data->mv_size;
4871 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4873 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4874 (dkey.mv_size & 1) + (data->mv_size & 1);
4876 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4879 xdata.mv_size = fp->mp_upper;
4884 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4885 /* See if we need to convert from fake page to subDB */
4887 unsigned int offset;
4890 fp = NODEDATA(leaf);
4891 if (flags == MDB_CURRENT) {
4893 fp->mp_flags |= P_DIRTY;
4894 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4895 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4899 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4900 offset = fp->mp_pad;
4901 if (SIZELEFT(fp) >= offset)
4903 offset *= 4; /* space for 4 more */
4905 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4907 offset += offset & 1;
4908 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4909 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4911 /* yes, convert it */
4913 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4914 dummy.md_pad = fp->mp_pad;
4915 dummy.md_flags = MDB_DUPFIXED;
4916 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4917 dummy.md_flags |= MDB_INTEGERKEY;
4920 dummy.md_branch_pages = 0;
4921 dummy.md_leaf_pages = 1;
4922 dummy.md_overflow_pages = 0;
4923 dummy.md_entries = NUMKEYS(fp);
4925 xdata.mv_size = sizeof(MDB_db);
4926 xdata.mv_data = &dummy;
4927 if ((rc = mdb_page_alloc(mc, 1, &mp)))
4929 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4930 flags |= F_DUPDATA|F_SUBDATA;
4931 dummy.md_root = mp->mp_pgno;
4933 /* no, just grow it */
4935 xdata.mv_size = NODEDSZ(leaf) + offset;
4936 xdata.mv_data = &pbuf;
4937 mp = (MDB_page *)&pbuf;
4938 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4941 mp->mp_flags = fp->mp_flags | P_DIRTY;
4942 mp->mp_pad = fp->mp_pad;
4943 mp->mp_lower = fp->mp_lower;
4944 mp->mp_upper = fp->mp_upper + offset;
4946 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4948 nsize = NODEDSZ(leaf) - fp->mp_upper;
4949 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4950 for (i=0; i<NUMKEYS(fp); i++)
4951 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4953 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4957 /* data is on sub-DB, just store it */
4958 flags |= F_DUPDATA|F_SUBDATA;
4962 /* overflow page overwrites need special handling */
4963 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4966 int ovpages, dpages;
4968 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4969 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4970 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4971 mdb_page_get(mc->mc_txn, pg, &omp);
4972 /* Is the ov page writable and large enough? */
4973 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4974 /* yes, overwrite it. Note in this case we don't
4975 * bother to try shrinking the node if the new data
4976 * is smaller than the overflow threshold.
4978 if (F_ISSET(flags, MDB_RESERVE))
4979 data->mv_data = METADATA(omp);
4981 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4984 /* no, free ovpages */
4986 mc->mc_db->md_overflow_pages -= ovpages;
4987 for (i=0; i<ovpages; i++) {
4988 DPRINTF("freed ov page %zu", pg);
4989 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4993 } else if (NODEDSZ(leaf) == data->mv_size) {
4994 /* same size, just replace it. Note that we could
4995 * also reuse this node if the new data is smaller,
4996 * but instead we opt to shrink the node in that case.
4998 if (F_ISSET(flags, MDB_RESERVE))
4999 data->mv_data = NODEDATA(leaf);
5001 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5004 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5005 mc->mc_db->md_entries--;
5007 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5014 nflags = flags & NODE_ADD_FLAGS;
5015 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5016 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5017 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5018 nflags &= ~MDB_APPEND;
5020 nflags |= MDB_SPLIT_REPLACE;
5021 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5023 /* There is room already in this leaf page. */
5024 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5025 if (rc == 0 && !do_sub && insert) {
5026 /* Adjust other cursors pointing to mp */
5027 MDB_cursor *m2, *m3;
5028 MDB_dbi dbi = mc->mc_dbi;
5029 unsigned i = mc->mc_top;
5030 MDB_page *mp = mc->mc_pg[i];
5032 if (mc->mc_flags & C_SUB)
5035 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5036 if (mc->mc_flags & C_SUB)
5037 m3 = &m2->mc_xcursor->mx_cursor;
5040 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5041 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5048 if (rc != MDB_SUCCESS)
5049 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5051 /* Now store the actual data in the child DB. Note that we're
5052 * storing the user data in the keys field, so there are strict
5053 * size limits on dupdata. The actual data fields of the child
5054 * DB are all zero size.
5061 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5062 if (flags & MDB_CURRENT) {
5063 xflags = MDB_CURRENT;
5065 mdb_xcursor_init1(mc, leaf);
5066 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5068 /* converted, write the original data first */
5070 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5074 /* Adjust other cursors pointing to mp */
5076 unsigned i = mc->mc_top;
5077 MDB_page *mp = mc->mc_pg[i];
5079 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5080 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5081 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5082 mdb_xcursor_init1(m2, leaf);
5087 if (flags & MDB_APPENDDUP)
5088 xflags |= MDB_APPEND;
5089 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5090 if (flags & F_SUBDATA) {
5091 void *db = NODEDATA(leaf);
5092 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5095 /* sub-writes might have failed so check rc again.
5096 * Don't increment count if we just replaced an existing item.
5098 if (!rc && !(flags & MDB_CURRENT))
5099 mc->mc_db->md_entries++;
5100 if (flags & MDB_MULTIPLE) {
5102 if (mcount < data[1].mv_size) {
5103 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5104 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5114 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5119 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5122 if (!mc->mc_flags & C_INITIALIZED)
5125 rc = mdb_cursor_touch(mc);
5129 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5131 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5132 if (flags != MDB_NODUPDATA) {
5133 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5134 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5136 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5137 /* If sub-DB still has entries, we're done */
5138 if (mc->mc_xcursor->mx_db.md_entries) {
5139 if (leaf->mn_flags & F_SUBDATA) {
5140 /* update subDB info */
5141 void *db = NODEDATA(leaf);
5142 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5144 /* shrink fake page */
5145 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5147 mc->mc_db->md_entries--;
5150 /* otherwise fall thru and delete the sub-DB */
5153 if (leaf->mn_flags & F_SUBDATA) {
5154 /* add all the child DB's pages to the free list */
5155 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5156 if (rc == MDB_SUCCESS) {
5157 mc->mc_db->md_entries -=
5158 mc->mc_xcursor->mx_db.md_entries;
5163 return mdb_cursor_del0(mc, leaf);
5166 /** Allocate and initialize new pages for a database.
