2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
36 #include <sys/types.h>
38 #include <sys/param.h>
44 #ifdef HAVE_SYS_FILE_H
61 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
62 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
65 #if defined(__APPLE__) || defined (BSD)
66 # define MDB_USE_POSIX_SEM 1
67 # define MDB_FDATASYNC fsync
68 #elif defined(ANDROID)
69 # define MDB_FDATASYNC fsync
74 #ifdef MDB_USE_POSIX_SEM
75 #include <semaphore.h>
80 #include <valgrind/memcheck.h>
81 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
82 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
83 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
84 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
85 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
87 #define VGMEMP_CREATE(h,r,z)
88 #define VGMEMP_ALLOC(h,a,s)
89 #define VGMEMP_FREE(h,a)
90 #define VGMEMP_DESTROY(h)
91 #define VGMEMP_DEFINED(a,s)
95 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
96 /* Solaris just defines one or the other */
97 # define LITTLE_ENDIAN 1234
98 # define BIG_ENDIAN 4321
99 # ifdef _LITTLE_ENDIAN
100 # define BYTE_ORDER LITTLE_ENDIAN
102 # define BYTE_ORDER BIG_ENDIAN
105 # define BYTE_ORDER __BYTE_ORDER
109 #ifndef LITTLE_ENDIAN
110 #define LITTLE_ENDIAN __LITTLE_ENDIAN
113 #define BIG_ENDIAN __BIG_ENDIAN
116 #if defined(__i386) || defined(__x86_64)
117 #define MISALIGNED_OK 1
123 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
124 # error "Unknown or unsupported endianness (BYTE_ORDER)"
125 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
126 # error "Two's complement, reasonably sized integer types, please"
129 /** @defgroup internal MDB Internals
132 /** @defgroup compat Windows Compatibility Macros
133 * A bunch of macros to minimize the amount of platform-specific ifdefs
134 * needed throughout the rest of the code. When the features this library
135 * needs are similar enough to POSIX to be hidden in a one-or-two line
136 * replacement, this macro approach is used.
140 #define pthread_t DWORD
141 #define pthread_mutex_t HANDLE
142 #define pthread_key_t DWORD
143 #define pthread_self() GetCurrentThreadId()
144 #define pthread_key_create(x,y) \
145 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
146 #define pthread_key_delete(x) TlsFree(x)
147 #define pthread_getspecific(x) TlsGetValue(x)
148 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
149 #define pthread_mutex_unlock(x) ReleaseMutex(x)
150 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
151 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
152 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
153 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
154 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
155 #define getpid() GetCurrentProcessId()
156 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
157 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
158 #define ErrCode() GetLastError()
159 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
160 #define close(fd) CloseHandle(fd)
161 #define munmap(ptr,len) UnmapViewOfFile(ptr)
164 #ifdef MDB_USE_POSIX_SEM
166 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
167 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
168 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
169 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
172 mdb_sem_wait(sem_t *sem)
175 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
180 /** Lock the reader mutex.
182 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
183 /** Unlock the reader mutex.
185 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
187 /** Lock the writer mutex.
188 * Only a single write transaction is allowed at a time. Other writers
189 * will block waiting for this mutex.
191 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
192 /** Unlock the writer mutex.
194 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
195 #endif /* MDB_USE_POSIX_SEM */
197 /** Get the error code for the last failed system function.
199 #define ErrCode() errno
201 /** An abstraction for a file handle.
202 * On POSIX systems file handles are small integers. On Windows
203 * they're opaque pointers.
207 /** A value for an invalid file handle.
208 * Mainly used to initialize file variables and signify that they are
211 #define INVALID_HANDLE_VALUE (-1)
213 /** Get the size of a memory page for the system.
214 * This is the basic size that the platform's memory manager uses, and is
215 * fundamental to the use of memory-mapped files.
217 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
220 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
223 #define MNAME_LEN (sizeof(pthread_mutex_t))
229 /** A flag for opening a file and requesting synchronous data writes.
230 * This is only used when writing a meta page. It's not strictly needed;
231 * we could just do a normal write and then immediately perform a flush.
232 * But if this flag is available it saves us an extra system call.
234 * @note If O_DSYNC is undefined but exists in /usr/include,
235 * preferably set some compiler flag to get the definition.
236 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
239 # define MDB_DSYNC O_DSYNC
243 /** Function for flushing the data of a file. Define this to fsync
244 * if fdatasync() is not supported.
246 #ifndef MDB_FDATASYNC
247 # define MDB_FDATASYNC fdatasync
251 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
262 /** A page number in the database.
263 * Note that 64 bit page numbers are overkill, since pages themselves
264 * already represent 12-13 bits of addressable memory, and the OS will
265 * always limit applications to a maximum of 63 bits of address space.
267 * @note In the #MDB_node structure, we only store 48 bits of this value,
268 * which thus limits us to only 60 bits of addressable data.
270 typedef MDB_ID pgno_t;
272 /** A transaction ID.
273 * See struct MDB_txn.mt_txnid for details.
275 typedef MDB_ID txnid_t;
277 /** @defgroup debug Debug Macros
281 /** Enable debug output.
282 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
283 * read from and written to the database (used for free space management).
288 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
289 # define DPRINTF (void) /* Vararg macros may be unsupported */
291 static int mdb_debug;
292 static txnid_t mdb_debug_start;
294 /** Print a debug message with printf formatting. */
295 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
296 ((void) ((mdb_debug) && \
297 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
299 # define DPRINTF(fmt, ...) ((void) 0)
300 # define MDB_DEBUG_SKIP
302 /** Print a debug string.
303 * The string is printed literally, with no format processing.
305 #define DPUTS(arg) DPRINTF("%s", arg)
308 /** A default memory page size.
309 * The actual size is platform-dependent, but we use this for
310 * boot-strapping. We probably should not be using this any more.
311 * The #GET_PAGESIZE() macro is used to get the actual size.
313 * Note that we don't currently support Huge pages. On Linux,
314 * regular data files cannot use Huge pages, and in general
315 * Huge pages aren't actually pageable. We rely on the OS
316 * demand-pager to read our data and page it out when memory
317 * pressure from other processes is high. So until OSs have
318 * actual paging support for Huge pages, they're not viable.
320 #define MDB_PAGESIZE 4096
322 /** The minimum number of keys required in a database page.
323 * Setting this to a larger value will place a smaller bound on the
324 * maximum size of a data item. Data items larger than this size will
325 * be pushed into overflow pages instead of being stored directly in
326 * the B-tree node. This value used to default to 4. With a page size
327 * of 4096 bytes that meant that any item larger than 1024 bytes would
328 * go into an overflow page. That also meant that on average 2-3KB of
329 * each overflow page was wasted space. The value cannot be lower than
330 * 2 because then there would no longer be a tree structure. With this
331 * value, items larger than 2KB will go into overflow pages, and on
332 * average only 1KB will be wasted.
334 #define MDB_MINKEYS 2
336 /** A stamp that identifies a file as an MDB file.
337 * There's nothing special about this value other than that it is easily
338 * recognizable, and it will reflect any byte order mismatches.
340 #define MDB_MAGIC 0xBEEFC0DE
342 /** The version number for a database's file format. */
343 #define MDB_VERSION 1
345 /** @brief The maximum size of a key in the database.
347 * We require that keys all fit onto a regular page. This limit
348 * could be raised a bit further if needed; to something just
349 * under #MDB_PAGESIZE / #MDB_MINKEYS.
351 * Note that data items in an #MDB_DUPSORT database are actually keys
352 * of a subDB, so they're also limited to this size.
354 #ifndef MDB_MAXKEYSIZE
355 #define MDB_MAXKEYSIZE 511
358 /** @brief The maximum size of a data item.
360 * We only store a 32 bit value for node sizes.
362 #define MAXDATASIZE 0xffffffffUL
367 * This is used for printing a hex dump of a key's contents.
369 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
370 /** Display a key in hex.
372 * Invoke a function to display a key in hex.
374 #define DKEY(x) mdb_dkey(x, kbuf)
376 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
380 /** An invalid page number.
381 * Mainly used to denote an empty tree.
383 #define P_INVALID (~(pgno_t)0)
385 /** Test if the flags \b f are set in a flag word \b w. */
386 #define F_ISSET(w, f) (((w) & (f)) == (f))
388 /** Used for offsets within a single page.
389 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
392 typedef uint16_t indx_t;
394 /** Default size of memory map.
395 * This is certainly too small for any actual applications. Apps should always set
396 * the size explicitly using #mdb_env_set_mapsize().
398 #define DEFAULT_MAPSIZE 1048576
400 /** @defgroup readers Reader Lock Table
401 * Readers don't acquire any locks for their data access. Instead, they
402 * simply record their transaction ID in the reader table. The reader
403 * mutex is needed just to find an empty slot in the reader table. The
404 * slot's address is saved in thread-specific data so that subsequent read
405 * transactions started by the same thread need no further locking to proceed.
407 * No reader table is used if the database is on a read-only filesystem.
409 * Since the database uses multi-version concurrency control, readers don't
410 * actually need any locking. This table is used to keep track of which
411 * readers are using data from which old transactions, so that we'll know
412 * when a particular old transaction is no longer in use. Old transactions
413 * that have discarded any data pages can then have those pages reclaimed
414 * for use by a later write transaction.
416 * The lock table is constructed such that reader slots are aligned with the
417 * processor's cache line size. Any slot is only ever used by one thread.
418 * This alignment guarantees that there will be no contention or cache
419 * thrashing as threads update their own slot info, and also eliminates
420 * any need for locking when accessing a slot.
422 * A writer thread will scan every slot in the table to determine the oldest
423 * outstanding reader transaction. Any freed pages older than this will be
424 * reclaimed by the writer. The writer doesn't use any locks when scanning
425 * this table. This means that there's no guarantee that the writer will
426 * see the most up-to-date reader info, but that's not required for correct
427 * operation - all we need is to know the upper bound on the oldest reader,
428 * we don't care at all about the newest reader. So the only consequence of
429 * reading stale information here is that old pages might hang around a
430 * while longer before being reclaimed. That's actually good anyway, because
431 * the longer we delay reclaiming old pages, the more likely it is that a
432 * string of contiguous pages can be found after coalescing old pages from
433 * many old transactions together.
435 * @todo We don't actually do such coalescing yet, we grab pages from one
436 * old transaction at a time.
439 /** Number of slots in the reader table.
440 * This value was chosen somewhat arbitrarily. 126 readers plus a
441 * couple mutexes fit exactly into 8KB on my development machine.
442 * Applications should set the table size using #mdb_env_set_maxreaders().
444 #define DEFAULT_READERS 126
446 /** The size of a CPU cache line in bytes. We want our lock structures
447 * aligned to this size to avoid false cache line sharing in the
449 * This value works for most CPUs. For Itanium this should be 128.
455 /** The information we store in a single slot of the reader table.
456 * In addition to a transaction ID, we also record the process and
457 * thread ID that owns a slot, so that we can detect stale information,
458 * e.g. threads or processes that went away without cleaning up.
459 * @note We currently don't check for stale records. We simply re-init
460 * the table when we know that we're the only process opening the
463 typedef struct MDB_rxbody {
464 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
465 * Multiple readers that start at the same time will probably have the
466 * same ID here. Again, it's not important to exclude them from
467 * anything; all we need to know is which version of the DB they
468 * started from so we can avoid overwriting any data used in that
469 * particular version.
472 /** The process ID of the process owning this reader txn. */
474 /** The thread ID of the thread owning this txn. */
478 /** The actual reader record, with cacheline padding. */
479 typedef struct MDB_reader {
482 /** shorthand for mrb_txnid */
483 #define mr_txnid mru.mrx.mrb_txnid
484 #define mr_pid mru.mrx.mrb_pid
485 #define mr_tid mru.mrx.mrb_tid
486 /** cache line alignment */
487 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
491 /** The header for the reader table.
492 * The table resides in a memory-mapped file. (This is a different file
493 * than is used for the main database.)
495 * For POSIX the actual mutexes reside in the shared memory of this
496 * mapped file. On Windows, mutexes are named objects allocated by the
497 * kernel; we store the mutex names in this mapped file so that other
498 * processes can grab them. This same approach is also used on
499 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
500 * process-shared POSIX mutexes. For these cases where a named object
501 * is used, the object name is derived from a 64 bit FNV hash of the
502 * environment pathname. As such, naming collisions are extremely
503 * unlikely. If a collision occurs, the results are unpredictable.
505 typedef struct MDB_txbody {
506 /** Stamp identifying this as an MDB file. It must be set
509 /** Version number of this lock file. Must be set to #MDB_VERSION. */
510 uint32_t mtb_version;
511 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
512 char mtb_rmname[MNAME_LEN];
514 /** Mutex protecting access to this table.
515 * This is the reader lock that #LOCK_MUTEX_R acquires.
517 pthread_mutex_t mtb_mutex;
519 /** The ID of the last transaction committed to the database.
520 * This is recorded here only for convenience; the value can always
521 * be determined by reading the main database meta pages.
524 /** The number of slots that have been used in the reader table.
525 * This always records the maximum count, it is not decremented
526 * when readers release their slots.
528 unsigned mtb_numreaders;
531 /** The actual reader table definition. */
532 typedef struct MDB_txninfo {
535 #define mti_magic mt1.mtb.mtb_magic
536 #define mti_version mt1.mtb.mtb_version
537 #define mti_mutex mt1.mtb.mtb_mutex
538 #define mti_rmname mt1.mtb.mtb_rmname
539 #define mti_txnid mt1.mtb.mtb_txnid
540 #define mti_numreaders mt1.mtb.mtb_numreaders
541 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
544 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
545 char mt2_wmname[MNAME_LEN];
546 #define mti_wmname mt2.mt2_wmname
548 pthread_mutex_t mt2_wmutex;
549 #define mti_wmutex mt2.mt2_wmutex
551 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
553 MDB_reader mti_readers[1];
557 /** Common header for all page types.
558 * Overflow records occupy a number of contiguous pages with no
559 * headers on any page after the first.
561 typedef struct MDB_page {
562 #define mp_pgno mp_p.p_pgno
563 #define mp_next mp_p.p_next
565 pgno_t p_pgno; /**< page number */
566 void * p_next; /**< for in-memory list of freed structs */
569 /** @defgroup mdb_page Page Flags
571 * Flags for the page headers.
574 #define P_BRANCH 0x01 /**< branch page */
575 #define P_LEAF 0x02 /**< leaf page */
576 #define P_OVERFLOW 0x04 /**< overflow page */
577 #define P_META 0x08 /**< meta page */
578 #define P_DIRTY 0x10 /**< dirty page */
579 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
580 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
582 uint16_t mp_flags; /**< @ref mdb_page */
583 #define mp_lower mp_pb.pb.pb_lower
584 #define mp_upper mp_pb.pb.pb_upper
585 #define mp_pages mp_pb.pb_pages
588 indx_t pb_lower; /**< lower bound of free space */
589 indx_t pb_upper; /**< upper bound of free space */
591 uint32_t pb_pages; /**< number of overflow pages */
593 indx_t mp_ptrs[1]; /**< dynamic size */
596 /** Size of the page header, excluding dynamic data at the end */
597 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
599 /** Address of first usable data byte in a page, after the header */
600 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
602 /** Number of nodes on a page */
603 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
605 /** The amount of space remaining in the page */
606 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
608 /** The percentage of space used in the page, in tenths of a percent. */
609 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
610 ((env)->me_psize - PAGEHDRSZ))
611 /** The minimum page fill factor, in tenths of a percent.