5167 * @param[in] mc a cursor on the database being added to.
5168 * @param[in] flags flags defining what type of page is being allocated.
5169 * @param[in] num the number of pages to allocate. This is usually 1,
5170 * unless allocating overflow pages for a large record.
5171 * @param[out] mp Address of a page, or NULL on failure.
5172 * @return 0 on success, non-zero on failure.
5175 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5180 if ((rc = mdb_page_alloc(mc, num, &np)))
5182 DPRINTF("allocated new mpage %zu, page size %u",
5183 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5184 np->mp_flags = flags | P_DIRTY;
5185 np->mp_lower = PAGEHDRSZ;
5186 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5189 mc->mc_db->md_branch_pages++;
5190 else if (IS_LEAF(np))
5191 mc->mc_db->md_leaf_pages++;
5192 else if (IS_OVERFLOW(np)) {
5193 mc->mc_db->md_overflow_pages += num;
5201 /** Calculate the size of a leaf node.
5202 * The size depends on the environment's page size; if a data item
5203 * is too large it will be put onto an overflow page and the node
5204 * size will only include the key and not the data. Sizes are always
5205 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5206 * of the #MDB_node headers.
5207 * @param[in] env The environment handle.
5208 * @param[in] key The key for the node.
5209 * @param[in] data The data for the node.
5210 * @return The number of bytes needed to store the node.
5213 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5217 sz = LEAFSIZE(key, data);
5218 if (sz >= env->me_psize / MDB_MINKEYS) {
5219 /* put on overflow page */
5220 sz -= data->mv_size - sizeof(pgno_t);
5224 return sz + sizeof(indx_t);
5227 /** Calculate the size of a branch node.
5228 * The size should depend on the environment's page size but since
5229 * we currently don't support spilling large keys onto overflow
5230 * pages, it's simply the size of the #MDB_node header plus the
5231 * size of the key. Sizes are always rounded up to an even number
5232 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5233 * @param[in] env The environment handle.
5234 * @param[in] key The key for the node.
5235 * @return The number of bytes needed to store the node.
5238 mdb_branch_size(MDB_env *env, MDB_val *key)
5243 if (sz >= env->me_psize / MDB_MINKEYS) {
5244 /* put on overflow page */
5245 /* not implemented */
5246 /* sz -= key->size - sizeof(pgno_t); */
5249 return sz + sizeof(indx_t);
5252 /** Add a node to the page pointed to by the cursor.
5253 * @param[in] mc The cursor for this operation.
5254 * @param[in] indx The index on the page where the new node should be added.
5255 * @param[in] key The key for the new node.
5256 * @param[in] data The data for the new node, if any.
5257 * @param[in] pgno The page number, if adding a branch node.
5258 * @param[in] flags Flags for the node.
5259 * @return 0 on success, non-zero on failure. Possible errors are:
5261 * <li>ENOMEM - failed to allocate overflow pages for the node.
5262 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5263 * should never happen since all callers already calculate the
5264 * page's free space before calling this function.
5268 mdb_node_add(MDB_cursor *mc, indx_t indx,
5269 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5272 size_t node_size = NODESIZE;
5275 MDB_page *mp = mc->mc_pg[mc->mc_top];
5276 MDB_page *ofp = NULL; /* overflow page */
5279 assert(mp->mp_upper >= mp->mp_lower);
5281 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5282 IS_LEAF(mp) ? "leaf" : "branch",
5283 IS_SUBP(mp) ? "sub-" : "",
5284 mp->mp_pgno, indx, data ? data->mv_size : 0,
5285 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5288 /* Move higher keys up one slot. */
5289 int ksize = mc->mc_db->md_pad, dif;
5290 char *ptr = LEAF2KEY(mp, indx, ksize);
5291 dif = NUMKEYS(mp) - indx;
5293 memmove(ptr+ksize, ptr, dif*ksize);
5294 /* insert new key */
5295 memcpy(ptr, key->mv_data, ksize);
5297 /* Just using these for counting */
5298 mp->mp_lower += sizeof(indx_t);
5299 mp->mp_upper -= ksize - sizeof(indx_t);
5304 node_size += key->mv_size;
5308 if (F_ISSET(flags, F_BIGDATA)) {
5309 /* Data already on overflow page. */
5310 node_size += sizeof(pgno_t);
5311 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5312 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5314 /* Put data on overflow page. */
5315 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5316 data->mv_size, node_size+data->mv_size);
5317 node_size += sizeof(pgno_t);
5318 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5320 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5323 node_size += data->mv_size;
5326 node_size += node_size & 1;
5328 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5329 DPRINTF("not enough room in page %zu, got %u ptrs",
5330 mp->mp_pgno, NUMKEYS(mp));
5331 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5332 mp->mp_upper - mp->mp_lower);
5333 DPRINTF("node size = %zu", node_size);
5334 return MDB_PAGE_FULL;
5337 /* Move higher pointers up one slot. */
5338 for (i = NUMKEYS(mp); i > indx; i--)
5339 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5341 /* Adjust free space offsets. */
5342 ofs = mp->mp_upper - node_size;
5343 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5344 mp->mp_ptrs[indx] = ofs;
5346 mp->mp_lower += sizeof(indx_t);
5348 /* Write the node data. */
5349 node = NODEPTR(mp, indx);
5350 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5351 node->mn_flags = flags;
5353 SETDSZ(node,data->mv_size);
5358 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5363 if (F_ISSET(flags, F_BIGDATA))
5364 memcpy(node->mn_data + key->mv_size, data->mv_data,
5366 else if (F_ISSET(flags, MDB_RESERVE))
5367 data->mv_data = node->mn_data + key->mv_size;
5369 memcpy(node->mn_data + key->mv_size, data->mv_data,
5372 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5374 if (F_ISSET(flags, MDB_RESERVE))
5375 data->mv_data = METADATA(ofp);
5377 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5384 /** Delete the specified node from a page.
5385 * @param[in] mp The page to operate on.
5386 * @param[in] indx The index of the node to delete.