612 * Pages emptier than this are candidates for merging.
614 #define FILL_THRESHOLD 250
616 /** Test if a page is a leaf page */
617 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
618 /** Test if a page is a LEAF2 page */
619 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
620 /** Test if a page is a branch page */
621 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
622 /** Test if a page is an overflow page */
623 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
624 /** Test if a page is a sub page */
625 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
627 /** The number of overflow pages needed to store the given size. */
628 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
630 /** Header for a single key/data pair within a page.
631 * We guarantee 2-byte alignment for nodes.
633 typedef struct MDB_node {
634 /** lo and hi are used for data size on leaf nodes and for
635 * child pgno on branch nodes. On 64 bit platforms, flags
636 * is also used for pgno. (Branch nodes have no flags).
637 * They are in host byte order in case that lets some
638 * accesses be optimized into a 32-bit word access.
640 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
641 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
642 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
643 /** @defgroup mdb_node Node Flags
645 * Flags for node headers.
648 #define F_BIGDATA 0x01 /**< data put on overflow page */
649 #define F_SUBDATA 0x02 /**< data is a sub-database */
650 #define F_DUPDATA 0x04 /**< data has duplicates */
652 /** valid flags for #mdb_node_add() */
653 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
656 unsigned short mn_flags; /**< @ref mdb_node */
657 unsigned short mn_ksize; /**< key size */
658 char mn_data[1]; /**< key and data are appended here */
661 /** Size of the node header, excluding dynamic data at the end */
662 #define NODESIZE offsetof(MDB_node, mn_data)
664 /** Bit position of top word in page number, for shifting mn_flags */
665 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
667 /** Size of a node in a branch page with a given key.
668 * This is just the node header plus the key, there is no data.
670 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
672 /** Size of a node in a leaf page with a given key and data.
673 * This is node header plus key plus data size.
675 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
677 /** Address of node \b i in page \b p */
678 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
680 /** Address of the key for the node */
681 #define NODEKEY(node) (void *)((node)->mn_data)
683 /** Address of the data for a node */
684 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
686 /** Get the page number pointed to by a branch node */
687 #define NODEPGNO(node) \
688 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
689 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
690 /** Set the page number in a branch node */
691 #define SETPGNO(node,pgno) do { \
692 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
693 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
695 /** Get the size of the data in a leaf node */
696 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
697 /** Set the size of the data for a leaf node */
698 #define SETDSZ(node,size) do { \
699 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
700 /** The size of a key in a node */
701 #define NODEKSZ(node) ((node)->mn_ksize)
703 /** Copy a page number from src to dst */
705 #define COPY_PGNO(dst,src) dst = src
707 #if SIZE_MAX > 4294967295UL
708 #define COPY_PGNO(dst,src) do { \
709 unsigned short *s, *d; \
710 s = (unsigned short *)&(src); \
711 d = (unsigned short *)&(dst); \
718 #define COPY_PGNO(dst,src) do { \
719 unsigned short *s, *d; \
720 s = (unsigned short *)&(src); \
721 d = (unsigned short *)&(dst); \
727 /** The address of a key in a LEAF2 page.
728 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
729 * There are no node headers, keys are stored contiguously.
731 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
733 /** Set the \b node's key into \b key, if requested. */
734 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
735 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
737 /** Information about a single database in the environment. */
738 typedef struct MDB_db {
739 uint32_t md_pad; /**< also ksize for LEAF2 pages */
740 uint16_t md_flags; /**< @ref mdb_dbi_open */
741 uint16_t md_depth; /**< depth of this tree */
742 pgno_t md_branch_pages; /**< number of internal pages */
743 pgno_t md_leaf_pages; /**< number of leaf pages */
744 pgno_t md_overflow_pages; /**< number of overflow pages */
745 size_t md_entries; /**< number of data items */
746 pgno_t md_root; /**< the root page of this tree */
749 /** Handle for the DB used to track free pages. */
751 /** Handle for the default DB. */
754 /** Meta page content. */
755 typedef struct MDB_meta {
756 /** Stamp identifying this as an MDB file. It must be set
759 /** Version number of this lock file. Must be set to #MDB_VERSION. */
761 void *mm_address; /**< address for fixed mapping */
762 size_t mm_mapsize; /**< size of mmap region */
763 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
764 /** The size of pages used in this DB */
765 #define mm_psize mm_dbs[0].md_pad
766 /** Any persistent environment flags. @ref mdb_env */
767 #define mm_flags mm_dbs[0].md_flags
768 pgno_t mm_last_pg; /**< last used page in file */
769 txnid_t mm_txnid; /**< txnid that committed this page */
772 /** Buffer for a stack-allocated dirty page.
773 * The members define size and alignment, and silence type
774 * aliasing warnings. They are not used directly; that could
775 * mean incorrectly using several union members in parallel.
777 typedef union MDB_pagebuf {
778 char mb_raw[MDB_PAGESIZE];
781 char mm_pad[PAGEHDRSZ];
786 /** Auxiliary DB info.
787 * The information here is mostly static/read-only. There is
788 * only a single copy of this record in the environment.
790 typedef struct MDB_dbx {
791 MDB_val md_name; /**< name of the database */
792 MDB_cmp_func *md_cmp; /**< function for comparing keys */
793 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
794 MDB_rel_func *md_rel; /**< user relocate function */
795 void *md_relctx; /**< user-provided context for md_rel */
798 /** A database transaction.
799 * Every operation requires a transaction handle.
802 MDB_txn *mt_parent; /**< parent of a nested txn */
803 MDB_txn *mt_child; /**< nested txn under this txn */
804 pgno_t mt_next_pgno; /**< next unallocated page */
805 /** The ID of this transaction. IDs are integers incrementing from 1.
806 * Only committed write transactions increment the ID. If a transaction
807 * aborts, the ID may be re-used by the next writer.
810 MDB_env *mt_env; /**< the DB environment */
811 /** The list of pages that became unused during this transaction.
815 MDB_ID2L dirty_list; /**< for write txns: modified pages */
816 MDB_reader *reader; /**< this thread's reader table slot or NULL */
818 /** Array of records for each DB known in the environment. */
820 /** Array of MDB_db records for each known DB */
822 /** @defgroup mt_dbflag Transaction DB Flags
826 #define DB_DIRTY 0x01 /**< DB was written in this txn */
827 #define DB_STALE 0x02 /**< DB record is older than txnID */
829 /** In write txns, array of cursors for each DB */
830 MDB_cursor **mt_cursors;
831 /** Array of flags for each DB */
832 unsigned char *mt_dbflags;
833 /** Number of DB records in use. This number only ever increments;
834 * we don't decrement it when individual DB handles are closed.
838 /** @defgroup mdb_txn Transaction Flags
842 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
843 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
844 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
846 unsigned int mt_flags; /**< @ref mdb_txn */
847 /** Tracks which of the two meta pages was used at the start
848 * of this transaction.
850 unsigned int mt_toggle;
853 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
854 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
855 * raise this on a 64 bit machine.
857 #define CURSOR_STACK 32
861 /** Cursors are used for all DB operations */
863 /** Next cursor on this DB in this txn */
865 /** Original cursor if this is a shadow */
867 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
868 struct MDB_xcursor *mc_xcursor;
869 /** The transaction that owns this cursor */
871 /** The database handle this cursor operates on */
873 /** The database record for this cursor */
875 /** The database auxiliary record for this cursor */
877 /** The @ref mt_dbflag for this database */
878 unsigned char *mc_dbflag;
879 unsigned short mc_snum; /**< number of pushed pages */
880 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
881 /** @defgroup mdb_cursor Cursor Flags
883 * Cursor state flags.