5387 * @param[in] ksize The size of a node. Only used if the page is
5388 * part of a #MDB_DUPFIXED database.
5391 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5394 indx_t i, j, numkeys, ptr;
5401 COPY_PGNO(pgno, mp->mp_pgno);
5402 DPRINTF("delete node %u on %s page %zu", indx,
5403 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5406 assert(indx < NUMKEYS(mp));
5409 int x = NUMKEYS(mp) - 1 - indx;
5410 base = LEAF2KEY(mp, indx, ksize);
5412 memmove(base, base + ksize, x * ksize);
5413 mp->mp_lower -= sizeof(indx_t);
5414 mp->mp_upper += ksize - sizeof(indx_t);
5418 node = NODEPTR(mp, indx);
5419 sz = NODESIZE + node->mn_ksize;
5421 if (F_ISSET(node->mn_flags, F_BIGDATA))
5422 sz += sizeof(pgno_t);
5424 sz += NODEDSZ(node);
5428 ptr = mp->mp_ptrs[indx];
5429 numkeys = NUMKEYS(mp);
5430 for (i = j = 0; i < numkeys; i++) {
5432 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5433 if (mp->mp_ptrs[i] < ptr)
5434 mp->mp_ptrs[j] += sz;
5439 base = (char *)mp + mp->mp_upper;
5440 memmove(base + sz, base, ptr - mp->mp_upper);
5442 mp->mp_lower -= sizeof(indx_t);
5446 /** Compact the main page after deleting a node on a subpage.
5447 * @param[in] mp The main page to operate on.
5448 * @param[in] indx The index of the subpage on the main page.
5451 mdb_node_shrink(MDB_page *mp, indx_t indx)
5458 indx_t i, numkeys, ptr;
5460 node = NODEPTR(mp, indx);
5461 sp = (MDB_page *)NODEDATA(node);
5462 osize = NODEDSZ(node);
5464 delta = sp->mp_upper - sp->mp_lower;
5465 SETDSZ(node, osize - delta);
5466 xp = (MDB_page *)((char *)sp + delta);
5468 /* shift subpage upward */
5470 nsize = NUMKEYS(sp) * sp->mp_pad;
5471 memmove(METADATA(xp), METADATA(sp), nsize);
5474 nsize = osize - sp->mp_upper;
5475 numkeys = NUMKEYS(sp);
5476 for (i=numkeys-1; i>=0; i--)
5477 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5479 xp->mp_upper = sp->mp_lower;
5480 xp->mp_lower = sp->mp_lower;
5481 xp->mp_flags = sp->mp_flags;
5482 xp->mp_pad = sp->mp_pad;
5483 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5485 /* shift lower nodes upward */
5486 ptr = mp->mp_ptrs[indx];
5487 numkeys = NUMKEYS(mp);
5488 for (i = 0; i < numkeys; i++) {
5489 if (mp->mp_ptrs[i] <= ptr)
5490 mp->mp_ptrs[i] += delta;
5493 base = (char *)mp + mp->mp_upper;
5494 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5495 mp->mp_upper += delta;
5498 /** Initial setup of a sorted-dups cursor.
5499 * Sorted duplicates are implemented as a sub-database for the given key.
5500 * The duplicate data items are actually keys of the sub-database.
5501 * Operations on the duplicate data items are performed using a sub-cursor
5502 * initialized when the sub-database is first accessed. This function does
5503 * the preliminary setup of the sub-cursor, filling in the fields that
5504 * depend only on the parent DB.
5505 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5508 mdb_xcursor_init0(MDB_cursor *mc)
5510 MDB_xcursor *mx = mc->mc_xcursor;
5512 mx->mx_cursor.mc_xcursor = NULL;
5513 mx->mx_cursor.mc_txn = mc->mc_txn;
5514 mx->mx_cursor.mc_db = &mx->mx_db;
5515 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5516 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5517 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5518 mx->mx_cursor.mc_snum = 0;
5519 mx->mx_cursor.mc_top = 0;
5520 mx->mx_cursor.mc_flags = C_SUB;
5521 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5522 mx->mx_dbx.md_dcmp = NULL;
5523 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5526 /** Final setup of a sorted-dups cursor.
5527 * Sets up the fields that depend on the data from the main cursor.
5528 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5529 * @param[in] node The data containing the #MDB_db record for the
5530 * sorted-dup database.
5533 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5535 MDB_xcursor *mx = mc->mc_xcursor;
5537 if (node->mn_flags & F_SUBDATA) {
5538 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5539 mx->mx_cursor.mc_pg[0] = 0;
5540 mx->mx_cursor.mc_snum = 0;
5541 mx->mx_cursor.mc_flags = C_SUB;
5543 MDB_page *fp = NODEDATA(node);
5544 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5545 mx->mx_db.md_flags = 0;
5546 mx->mx_db.md_depth = 1;
5547 mx->mx_db.md_branch_pages = 0;
5548 mx->mx_db.md_leaf_pages = 1;
5549 mx->mx_db.md_overflow_pages = 0;
5550 mx->mx_db.md_entries = NUMKEYS(fp);
5551 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5552 mx->mx_cursor.mc_snum = 1;
5553 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5554 mx->mx_cursor.mc_top = 0;
5555 mx->mx_cursor.mc_pg[0] = fp;
5556 mx->mx_cursor.mc_ki[0] = 0;
5557 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5558 mx->mx_db.md_flags = MDB_DUPFIXED;
5559 mx->mx_db.md_pad = fp->mp_pad;
5560 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5561 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5564 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5566 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5568 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5569 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5570 #if UINT_MAX < SIZE_MAX
5571 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5572 #ifdef MISALIGNED_OK
5573 mx->mx_dbx.md_cmp = mdb_cmp_long;
5575 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5580 /** Initialize a cursor for a given transaction and database. */
5582 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5587 mc->mc_db = &txn->mt_dbs[dbi];
5588 mc->mc_dbx = &txn->mt_dbxs[dbi];
5589 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5594 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5596 mc->mc_xcursor = mx;
5597 mdb_xcursor_init0(mc);
5599 mc->mc_xcursor = NULL;
5601 if (*mc->mc_dbflag & DB_STALE) {
5602 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5607 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5610 MDB_xcursor *mx = NULL;
5611 size_t size = sizeof(MDB_cursor);
5613 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5616 /* Allow read access to the freelist */
5617 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5620 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5621 size += sizeof(MDB_xcursor);
5623 if ((mc = malloc(size)) != NULL) {
5624 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5625 mx = (MDB_xcursor *)(mc + 1);
5627 mdb_cursor_init(mc, txn, dbi, mx);
5628 if (txn->mt_cursors) {
5629 mc->mc_next = txn->mt_cursors[dbi];
5630 txn->mt_cursors[dbi] = mc;
5632 mc->mc_flags |= C_ALLOCD;
5643 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5645 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5648 if (txn->mt_cursors)
5651 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5655 /* Return the count of duplicate data items for the current key */
5657 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5661 if (mc == NULL || countp == NULL)
5664 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5667 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5668 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5671 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5674 *countp = mc->mc_xcursor->mx_db.md_entries;
5680 mdb_cursor_close(MDB_cursor *mc)
5683 /* remove from txn, if tracked */
5684 if (mc->mc_txn->mt_cursors) {
5685 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5686 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5688 *prev = mc->mc_next;
5690 if (mc->mc_flags & C_ALLOCD)
5696 mdb_cursor_txn(MDB_cursor *mc)
5698 if (!mc) return NULL;
5703 mdb_cursor_dbi(MDB_cursor *mc)
5709 /** Replace the key for a node with a new key.