886 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
887 #define C_EOF 0x02 /**< No more data */
888 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
889 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
890 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
891 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
893 unsigned int mc_flags; /**< @ref mdb_cursor */
894 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
895 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
898 /** Context for sorted-dup records.
899 * We could have gone to a fully recursive design, with arbitrarily
900 * deep nesting of sub-databases. But for now we only handle these
901 * levels - main DB, optional sub-DB, sorted-duplicate DB.
903 typedef struct MDB_xcursor {
904 /** A sub-cursor for traversing the Dup DB */
905 MDB_cursor mx_cursor;
906 /** The database record for this Dup DB */
908 /** The auxiliary DB record for this Dup DB */
910 /** The @ref mt_dbflag for this Dup DB */
911 unsigned char mx_dbflag;
914 /** The database environment. */
916 HANDLE me_fd; /**< The main data file */
917 HANDLE me_lfd; /**< The lock file */
918 HANDLE me_mfd; /**< just for writing the meta pages */
919 /** Failed to update the meta page. Probably an I/O error. */
920 #define MDB_FATAL_ERROR 0x80000000U
921 /** Read-only Filesystem. Allow read access, no locking. */
922 #define MDB_ROFS 0x40000000U
923 /** Some fields are initialized. */
924 #define MDB_ENV_ACTIVE 0x20000000U
925 uint32_t me_flags; /**< @ref mdb_env */
926 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
927 unsigned int me_maxreaders; /**< size of the reader table */
928 unsigned int me_numreaders; /**< max numreaders set by this env */
929 MDB_dbi me_numdbs; /**< number of DBs opened */
930 MDB_dbi me_maxdbs; /**< size of the DB table */
931 pid_t me_pid; /**< process ID of this env */
932 char *me_path; /**< path to the DB files */
933 char *me_map; /**< the memory map of the data file */
934 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
935 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
936 MDB_txn *me_txn; /**< current write transaction */
937 size_t me_mapsize; /**< size of the data memory map */
938 off_t me_size; /**< current file size */
939 pgno_t me_maxpg; /**< me_mapsize / me_psize */
940 txnid_t me_pglast; /**< ID of last old page record we used */
941 MDB_dbx *me_dbxs; /**< array of static DB info */
942 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
943 pgno_t *me_pghead; /**< old pages reclaimed from freelist */
944 pgno_t *me_pgfree; /**< memory to free when dropping me_pghead */
945 pthread_key_t me_txkey; /**< thread-key for readers */
946 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
947 /** IDL of pages that became unused in a write txn */
949 /** ID2L of pages that were written during a write txn */
950 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
951 /** Max number of freelist items that can fit in a single overflow page */
952 unsigned int me_maxfree_1pg;
954 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
956 #elif defined(MDB_USE_POSIX_SEM)
957 sem_t *me_rmutex; /* Shared mutexes are not supported */
961 /** max number of pages to commit in one writev() call */
962 #define MDB_COMMIT_PAGES 64
963 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
964 #undef MDB_COMMIT_PAGES
965 #define MDB_COMMIT_PAGES IOV_MAX
968 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
969 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
970 static int mdb_page_touch(MDB_cursor *mc);
972 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
973 static int mdb_page_search_root(MDB_cursor *mc,
974 MDB_val *key, int modify);
975 #define MDB_PS_MODIFY 1
976 #define MDB_PS_ROOTONLY 2
977 static int mdb_page_search(MDB_cursor *mc,
978 MDB_val *key, int flags);
979 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
981 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
982 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
983 pgno_t newpgno, unsigned int nflags);
985 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
986 static int mdb_env_pick_meta(const MDB_env *env);
987 static int mdb_env_write_meta(MDB_txn *txn);
988 static void mdb_env_close0(MDB_env *env, int excl);
990 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
991 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
992 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
993 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
994 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
995 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
996 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
997 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
998 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1000 static int mdb_rebalance(MDB_cursor *mc);
1001 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1003 static void mdb_cursor_pop(MDB_cursor *mc);
1004 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1006 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1007 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1008 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1009 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1010 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1012 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1013 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1015 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1016 static void mdb_xcursor_init0(MDB_cursor *mc);
1017 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1019 static int mdb_drop0(MDB_cursor *mc, int subs);
1020 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1023 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1027 static SECURITY_DESCRIPTOR mdb_null_sd;
1028 static SECURITY_ATTRIBUTES mdb_all_sa;
1029 static int mdb_sec_inited;
1032 /** Return the library version info. */
1034 mdb_version(int *major, int *minor, int *patch)
1036 if (major) *major = MDB_VERSION_MAJOR;
1037 if (minor) *minor = MDB_VERSION_MINOR;
1038 if (patch) *patch = MDB_VERSION_PATCH;
1039 return MDB_VERSION_STRING;
1042 /** Table of descriptions for MDB @ref errors */
1043 static char *const mdb_errstr[] = {
1044 "MDB_KEYEXIST: Key/data pair already exists",
1045 "MDB_NOTFOUND: No matching key/data pair found",
1046 "MDB_PAGE_NOTFOUND: Requested page not found",
1047 "MDB_CORRUPTED: Located page was wrong type",
1048 "MDB_PANIC: Update of meta page failed",
1049 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1050 "MDB_INVALID: File is not an MDB file",
1051 "MDB_MAP_FULL: Environment mapsize limit reached",
1052 "MDB_DBS_FULL: Environment maxdbs limit reached",
1053 "MDB_READERS_FULL: Environment maxreaders limit reached",
1054 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1055 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1056 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1057 "MDB_PAGE_FULL: Internal error - page has no more space",
1058 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1062 mdb_strerror(int err)
1066 return ("Successful return: 0");
1068 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1069 i = err - MDB_KEYEXIST;
1070 return mdb_errstr[i];
1073 return strerror(err);
1077 /** Display a key in hexadecimal and return the address of the result.
1078 * @param[in] key the key to display
1079 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1080 * @return The key in hexadecimal form.
1083 mdb_dkey(MDB_val *key, char *buf)
1086 unsigned char *c = key->mv_data;
1092 if (key->mv_size > MDB_MAXKEYSIZE)
1093 return "MDB_MAXKEYSIZE";
1094 /* may want to make this a dynamic check: if the key is mostly
1095 * printable characters, print it as-is instead of converting to hex.
1099 for (i=0; i<key->mv_size; i++)
1100 ptr += sprintf(ptr, "%02x", *c++);
1102 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1107 /** Display all the keys in the page. */
1109 mdb_page_list(MDB_page *mp)
1112 unsigned int i, nkeys, nsize;
1116 nkeys = NUMKEYS(mp);
1117 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1118 for (i=0; i<nkeys; i++) {
1119 node = NODEPTR(mp, i);
1120 key.mv_size = node->mn_ksize;
1121 key.mv_data = node->mn_data;
1122 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1123 if (IS_BRANCH(mp)) {
1124 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1127 if (F_ISSET(node->mn_flags, F_BIGDATA))
1128 nsize += sizeof(pgno_t);
1130 nsize += NODEDSZ(node);
1131 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1137 mdb_cursor_chk(MDB_cursor *mc)
1143 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1144 for (i=0; i<mc->mc_top; i++) {
1146 node = NODEPTR(mp, mc->mc_ki[i]);
1147 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1150 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1156 /** Count all the pages in each DB and in the freelist
1157 * and make sure it matches the actual number of pages
1160 static void mdb_audit(MDB_txn *txn)
1164 MDB_ID freecount, count;
1169 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1170 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1171 freecount += *(MDB_ID *)data.mv_data;
1174 for (i = 0; i<txn->mt_numdbs; i++) {
1175 MDB_xcursor mx, *mxp;
1176 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1177 mdb_cursor_init(&mc, txn, i, mxp);
1178 if (txn->mt_dbs[i].md_root == P_INVALID)
1180 count += txn->mt_dbs[i].md_branch_pages +
1181 txn->mt_dbs[i].md_leaf_pages +
1182 txn->mt_dbs[i].md_overflow_pages;
1183 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1184 mdb_page_search(&mc, NULL, 0);
1188 mp = mc.mc_pg[mc.mc_top];
1189 for (j=0; j<NUMKEYS(mp); j++) {
1190 MDB_node *leaf = NODEPTR(mp, j);
1191 if (leaf->mn_flags & F_SUBDATA) {
1193 memcpy(&db, NODEDATA(leaf), sizeof(db));
1194 count += db.md_branch_pages + db.md_leaf_pages +
1195 db.md_overflow_pages;
1199 while (mdb_cursor_sibling(&mc, 1) == 0);
1202 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1203 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1204 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1210 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1212 return txn->mt_dbxs[dbi].md_cmp(a, b);
1216 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1218 if (txn->mt_dbxs[dbi].md_dcmp)
1219 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1221 return EINVAL; /* too bad you can't distinguish this from a valid result */
1224 /** Allocate a single page.