5710 * @param[in] mp The page containing the node to operate on.
5711 * @param[in] indx The index of the node to operate on.
5712 * @param[in] key The new key to use.
5713 * @return 0 on success, non-zero on failure.
5716 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5722 indx_t ptr, i, numkeys;
5725 node = NODEPTR(mp, indx);
5726 ptr = mp->mp_ptrs[indx];
5730 char kbuf2[(MAXKEYSIZE*2+1)];
5731 k2.mv_data = NODEKEY(node);
5732 k2.mv_size = node->mn_ksize;
5733 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5735 mdb_dkey(&k2, kbuf2),
5741 delta0 = delta = key->mv_size - node->mn_ksize;
5743 /* Must be 2-byte aligned. If new key is
5744 * shorter by 1, the shift will be skipped.
5746 delta += (delta & 1);
5748 if (delta > 0 && SIZELEFT(mp) < delta) {
5749 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5750 return MDB_PAGE_FULL;
5753 numkeys = NUMKEYS(mp);
5754 for (i = 0; i < numkeys; i++) {
5755 if (mp->mp_ptrs[i] <= ptr)
5756 mp->mp_ptrs[i] -= delta;
5759 base = (char *)mp + mp->mp_upper;
5760 len = ptr - mp->mp_upper + NODESIZE;
5761 memmove(base - delta, base, len);
5762 mp->mp_upper -= delta;
5764 node = NODEPTR(mp, indx);
5767 /* But even if no shift was needed, update ksize */
5769 node->mn_ksize = key->mv_size;
5772 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5777 /** Move a node from csrc to cdst.
5780 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5786 unsigned short flags;
5790 /* Mark src and dst as dirty. */
5791 if ((rc = mdb_page_touch(csrc)) ||
5792 (rc = mdb_page_touch(cdst)))
5795 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5796 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5797 key.mv_size = csrc->mc_db->md_pad;
5798 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5800 data.mv_data = NULL;
5804 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5805 assert(!((long)srcnode&1));
5806 srcpg = NODEPGNO(srcnode);
5807 flags = srcnode->mn_flags;
5808 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5809 unsigned int snum = csrc->mc_snum;
5811 /* must find the lowest key below src */
5812 mdb_page_search_root(csrc, NULL, 0);
5813 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5814 key.mv_size = csrc->mc_db->md_pad;
5815 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5817 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5818 key.mv_size = NODEKSZ(s2);
5819 key.mv_data = NODEKEY(s2);
5821 csrc->mc_snum = snum--;
5822 csrc->mc_top = snum;
5824 key.mv_size = NODEKSZ(srcnode);
5825 key.mv_data = NODEKEY(srcnode);
5827 data.mv_size = NODEDSZ(srcnode);
5828 data.mv_data = NODEDATA(srcnode);
5830 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5831 unsigned int snum = cdst->mc_snum;
5834 /* must find the lowest key below dst */
5835 mdb_page_search_root(cdst, NULL, 0);
5836 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5837 bkey.mv_size = cdst->mc_db->md_pad;
5838 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5840 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5841 bkey.mv_size = NODEKSZ(s2);
5842 bkey.mv_data = NODEKEY(s2);
5844 cdst->mc_snum = snum--;
5845 cdst->mc_top = snum;
5846 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5849 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5850 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5851 csrc->mc_ki[csrc->mc_top],
5853 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5854 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5856 /* Add the node to the destination page.
5858 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5859 if (rc != MDB_SUCCESS)
5862 /* Delete the node from the source page.
5864 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5867 /* Adjust other cursors pointing to mp */
5868 MDB_cursor *m2, *m3;
5869 MDB_dbi dbi = csrc->mc_dbi;
5870 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5872 if (csrc->mc_flags & C_SUB)
5875 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5876 if (m2 == csrc) continue;
5877 if (csrc->mc_flags & C_SUB)
5878 m3 = &m2->mc_xcursor->mx_cursor;
5881 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5882 csrc->mc_ki[csrc->mc_top]) {
5883 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5884 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5889 /* Update the parent separators.
5891 if (csrc->mc_ki[csrc->mc_top] == 0) {
5892 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5893 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5894 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5896 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5897 key.mv_size = NODEKSZ(srcnode);
5898 key.mv_data = NODEKEY(srcnode);
5900 DPRINTF("update separator for source page %zu to [%s]",
5901 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5902 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5903 &key)) != MDB_SUCCESS)
5906 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5908 nullkey.mv_size = 0;
5909 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5910 assert(rc == MDB_SUCCESS);
5914 if (cdst->mc_ki[cdst->mc_top] == 0) {
5915 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5916 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5917 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5919 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5920 key.mv_size = NODEKSZ(srcnode);
5921 key.mv_data = NODEKEY(srcnode);
5923 DPRINTF("update separator for destination page %zu to [%s]",
5924 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5925 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5926 &key)) != MDB_SUCCESS)
5929 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5931 nullkey.mv_size = 0;
5932 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5933 assert(rc == MDB_SUCCESS);
5940 /** Merge one page into another.