1225 * Re-use old malloc'd pages first, otherwise just malloc.
1228 mdb_page_malloc(MDB_cursor *mc) {
1230 size_t sz = mc->mc_txn->mt_env->me_psize;
1231 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1232 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1233 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1234 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1235 } else if ((ret = malloc(sz)) != NULL) {
1236 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1242 mdb_page_free(MDB_env *env, MDB_page *mp)
1244 mp->mp_next = env->me_dpages;
1245 VGMEMP_FREE(env, mp);
1246 env->me_dpages = mp;
1249 /** Allocate pages for writing.
1250 * If there are free pages available from older transactions, they
1251 * will be re-used first. Otherwise a new page will be allocated.
1252 * @param[in] mc cursor A cursor handle identifying the transaction and
1253 * database for which we are allocating.
1254 * @param[in] num the number of pages to allocate.
1255 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1256 * will always be satisfied by a single contiguous chunk of memory.
1257 * @return 0 on success, non-zero on failure.
1260 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1262 MDB_txn *txn = mc->mc_txn;
1264 pgno_t pgno = P_INVALID;
1266 txnid_t oldest = 0, last;
1271 /* If our dirty list is already full, we can't do anything */
1272 if (txn->mt_u.dirty_list[0].mid >= MDB_IDL_UM_MAX)
1273 return MDB_TXN_FULL;
1275 /* The free list won't have any content at all until txn 2 has
1276 * committed. The pages freed by txn 2 will be unreferenced
1277 * after txn 3 commits, and so will be safe to re-use in txn 4.
1279 if (txn->mt_txnid > 3) {
1280 if (!txn->mt_env->me_pghead &&
1281 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1282 /* See if there's anything in the free DB */
1289 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1290 if (!txn->mt_env->me_pglast) {
1291 mdb_page_search(&m2, NULL, 0);
1292 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1293 kptr = (txnid_t *)NODEKEY(leaf);
1300 last = txn->mt_env->me_pglast + 1;
1302 key.mv_data = &last;
1303 key.mv_size = sizeof(last);
1304 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1307 last = *(txnid_t *)key.mv_data;
1313 oldest = txn->mt_txnid - 1;
1314 nr = txn->mt_env->me_txns->mti_numreaders;
1315 r = txn->mt_env->me_txns->mti_readers;
1316 for (i=0; i<nr; i++) {
1317 if (!r[i].mr_pid) continue;
1324 if (oldest > last) {
1325 /* It's usable, grab it.
1329 if (!txn->mt_env->me_pglast) {
1330 mdb_node_read(txn, leaf, &data);
1332 idl = (MDB_ID *) data.mv_data;
1333 /* We might have a zero-length IDL due to freelist growth
1334 * during a prior commit
1337 txn->mt_env->me_pglast = last;
1340 mop = malloc(MDB_IDL_SIZEOF(idl));
1343 txn->mt_env->me_pglast = last;
1344 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop;
1345 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1350 DPRINTF("IDL read txn %zu root %zu num %zu",
1351 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1352 for (i=0; i<idl[0]; i++) {
1353 DPRINTF("IDL %zu", idl[i+1]);
1360 if (txn->mt_env->me_pghead) {
1361 pgno_t *mop = txn->mt_env->me_pghead;
1364 int retry = 1, readit = 0, n2 = num-1;
1365 unsigned int i, j, k;
1367 /* If current list is too short, must fetch more and coalesce */
1368 if (mop[0] < (unsigned)num)
1371 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1373 /* If on freelist, don't try to read more. If what we have
1374 * right now isn't enough just use new pages.
1375 * TODO: get all of this working. Many circular dependencies...
1377 if (mc->mc_dbi == FREE_DBI) {
1386 last = txn->mt_env->me_pglast + 1;
1388 /* We haven't hit the readers list yet? */
1394 oldest = txn->mt_txnid - 1;
1395 nr = txn->mt_env->me_txns->mti_numreaders;
1396 r = txn->mt_env->me_txns->mti_readers;
1397 for (i=0; i<nr; i++) {
1398 if (!r[i].mr_pid) continue;
1405 /* There's nothing we can use on the freelist */
1406 if (oldest - last < 1)
1410 key.mv_data = &last;
1411 key.mv_size = sizeof(last);
1412 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1415 idl = (MDB_ID *) data.mv_data;
1416 mop2 = malloc(MDB_IDL_SIZEOF(idl) + MDB_IDL_SIZEOF(mop));
1419 /* merge in sorted order */
1420 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1422 while (i>0 || j>0) {
1423 if (i && idl[i] < mop[j])
1424 mop2[k--] = idl[i--];
1426 mop2[k--] = mop[j--];
1428 txn->mt_env->me_pglast = last;
1429 free(txn->mt_env->me_pgfree);
1430 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop2;
1432 /* Keep trying to read until we have enough */
1433 if (mop[0] < (unsigned)num) {
1438 /* current list has enough pages, but are they contiguous? */
1439 for (i=mop[0]; i>=(unsigned)num; i--) {
1440 if (mop[i-n2] == mop[i] + n2) {
1443 /* move any stragglers down */
1444 for (j=i+num; j<=mop[0]; j++)
1451 /* Stop if we succeeded, or no retries */
1452 if (!retry || pgno != P_INVALID)
1458 /* peel pages off tail, so we only have to truncate the list */
1459 pgno = MDB_IDL_LAST(mop);
1462 if (MDB_IDL_IS_ZERO(mop)) {
1463 free(txn->mt_env->me_pgfree);
1464 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1469 if (pgno == P_INVALID) {
1470 /* DB size is maxed out */
1471 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1472 DPUTS("DB size maxed out");
1473 return MDB_MAP_FULL;
1476 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1477 if (pgno == P_INVALID) {
1478 pgno = txn->mt_next_pgno;
1479 txn->mt_next_pgno += num;
1481 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1484 if (txn->mt_env->me_dpages && num == 1) {
1485 np = txn->mt_env->me_dpages;
1486 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1487 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1488 txn->mt_env->me_dpages = np->mp_next;
1490 size_t sz = txn->mt_env->me_psize * num;
1491 if ((np = malloc(sz)) == NULL)
1493 VGMEMP_ALLOC(txn->mt_env, np, sz);
1495 if (pgno == P_INVALID) {
1496 np->mp_pgno = txn->mt_next_pgno;
1497 txn->mt_next_pgno += num;
1502 mid.mid = np->mp_pgno;
1504 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1505 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1507 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1514 /** Copy a page: avoid copying unused portions of the page.
1515 * @param[in] dst page to copy into
1516 * @param[in] src page to copy from
1519 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1521 dst->mp_flags = src->mp_flags | P_DIRTY;
1522 dst->mp_pages = src->mp_pages;
1524 if (IS_LEAF2(src)) {
1525 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1527 unsigned int i, nkeys = NUMKEYS(src);
1528 for (i=0; i<nkeys; i++)
1529 dst->mp_ptrs[i] = src->mp_ptrs[i];
1530 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1531 psize - src->mp_upper);
1535 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1536 * @param[in] mc cursor pointing to the page to be touched
1537 * @return 0 on success, non-zero on failure.
1540 mdb_page_touch(MDB_cursor *mc)
1542 MDB_page *mp = mc->mc_pg[mc->mc_top];
1546 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1548 if ((rc = mdb_page_alloc(mc, 1, &np)))
1550 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1551 assert(mp->mp_pgno != np->mp_pgno);
1552 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1554 /* If page isn't full, just copy the used portion */
1555 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1558 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1560 np->mp_flags |= P_DIRTY;
1565 /* Adjust other cursors pointing to mp */
1566 if (mc->mc_flags & C_SUB) {
1567 MDB_cursor *m2, *m3;
1568 MDB_dbi dbi = mc->mc_dbi-1;
1570 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1571 if (m2 == mc) continue;
1572 m3 = &m2->mc_xcursor->mx_cursor;
1573 if (m3->mc_snum < mc->mc_snum) continue;
1574 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1575 m3->mc_pg[mc->mc_top] = mp;
1581 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1582 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1583 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1584 m2->mc_pg[mc->mc_top] = mp;
1588 mc->mc_pg[mc->mc_top] = mp;
1589 /** If this page has a parent, update the parent to point to
1593 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1595 mc->mc_db->md_root = mp->mp_pgno;
1596 } else if (mc->mc_txn->mt_parent) {
1599 /* If txn has a parent, make sure the page is in our
1602 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1603 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1604 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1605 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1606 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1607 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1608 mc->mc_pg[mc->mc_top] = mp;
1613 if (mc->mc_txn->mt_u.dirty_list[0].mid >= MDB_IDL_UM_MAX)
1614 return MDB_TXN_FULL;
1616 np = mdb_page_malloc(mc);
1619 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1620 mid.mid = np->mp_pgno;
1622 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1630 mdb_env_sync(MDB_env *env, int force)
1633 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1634 if (env->me_flags & MDB_WRITEMAP) {
1635 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1636 ? MS_ASYNC : MS_SYNC;
1637 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1640 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1644 if (MDB_FDATASYNC(env->me_fd))
1651 /** Make shadow copies of all of parent txn's cursors */
1653 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1655 MDB_cursor *mc, *m2;
1656 unsigned int i, j, size;
1658 for (i=0;i<src->mt_numdbs; i++) {
1659 if (src->mt_cursors[i]) {
1660 size = sizeof(MDB_cursor);
1661 if (src->mt_cursors[i]->mc_xcursor)
1662 size += sizeof(MDB_xcursor);
1663 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1670 mc->mc_db = &dst->mt_dbs[i];
1671 mc->mc_dbx = m2->mc_dbx;
1672 mc->mc_dbflag = &dst->mt_dbflags[i];
1673 mc->mc_snum = m2->mc_snum;
1674 mc->mc_top = m2->mc_top;
1675 mc->mc_flags = m2->mc_flags | C_SHADOW;
1676 for (j=0; j<mc->mc_snum; j++) {
1677 mc->mc_pg[j] = m2->mc_pg[j];
1678 mc->mc_ki[j] = m2->mc_ki[j];
1680 if (m2->mc_xcursor) {
1681 MDB_xcursor *mx, *mx2;
1682 mx = (MDB_xcursor *)(mc+1);
1683 mc->mc_xcursor = mx;
1684 mx2 = m2->mc_xcursor;
1685 mx->mx_db = mx2->mx_db;
1686 mx->mx_dbx = mx2->mx_dbx;
1687 mx->mx_dbflag = mx2->mx_dbflag;
1688 mx->mx_cursor.mc_txn = dst;
1689 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1690 mx->mx_cursor.mc_db = &mx->mx_db;
1691 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1692 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1693 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1694 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1695 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1696 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1697 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1698 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1701 mc->mc_xcursor = NULL;
1703 mc->mc_next = dst->mt_cursors[i];
1704 dst->mt_cursors[i] = mc;
1711 /** Merge shadow cursors back into parent's */
1713 mdb_cursor_merge(MDB_txn *txn)
1716 for (i=0; i<txn->mt_numdbs; i++) {
1717 if (txn->mt_cursors[i]) {
1719 while ((mc = txn->mt_cursors[i])) {
1720 txn->mt_cursors[i] = mc->mc_next;
1721 if (mc->mc_flags & C_SHADOW) {
1722 MDB_cursor *m2 = mc->mc_orig;
1724 m2->mc_snum = mc->mc_snum;
1725 m2->mc_top = mc->mc_top;
1726 for (j=0; j<mc->mc_snum; j++) {
1727 m2->mc_pg[j] = mc->mc_pg[j];
1728 m2->mc_ki[j] = mc->mc_ki[j];
1731 if (mc->mc_flags & C_ALLOCD)
1739 mdb_txn_reset0(MDB_txn *txn);
1741 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1742 * @param[in] txn the transaction handle to initialize
1743 * @return 0 on success, non-zero on failure. This can only
1744 * fail for read-only transactions, and then only if the
1745 * reader table is full.