5941 * The nodes from the page pointed to by \b csrc will
5942 * be copied to the page pointed to by \b cdst and then
5943 * the \b csrc page will be freed.
5944 * @param[in] csrc Cursor pointing to the source page.
5945 * @param[in] cdst Cursor pointing to the destination page.
5948 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5956 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5957 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5959 assert(csrc->mc_snum > 1); /* can't merge root page */
5960 assert(cdst->mc_snum > 1);
5962 /* Mark dst as dirty. */
5963 if ((rc = mdb_page_touch(cdst)))
5966 /* Move all nodes from src to dst.
5968 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5969 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5970 key.mv_size = csrc->mc_db->md_pad;
5971 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5972 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5973 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5974 if (rc != MDB_SUCCESS)
5976 key.mv_data = (char *)key.mv_data + key.mv_size;
5979 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5980 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5981 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5982 unsigned int snum = csrc->mc_snum;
5984 /* must find the lowest key below src */
5985 mdb_page_search_root(csrc, NULL, 0);
5986 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5987 key.mv_size = csrc->mc_db->md_pad;
5988 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5990 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5991 key.mv_size = NODEKSZ(s2);
5992 key.mv_data = NODEKEY(s2);
5994 csrc->mc_snum = snum--;
5995 csrc->mc_top = snum;
5997 key.mv_size = srcnode->mn_ksize;
5998 key.mv_data = NODEKEY(srcnode);
6001 data.mv_size = NODEDSZ(srcnode);
6002 data.mv_data = NODEDATA(srcnode);
6003 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6004 if (rc != MDB_SUCCESS)
6009 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6010 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);
6012 /* Unlink the src page from parent and add to free list.
6014 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6015 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6017 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
6021 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6022 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6023 csrc->mc_db->md_leaf_pages--;
6025 csrc->mc_db->md_branch_pages--;
6027 /* Adjust other cursors pointing to mp */
6028 MDB_cursor *m2, *m3;
6029 MDB_dbi dbi = csrc->mc_dbi;
6030 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6032 if (csrc->mc_flags & C_SUB)
6035 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6036 if (csrc->mc_flags & C_SUB)
6037 m3 = &m2->mc_xcursor->mx_cursor;
6040 if (m3 == csrc) continue;
6041 if (m3->mc_snum < csrc->mc_snum) continue;
6042 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6043 m3->mc_pg[csrc->mc_top] = mp;
6044 m3->mc_ki[csrc->mc_top] += nkeys;
6048 mdb_cursor_pop(csrc);
6050 return mdb_rebalance(csrc);
6053 /** Copy the contents of a cursor.
6054 * @param[in] csrc The cursor to copy from.
6055 * @param[out] cdst The cursor to copy to.
6058 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6062 cdst->mc_txn = csrc->mc_txn;
6063 cdst->mc_dbi = csrc->mc_dbi;
6064 cdst->mc_db = csrc->mc_db;
6065 cdst->mc_dbx = csrc->mc_dbx;
6066 cdst->mc_snum = csrc->mc_snum;
6067 cdst->mc_top = csrc->mc_top;
6068 cdst->mc_flags = csrc->mc_flags;
6070 for (i=0; i<csrc->mc_snum; i++) {
6071 cdst->mc_pg[i] = csrc->mc_pg[i];
6072 cdst->mc_ki[i] = csrc->mc_ki[i];
6076 /** Rebalance the tree after a delete operation.
6077 * @param[in] mc Cursor pointing to the page where rebalancing
6079 * @return 0 on success, non-zero on failure.
6082 mdb_rebalance(MDB_cursor *mc)
6092 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6093 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6094 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6095 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6099 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
6102 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6103 DPRINTF("no need to rebalance page %zu, above fill threshold",
6109 if (mc->mc_snum < 2) {
6110 MDB_page *mp = mc->mc_pg[0];
6111 if (NUMKEYS(mp) == 0) {
6112 DPUTS("tree is completely empty");
6113 mc->mc_db->md_root = P_INVALID;
6114 mc->mc_db->md_depth = 0;
6115 mc->mc_db->md_leaf_pages = 0;
6116 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6120 /* Adjust other cursors pointing to mp */
6121 MDB_cursor *m2, *m3;
6122 MDB_dbi dbi = mc->mc_dbi;
6124 if (mc->mc_flags & C_SUB)
6127 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6128 if (m2 == mc) continue;
6129 if (mc->mc_flags & C_SUB)
6130 m3 = &m2->mc_xcursor->mx_cursor;
6133 if (m3->mc_snum < mc->mc_snum) continue;
6134 if (m3->mc_pg[0] == mp) {
6140 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6141 DPUTS("collapsing root page!");
6142 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6143 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6144 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6147 mc->mc_db->md_depth--;
6148 mc->mc_db->md_branch_pages--;
6150 /* Adjust other cursors pointing to mp */
6151 MDB_cursor *m2, *m3;
6152 MDB_dbi dbi = mc->mc_dbi;
6154 if (mc->mc_flags & C_SUB)
6157 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6158 if (m2 == mc) continue;
6159 if (mc->mc_flags & C_SUB)
6160 m3 = &m2->mc_xcursor->mx_cursor;
6163 if (m3->mc_snum < mc->mc_snum) continue;
6164 if (m3->mc_pg[0] == mp) {
6165 m3->mc_pg[0] = mc->mc_pg[0];
6170 DPUTS("root page doesn't need rebalancing");
6174 /* The parent (branch page) must have at least 2 pointers,
6175 * otherwise the tree is invalid.
6177 ptop = mc->mc_top-1;
6178 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6180 /* Leaf page fill factor is below the threshold.
6181 * Try to move keys from left or right neighbor, or
6182 * merge with a neighbor page.
6187 mdb_cursor_copy(mc, &mn);
6188 mn.mc_xcursor = NULL;
6190 if (mc->mc_ki[ptop] == 0) {
6191 /* We're the leftmost leaf in our parent.
6193 DPUTS("reading right neighbor");
6195 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6196 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6198 mn.mc_ki[mn.mc_top] = 0;
6199 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6201 /* There is at least one neighbor to the left.
6203 DPUTS("reading left neighbor");
6205 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6206 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6208 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6209 mc->mc_ki[mc->mc_top] = 0;
6212 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6213 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);
6215 /* If the neighbor page is above threshold and has at least two
6216 * keys, move one key from it.
6218 * Otherwise we should try to merge them.
6220 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6221 return mdb_node_move(&mn, mc);
6222 else { /* FIXME: if (has_enough_room()) */
6223 mc->mc_flags &= ~C_INITIALIZED;
6224 if (mc->mc_ki[ptop] == 0)
6225 return mdb_page_merge(&mn, mc);
6227 return mdb_page_merge(mc, &mn);
6231 /** Complete a delete operation started by #mdb_cursor_del(). */
6233 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6237 /* add overflow pages to free list */
6238 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6242 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6243 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6244 mc->mc_db->md_overflow_pages -= ovpages;
6245 for (i=0; i<ovpages; i++) {
6246 DPRINTF("freed ov page %zu", pg);
6247 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6251 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6252 mc->mc_db->md_entries--;
6253 rc = mdb_rebalance(mc);
6254 if (rc != MDB_SUCCESS)
6255 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6261 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6262 MDB_val *key, MDB_val *data)
6267 MDB_val rdata, *xdata;
6271 assert(key != NULL);
6273 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6275 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6278 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6282 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6286 mdb_cursor_init(&mc, txn, dbi, &mx);
6297 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6299 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6303 /** Split a page and insert a new node.
6304 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6305 * The cursor will be updated to point to the actual page and index where
6306 * the node got inserted after the split.
6307 * @param[in] newkey The key for the newly inserted node.
6308 * @param[in] newdata The data for the newly inserted node.
6309 * @param[in] newpgno The page number, if the new node is a branch node.
6310 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6311 * @return 0 on success, non-zero on failure.
6314 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6315 unsigned int nflags)
6318 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6321 unsigned int i, j, split_indx, nkeys, pmax;
6323 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6325 MDB_page *mp, *rp, *pp;
6330 mp = mc->mc_pg[mc->mc_top];
6331 newindx = mc->mc_ki[mc->mc_top];
6333 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6334 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6335 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6337 /* Create a right sibling. */
6338 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6340 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6342 if (mc->mc_snum < 2) {
6343 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6345 /* shift current top to make room for new parent */
6346 mc->mc_pg[1] = mc->mc_pg[0];
6347 mc->mc_ki[1] = mc->mc_ki[0];
6350 mc->mc_db->md_root = pp->mp_pgno;
6351 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6352 mc->mc_db->md_depth++;
6355 /* Add left (implicit) pointer. */
6356 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6357 /* undo the pre-push */
6358 mc->mc_pg[0] = mc->mc_pg[1];
6359 mc->mc_ki[0] = mc->mc_ki[1];
6360 mc->mc_db->md_root = mp->mp_pgno;
6361 mc->mc_db->md_depth--;
6368 ptop = mc->mc_top-1;
6369 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6372 mc->mc_flags |= C_SPLITTING;
6373 mdb_cursor_copy(mc, &mn);
6374 mn.mc_pg[mn.mc_top] = rp;
6375 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6377 if (nflags & MDB_APPEND) {
6378 mn.mc_ki[mn.mc_top] = 0;
6380 split_indx = newindx;
6385 nkeys = NUMKEYS(mp);
6386 split_indx = nkeys / 2;
6387 if (newindx < split_indx)
6393 unsigned int lsize, rsize, ksize;
6394 /* Move half of the keys to the right sibling */
6396 x = mc->mc_ki[mc->mc_top] - split_indx;
6397 ksize = mc->mc_db->md_pad;
6398 split = LEAF2KEY(mp, split_indx, ksize);
6399 rsize = (nkeys - split_indx) * ksize;
6400 lsize = (nkeys - split_indx) * sizeof(indx_t);
6401 mp->mp_lower -= lsize;
6402 rp->mp_lower += lsize;
6403 mp->mp_upper += rsize - lsize;
6404 rp->mp_upper -= rsize - lsize;
6405 sepkey.mv_size = ksize;
6406 if (newindx == split_indx) {
6407 sepkey.mv_data = newkey->mv_data;
6409 sepkey.mv_data = split;
6412 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6413 memcpy(rp->mp_ptrs, split, rsize);
6414 sepkey.mv_data = rp->mp_ptrs;
6415 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6416 memcpy(ins, newkey->mv_data, ksize);
6417 mp->mp_lower += sizeof(indx_t);
6418 mp->mp_upper -= ksize - sizeof(indx_t);
6421 memcpy(rp->mp_ptrs, split, x * ksize);
6422 ins = LEAF2KEY(rp, x, ksize);
6423 memcpy(ins, newkey->mv_data, ksize);
6424 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6425 rp->mp_lower += sizeof(indx_t);
6426 rp->mp_upper -= ksize - sizeof(indx_t);
6427 mc->mc_ki[mc->mc_top] = x;
6428 mc->mc_pg[mc->mc_top] = rp;
6433 /* For leaf pages, check the split point based on what
6434 * fits where, since otherwise mdb_node_add can fail.
6436 * This check is only needed when the data items are
6437 * relatively large, such that being off by one will
6438 * make the difference between success or failure.
6440 * It's also relevant if a page happens to be laid out
6441 * such that one half of its nodes are all "small" and
6442 * the other half of its nodes are "large." If the new
6443 * item is also "large" and falls on the half with
6444 * "large" nodes, it also may not fit.
6447 unsigned int psize, nsize;
6448 /* Maximum free space in an empty page */
6449 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6450 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6451 if ((nkeys < 20) || (nsize > pmax/16)) {
6452 if (newindx <= split_indx) {
6455 for (i=0; i<split_indx; i++) {
6456 node = NODEPTR(mp, i);
6457 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6458 if (F_ISSET(node->mn_flags, F_BIGDATA))
6459 psize += sizeof(pgno_t);
6461 psize += NODEDSZ(node);
6465 split_indx = newindx;
6476 for (i=nkeys-1; i>=split_indx; i--) {
6477 node = NODEPTR(mp, i);
6478 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6479 if (F_ISSET(node->mn_flags, F_BIGDATA))
6480 psize += sizeof(pgno_t);
6482 psize += NODEDSZ(node);
6486 split_indx = newindx;
6497 /* First find the separating key between the split pages.