1748 mdb_txn_renew0(MDB_txn *txn)
1750 MDB_env *env = txn->mt_env;
1755 txn->mt_numdbs = env->me_numdbs;
1756 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1758 if (txn->mt_flags & MDB_TXN_RDONLY) {
1759 if (env->me_flags & MDB_ROFS) {
1760 i = mdb_env_pick_meta(env);
1761 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1762 txn->mt_u.reader = NULL;
1764 MDB_reader *r = pthread_getspecific(env->me_txkey);
1766 pid_t pid = env->me_pid;
1767 pthread_t tid = pthread_self();
1770 for (i=0; i<env->me_txns->mti_numreaders; i++)
1771 if (env->me_txns->mti_readers[i].mr_pid == 0)
1773 if (i == env->me_maxreaders) {
1774 UNLOCK_MUTEX_R(env);
1775 return MDB_READERS_FULL;
1777 env->me_txns->mti_readers[i].mr_pid = pid;
1778 env->me_txns->mti_readers[i].mr_tid = tid;
1779 if (i >= env->me_txns->mti_numreaders)
1780 env->me_txns->mti_numreaders = i+1;
1781 /* Save numreaders for un-mutexed mdb_env_close() */
1782 env->me_numreaders = env->me_txns->mti_numreaders;
1783 UNLOCK_MUTEX_R(env);
1784 r = &env->me_txns->mti_readers[i];
1785 if ((rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1786 env->me_txns->mti_readers[i].mr_pid = 0;
1790 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1791 txn->mt_u.reader = r;
1793 txn->mt_toggle = txn->mt_txnid & 1;
1794 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1798 txn->mt_txnid = env->me_txns->mti_txnid;
1799 txn->mt_toggle = txn->mt_txnid & 1;
1800 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1803 if (txn->mt_txnid == mdb_debug_start)
1806 txn->mt_u.dirty_list = env->me_dirty_list;
1807 txn->mt_u.dirty_list[0].mid = 0;
1808 txn->mt_free_pgs = env->me_free_pgs;
1809 txn->mt_free_pgs[0] = 0;
1813 /* Copy the DB info and flags */
1814 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1815 for (i=2; i<txn->mt_numdbs; i++)
1816 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1817 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1818 if (txn->mt_numdbs > 2)
1819 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1821 if (env->me_maxpg < txn->mt_next_pgno) {
1822 mdb_txn_reset0(txn);
1823 return MDB_MAP_RESIZED;
1830 mdb_txn_renew(MDB_txn *txn)
1834 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1837 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1838 DPUTS("environment had fatal error, must shutdown!");
1842 rc = mdb_txn_renew0(txn);
1843 if (rc == MDB_SUCCESS) {
1844 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1845 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1846 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1852 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1857 if (env->me_flags & MDB_FATAL_ERROR) {
1858 DPUTS("environment had fatal error, must shutdown!");
1861 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1864 /* Nested transactions: Max 1 child, write txns only, no writemap */
1865 if (parent->mt_child ||
1866 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1867 (env->me_flags & MDB_WRITEMAP))
1872 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1873 if (!(flags & MDB_RDONLY))
1874 size += env->me_maxdbs * sizeof(MDB_cursor *);
1876 if ((txn = calloc(1, size)) == NULL) {
1877 DPRINTF("calloc: %s", strerror(ErrCode()));
1880 txn->mt_dbs = (MDB_db *)(txn+1);
1881 if (flags & MDB_RDONLY) {
1882 txn->mt_flags |= MDB_TXN_RDONLY;
1883 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1885 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1886 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1891 txn->mt_free_pgs = mdb_midl_alloc();
1892 if (!txn->mt_free_pgs) {
1896 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1897 if (!txn->mt_u.dirty_list) {
1898 free(txn->mt_free_pgs);
1902 txn->mt_txnid = parent->mt_txnid;
1903 txn->mt_toggle = parent->mt_toggle;
1904 txn->mt_u.dirty_list[0].mid = 0;
1905 txn->mt_free_pgs[0] = 0;
1906 txn->mt_next_pgno = parent->mt_next_pgno;
1907 parent->mt_child = txn;
1908 txn->mt_parent = parent;
1909 txn->mt_numdbs = parent->mt_numdbs;
1910 txn->mt_dbxs = parent->mt_dbxs;
1911 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1912 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1913 mdb_cursor_shadow(parent, txn);
1916 rc = mdb_txn_renew0(txn);
1922 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1923 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1924 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1930 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1931 * @param[in] txn the transaction handle to reset
1934 mdb_txn_reset0(MDB_txn *txn)
1936 MDB_env *env = txn->mt_env;
1938 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1939 if (!(env->me_flags & MDB_ROFS))
1940 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1945 /* close(free) all cursors */
1946 for (i=0; i<txn->mt_numdbs; i++) {
1947 if (txn->mt_cursors[i]) {
1949 while ((mc = txn->mt_cursors[i])) {
1950 txn->mt_cursors[i] = mc->mc_next;
1951 if (mc->mc_flags & C_ALLOCD)
1957 if (!(env->me_flags & MDB_WRITEMAP)) {
1958 /* return all dirty pages to dpage list */
1959 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1960 dp = txn->mt_u.dirty_list[i].mptr;
1961 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1962 mdb_page_free(txn->mt_env, dp);
1964 /* large pages just get freed directly */
1965 VGMEMP_FREE(txn->mt_env, dp);
1971 if (txn->mt_parent) {
1972 txn->mt_parent->mt_child = NULL;
1973 mdb_midl_free(txn->mt_free_pgs);
1974 free(txn->mt_u.dirty_list);
1977 if (mdb_midl_shrink(&txn->mt_free_pgs))
1978 env->me_free_pgs = txn->mt_free_pgs;
1981 free(txn->mt_env->me_pgfree);
1982 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1983 txn->mt_env->me_pglast = 0;
1986 /* The writer mutex was locked in mdb_txn_begin. */
1987 UNLOCK_MUTEX_W(env);
1992 mdb_txn_reset(MDB_txn *txn)
1997 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1998 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1999 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2001 mdb_txn_reset0(txn);
2005 mdb_txn_abort(MDB_txn *txn)
2010 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2011 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2012 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2015 mdb_txn_abort(txn->mt_child);
2017 mdb_txn_reset0(txn);
2022 mdb_txn_commit(MDB_txn *txn)
2030 pgno_t next, freecnt;
2031 txnid_t oldpg_txnid, id;
2034 assert(txn != NULL);
2035 assert(txn->mt_env != NULL);
2037 if (txn->mt_child) {
2038 mdb_txn_commit(txn->mt_child);
2039 txn->mt_child = NULL;
2044 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2045 if (txn->mt_numdbs > env->me_numdbs) {
2046 /* update the DB flags */
2048 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2049 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2056 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2057 DPUTS("error flag is set, can't commit");
2059 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2064 if (txn->mt_parent) {
2065 MDB_txn *parent = txn->mt_parent;
2069 parent->mt_next_pgno = txn->mt_next_pgno;
2070 parent->mt_flags = txn->mt_flags;
2072 /* Merge (and close) our cursors with parent's */
2073 mdb_cursor_merge(txn);
2075 /* Update parent's DB table. */
2076 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2077 memcpy(parent->mt_dbflags, txn->mt_dbflags, txn->mt_numdbs);
2078 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2080 /* Append our free list to parent's */
2081 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
2083 mdb_midl_free(txn->mt_free_pgs);
2085 /* Merge our dirty list with parent's */
2086 dst = txn->mt_parent->mt_u.dirty_list;
2087 src = txn->mt_u.dirty_list;
2088 x = mdb_mid2l_search(dst, src[1].mid);
2089 for (y=1; y<=src[0].mid; y++) {
2090 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
2094 dst[x].mptr = src[y].mptr;
2097 for (; y<=src[0].mid; y++) {
2098 if (++x >= MDB_IDL_UM_MAX) {
2100 return MDB_TXN_FULL;
2105 free(txn->mt_u.dirty_list);
2106 txn->mt_parent->mt_child = NULL;
2111 if (txn != env->me_txn) {
2112 DPUTS("attempt to commit unknown transaction");
2117 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2120 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2121 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2123 /* Update DB root pointers */
2124 if (txn->mt_numdbs > 2) {
2127 data.mv_size = sizeof(MDB_db);
2129 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2130 for (i = 2; i < txn->mt_numdbs; i++) {
2131 if (txn->mt_dbflags[i] & DB_DIRTY) {
2132 data.mv_data = &txn->mt_dbs[i];
2133 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2140 /* Save the freelist as of this transaction to the freeDB. This
2141 * can change the freelist, so keep trying until it stabilizes.
2143 * env->me_pglast and the length of txn->mt_free_pgs cannot decrease,
2144 * except the code below can decrease env->me_pglast to split pghead.
2145 * Page numbers cannot disappear from txn->mt_free_pgs. New pages
2146 * can only appear in env->me_pghead when env->me_pglast increases.
2147 * Until then, the me_pghead pointer won't move but can become NULL.
2150 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2151 oldpg_txnid = id = 0;
2154 /* should only be one record now */
2155 if (env->me_pghead || env->me_pglast) {
2156 /* make sure first page of freeDB is touched and on freelist */
2157 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2158 if (rc && rc != MDB_NOTFOUND) {
2165 /* Delete IDLs we used from the free list */
2166 if (env->me_pglast) {
2171 rc = mdb_cursor_first(&mc, &key, NULL);
2174 oldpg_txnid = *(txnid_t *)key.mv_data;
2176 assert(oldpg_txnid <= env->me_pglast);
2178 rc = mdb_cursor_del(&mc, 0);
2181 } while (oldpg_txnid < env->me_pglast);
2184 /* Save IDL of pages freed by this txn, to freeDB */
2186 if (freecnt != txn->mt_free_pgs[0]) {
2189 /* make sure last page of freeDB is touched and on freelist */
2190 key.mv_size = MDB_MAXKEYSIZE+1;
2192 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2193 if (rc && rc != MDB_NOTFOUND)
2199 MDB_IDL idl = txn->mt_free_pgs;
2200 mdb_midl_sort(txn->mt_free_pgs);
2201 DPRINTF("IDL write txn %zu root %zu num %zu",
2202 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2203 for (i=1; i<=idl[0]; i++) {
2204 DPRINTF("IDL %zu", idl[i]);
2208 /* write to last page of freeDB */
2209 key.mv_size = sizeof(pgno_t);
2210 key.mv_data = &txn->mt_txnid;
2211 /* The free list can still grow during this call,
2212 * despite the pre-emptive touches above. So retry
2213 * until the reserved space remains big enough.