6498 * The case where newindx == split_indx is ambiguous; the
6499 * new item could go to the new page or stay on the original
6500 * page. If newpos == 1 it goes to the new page.
6502 if (newindx == split_indx && newpos) {
6503 sepkey.mv_size = newkey->mv_size;
6504 sepkey.mv_data = newkey->mv_data;
6506 node = NODEPTR(mp, split_indx);
6507 sepkey.mv_size = node->mn_ksize;
6508 sepkey.mv_data = NODEKEY(node);
6512 DPRINTF("separator is [%s]", DKEY(&sepkey));
6514 /* Copy separator key to the parent.
6516 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6520 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6523 if (mn.mc_snum == mc->mc_snum) {
6524 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6525 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6526 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6527 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6532 /* Right page might now have changed parent.
6533 * Check if left page also changed parent.
6535 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6536 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6537 for (i=0; i<ptop; i++) {
6538 mc->mc_pg[i] = mn.mc_pg[i];
6539 mc->mc_ki[i] = mn.mc_ki[i];
6541 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6542 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6546 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6549 mc->mc_flags ^= C_SPLITTING;
6550 if (rc != MDB_SUCCESS) {
6553 if (nflags & MDB_APPEND) {
6554 mc->mc_pg[mc->mc_top] = rp;
6555 mc->mc_ki[mc->mc_top] = 0;
6556 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6559 for (i=0; i<mc->mc_top; i++)
6560 mc->mc_ki[i] = mn.mc_ki[i];
6567 /* Move half of the keys to the right sibling. */
6569 /* grab a page to hold a temporary copy */
6570 copy = mdb_page_malloc(mc);
6574 copy->mp_pgno = mp->mp_pgno;
6575 copy->mp_flags = mp->mp_flags;
6576 copy->mp_lower = PAGEHDRSZ;
6577 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6578 mc->mc_pg[mc->mc_top] = copy;
6579 for (i = j = 0; i <= nkeys; j++) {
6580 if (i == split_indx) {
6581 /* Insert in right sibling. */
6582 /* Reset insert index for right sibling. */
6583 if (i != newindx || (newpos ^ ins_new)) {
6585 mc->mc_pg[mc->mc_top] = rp;
6589 if (i == newindx && !ins_new) {
6590 /* Insert the original entry that caused the split. */
6591 rkey.mv_data = newkey->mv_data;
6592 rkey.mv_size = newkey->mv_size;
6601 /* Update index for the new key. */
6602 mc->mc_ki[mc->mc_top] = j;
6603 } else if (i == nkeys) {
6606 node = NODEPTR(mp, i);
6607 rkey.mv_data = NODEKEY(node);
6608 rkey.mv_size = node->mn_ksize;
6610 xdata.mv_data = NODEDATA(node);
6611 xdata.mv_size = NODEDSZ(node);
6614 pgno = NODEPGNO(node);
6615 flags = node->mn_flags;
6620 if (!IS_LEAF(mp) && j == 0) {
6621 /* First branch index doesn't need key data. */
6625 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6629 nkeys = NUMKEYS(copy);
6630 for (i=0; i<nkeys; i++)
6631 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6632 mp->mp_lower = copy->mp_lower;
6633 mp->mp_upper = copy->mp_upper;
6634 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6635 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6637 /* reset back to original page */
6638 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6639 mc->mc_pg[mc->mc_top] = mp;
6640 if (nflags & MDB_RESERVE) {
6641 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6642 if (!(node->mn_flags & F_BIGDATA))
6643 newdata->mv_data = NODEDATA(node);
6649 /* return tmp page to freelist */
6650 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
6651 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
6652 mc->mc_txn->mt_env->me_dpages = copy;
6655 /* Adjust other cursors pointing to mp */
6656 MDB_cursor *m2, *m3;
6657 MDB_dbi dbi = mc->mc_dbi;
6658 int fixup = NUMKEYS(mp);
6660 if (mc->mc_flags & C_SUB)
6663 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6664 if (m2 == mc) continue;
6665 if (mc->mc_flags & C_SUB)
6666 m3 = &m2->mc_xcursor->mx_cursor;
6669 if (!(m3->mc_flags & C_INITIALIZED))
6671 if (m3->mc_flags & C_SPLITTING)
6676 for (k=m3->mc_top; k>=0; k--) {
6677 m3->mc_ki[k+1] = m3->mc_ki[k];
6678 m3->mc_pg[k+1] = m3->mc_pg[k];
6680 if (m3->mc_ki[0] >= split_indx) {
6685 m3->mc_pg[0] = mc->mc_pg[0];
6689 if (m3->mc_pg[mc->mc_top] == mp) {
6690 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6691 m3->mc_ki[mc->mc_top]++;
6692 if (m3->mc_ki[mc->mc_top] >= fixup) {
6693 m3->mc_pg[mc->mc_top] = rp;
6694 m3->mc_ki[mc->mc_top] -= fixup;
6695 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6697 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6698 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6707 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6708 MDB_val *key, MDB_val *data, unsigned int flags)
6713 assert(key != NULL);
6714 assert(data != NULL);
6716 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6719 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6723 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6727 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6730 mdb_cursor_init(&mc, txn, dbi, &mx);
6731 return mdb_cursor_put(&mc, key, data, flags);
6735 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6737 if ((flag & CHANGEABLE) != flag)
6740 env->me_flags |= flag;
6742 env->me_flags &= ~flag;
6747 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6752 *arg = env->me_flags;
6757 mdb_env_get_path(MDB_env *env, const char **arg)
6762 *arg = env->me_path;
6766 /** Common code for #mdb_stat() and #mdb_env_stat().
6767 * @param[in] env the environment to operate in.
6768 * @param[in] db the #MDB_db record containing the stats to return.
6769 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6770 * @return 0, this function always succeeds.
6773 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6775 arg->ms_psize = env->me_psize;
6776 arg->ms_depth = db->md_depth;
6777 arg->ms_branch_pages = db->md_branch_pages;
6778 arg->ms_leaf_pages = db->md_leaf_pages;
6779 arg->ms_overflow_pages = db->md_overflow_pages;
6780 arg->ms_entries = db->md_entries;
6785 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6789 if (env == NULL || arg == NULL)
6792 toggle = mdb_env_pick_meta(env);
6794 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6798 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
6802 if (env == NULL || arg == NULL)
6805 toggle = mdb_env_pick_meta(env);
6806 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
6807 arg->me_mapsize = env->me_mapsize;
6808 arg->me_maxreaders = env->me_maxreaders;
6809 arg->me_numreaders = env->me_numreaders;
6810 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
6811 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
6815 /** Set the default comparison functions for a database.
6816 * Called immediately after a database is opened to set the defaults.
6817 * The user can then override them with #mdb_set_compare() or
6818 * #mdb_set_dupsort().
6819 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6820 * @param[in] dbi A database handle returned by #mdb_dbi_open()
6823 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6825 uint16_t f = txn->mt_dbs[dbi].md_flags;
6827 txn->mt_dbxs[dbi].md_cmp =
6828 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6829 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6831 txn->mt_dbxs[dbi].md_dcmp =
6832 !(f & MDB_DUPSORT) ? 0 :
6833 ((f & MDB_INTEGERDUP)
6834 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6835 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6838 #define PERSISTENT_FLAGS 0xffff
6839 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
6840 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
6841 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6846 int rc, dbflag, exact;
6847 unsigned int unused = 0;
6850 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6851 mdb_default_cmp(txn, FREE_DBI);
6854 if ((flags & VALID_FLAGS) != flags)
6860 if (flags & PERSISTENT_FLAGS) {
6861 uint16_t f2 = flags & PERSISTENT_FLAGS;
6862 /* make sure flag changes get committed */
6863 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
6864 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
6865 txn->mt_flags |= MDB_TXN_DIRTY;
6868 mdb_default_cmp(txn, MAIN_DBI);
6872 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6873 mdb_default_cmp(txn, MAIN_DBI);
6876 /* Is the DB already open? */
6878 for (i=2; i<txn->mt_numdbs; i++) {
6879 if (!txn->mt_dbxs[i].md_name.mv_size) {
6880 /* Remember this free slot */
6881 if (!unused) unused = i;
6884 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6885 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6891 /* If no free slot and max hit, fail */
6892 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
6893 return MDB_DBS_FULL;
6895 /* Find the DB info */
6899 key.mv_data = (void *)name;
6900 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6901 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6902 if (rc == MDB_SUCCESS) {
6903 /* make sure this is actually a DB */
6904 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6905 if (!(node->mn_flags & F_SUBDATA))
6907 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6908 /* Create if requested */
6910 data.mv_size = sizeof(MDB_db);
6911 data.mv_data = &dummy;
6912 memset(&dummy, 0, sizeof(dummy));
6913 dummy.md_root = P_INVALID;
6914 dummy.md_flags = flags & PERSISTENT_FLAGS;
6915 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6919 /* OK, got info, add to table */
6920 if (rc == MDB_SUCCESS) {
6921 unsigned int slot = unused ? unused : txn->mt_numdbs;
6922 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
6923 txn->mt_dbxs[slot].md_name.mv_size = len;
6924 txn->mt_dbxs[slot].md_rel = NULL;
6925 txn->mt_dbflags[slot] = dbflag;
6926 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
6928 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
6929 mdb_default_cmp(txn, slot);
6932 txn->mt_env->me_numdbs++;
6939 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6941 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6944 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6947 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
6950 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6952 ptr = env->me_dbxs[dbi].md_name.mv_data;
6953 env->me_dbxs[dbi].md_name.mv_data = NULL;
6954 env->me_dbxs[dbi].md_name.mv_size = 0;
6958 /** Add all the DB's pages to the free list.
6959 * @param[in] mc Cursor on the DB to free.
6960 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6961 * @return 0 on success, non-zero on failure.
6964 mdb_drop0(MDB_cursor *mc, int subs)
6968 rc = mdb_page_search(mc, NULL, 0);
6969 if (rc == MDB_SUCCESS) {
6974 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6975 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6978 mdb_cursor_copy(mc, &mx);
6979 while (mc->mc_snum > 0) {
6980 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6981 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6982 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6983 if (ni->mn_flags & F_SUBDATA) {
6984 mdb_xcursor_init1(mc, ni);
6985 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6991 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6993 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6996 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7001 rc = mdb_cursor_sibling(mc, 1);
7003 /* no more siblings, go back to beginning
7004 * of previous level.
7007 for (i=1; i<mc->mc_top; i++)
7008 mc->mc_pg[i] = mx.mc_pg[i];
7012 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7013 mc->mc_db->md_root);
7018 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7023 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1)
7026 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7029 rc = mdb_cursor_open(txn, dbi, &mc);
7033 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7037 /* Can't delete the main DB */
7038 if (del && dbi > MAIN_DBI) {
7039 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7041 mdb_dbi_close(txn->mt_env, dbi);
7043 /* reset the DB record, mark it dirty */
7044 txn->mt_dbflags[dbi] |= DB_DIRTY;
7045 txn->mt_dbs[dbi].md_depth = 0;
7046 txn->mt_dbs[dbi].md_branch_pages = 0;
7047 txn->mt_dbs[dbi].md_leaf_pages = 0;
7048 txn->mt_dbs[dbi].md_overflow_pages = 0;
7049 txn->mt_dbs[dbi].md_entries = 0;
7050 txn->mt_dbs[dbi].md_root = P_INVALID;
7052 if (!txn->mt_u.dirty_list[0].mid) {
7055 /* make sure we have at least one dirty page in this txn
7056 * otherwise these changes will be ignored.
7058 key.mv_size = sizeof(txnid_t);
7059 key.mv_data = &txn->mt_txnid;
7060 data.mv_size = sizeof(MDB_ID);
7061 data.mv_data = txn->mt_free_pgs;
7062 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
7063 rc = mdb_cursor_put(&m2, &key, &data, 0);
7067 mdb_cursor_close(mc);
7071 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7073 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7076 txn->mt_dbxs[dbi].md_cmp = cmp;
7080 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7082 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7085 txn->mt_dbxs[dbi].md_dcmp = cmp;
7089 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7091 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7094 txn->mt_dbxs[dbi].md_rel = rel;
7098 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7100 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7103 txn->mt_dbxs[dbi].md_relctx = ctx;