2216 assert(freecnt < txn->mt_free_pgs[0]);
2217 freecnt = txn->mt_free_pgs[0];
2218 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2219 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2222 } while (freecnt != txn->mt_free_pgs[0]);
2223 mdb_midl_sort(txn->mt_free_pgs);
2224 memcpy(data.mv_data, txn->mt_free_pgs, data.mv_size);
2225 if (oldpg_txnid < env->me_pglast || (!env->me_pghead && id))
2226 goto free_pgfirst; /* used up freeDB[oldpg_txnid] */
2229 /* Put back page numbers we took from freeDB but did not use */
2230 if (env->me_pghead) {
2235 mop = env->me_pghead;
2236 id = env->me_pglast;
2237 key.mv_size = sizeof(id);
2239 /* These steps may grow the freelist again
2240 * due to freed overflow pages...
2245 if (orig > env->me_maxfree_1pg && id > 4)
2246 orig = env->me_maxfree_1pg; /* Do not use more than 1 page */
2247 data.mv_size = (orig + 1) * sizeof(pgno_t);
2248 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2251 assert(!env->me_pghead || env->me_pglast);
2252 /* mop could have been used again here */
2253 if (id != env->me_pglast || env->me_pghead == NULL)
2254 goto again; /* was completely used up */
2255 assert(mop == env->me_pghead);
2256 } while (mop[0] < orig && --i);
2257 memcpy(data.mv_data, mop, data.mv_size);
2260 *(pgno_t *)data.mv_data = orig;
2261 mop[orig] = mop[0] - orig;
2262 env->me_pghead = mop += orig;
2263 /* Save more oldpages at the previous txnid. */
2264 assert(env->me_pglast == id && id == oldpg_txnid);
2265 env->me_pglast = --oldpg_txnid;
2269 /* Check for growth of freelist again */
2270 if (freecnt != txn->mt_free_pgs[0])
2273 free(env->me_pgfree);
2274 env->me_pghead = env->me_pgfree = NULL;
2276 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2277 if (mdb_midl_shrink(&txn->mt_free_pgs))
2278 env->me_free_pgs = txn->mt_free_pgs;
2285 if (env->me_flags & MDB_WRITEMAP) {
2286 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2287 dp = txn->mt_u.dirty_list[i].mptr;
2288 /* clear dirty flag */
2289 dp->mp_flags &= ~P_DIRTY;
2290 txn->mt_u.dirty_list[i].mid = 0;
2292 txn->mt_u.dirty_list[0].mid = 0;
2296 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2302 /* Windows actually supports scatter/gather I/O, but only on
2303 * unbuffered file handles. Since we're relying on the OS page
2304 * cache for all our data, that's self-defeating. So we just
2305 * write pages one at a time. We use the ov structure to set
2306 * the write offset, to at least save the overhead of a Seek
2310 memset(&ov, 0, sizeof(ov));
2311 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2313 dp = txn->mt_u.dirty_list[i].mptr;
2314 DPRINTF("committing page %zu", dp->mp_pgno);
2315 size = dp->mp_pgno * env->me_psize;
2316 ov.Offset = size & 0xffffffff;
2317 ov.OffsetHigh = size >> 16;
2318 ov.OffsetHigh >>= 16;
2319 /* clear dirty flag */
2320 dp->mp_flags &= ~P_DIRTY;
2321 wsize = env->me_psize;
2322 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2323 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2326 DPRINTF("WriteFile: %d", n);
2333 struct iovec iov[MDB_COMMIT_PAGES];
2337 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2338 dp = txn->mt_u.dirty_list[i].mptr;
2339 if (dp->mp_pgno != next) {
2341 rc = writev(env->me_fd, iov, n);
2345 DPUTS("short write, filesystem full?");
2347 DPRINTF("writev: %s", strerror(n));
2354 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2357 DPRINTF("committing page %zu", dp->mp_pgno);
2358 iov[n].iov_len = env->me_psize;
2359 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2360 iov[n].iov_base = (char *)dp;
2361 size += iov[n].iov_len;
2362 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2363 /* clear dirty flag */
2364 dp->mp_flags &= ~P_DIRTY;
2365 if (++n >= MDB_COMMIT_PAGES) {
2375 rc = writev(env->me_fd, iov, n);
2379 DPUTS("short write, filesystem full?");
2381 DPRINTF("writev: %s", strerror(n));
2388 /* Drop the dirty pages.
2390 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2391 dp = txn->mt_u.dirty_list[i].mptr;
2392 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2393 mdb_page_free(txn->mt_env, dp);
2395 VGMEMP_FREE(txn->mt_env, dp);
2398 txn->mt_u.dirty_list[i].mid = 0;
2400 txn->mt_u.dirty_list[0].mid = 0;
2403 if ((n = mdb_env_sync(env, 0)) != 0 ||
2404 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2412 if (txn->mt_numdbs > env->me_numdbs) {
2413 /* update the DB flags */
2415 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2416 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2420 UNLOCK_MUTEX_W(env);
2426 /** Read the environment parameters of a DB environment before
2427 * mapping it into memory.
2428 * @param[in] env the environment handle
2429 * @param[out] meta address of where to store the meta information
2430 * @return 0 on success, non-zero on failure.
2433 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2440 /* We don't know the page size yet, so use a minimum value.
2441 * Read both meta pages so we can use the latest one.
2444 for (i=0; i<2; i++) {
2446 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2448 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2453 else if (rc != MDB_PAGESIZE) {
2457 DPRINTF("read: %s", strerror(err));
2461 p = (MDB_page *)&pbuf;
2463 if (!F_ISSET(p->mp_flags, P_META)) {
2464 DPRINTF("page %zu not a meta page", p->mp_pgno);
2469 if (m->mm_magic != MDB_MAGIC) {
2470 DPUTS("meta has invalid magic");
2474 if (m->mm_version != MDB_VERSION) {
2475 DPRINTF("database is version %u, expected version %u",
2476 m->mm_version, MDB_VERSION);
2477 return MDB_VERSION_MISMATCH;
2481 if (m->mm_txnid > meta->mm_txnid)
2482 memcpy(meta, m, sizeof(*m));
2484 memcpy(meta, m, sizeof(*m));
2486 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2488 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2496 /** Write the environment parameters of a freshly created DB environment.
2497 * @param[in] env the environment handle
2498 * @param[out] meta address of where to store the meta information
2499 * @return 0 on success, non-zero on failure.
2502 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2509 DPUTS("writing new meta page");
2511 GET_PAGESIZE(psize);
2513 meta->mm_magic = MDB_MAGIC;
2514 meta->mm_version = MDB_VERSION;
2515 meta->mm_mapsize = env->me_mapsize;
2516 meta->mm_psize = psize;
2517 meta->mm_last_pg = 1;
2518 meta->mm_flags = env->me_flags & 0xffff;
2519 meta->mm_flags |= MDB_INTEGERKEY;
2520 meta->mm_dbs[0].md_root = P_INVALID;
2521 meta->mm_dbs[1].md_root = P_INVALID;
2523 p = calloc(2, psize);
2525 p->mp_flags = P_META;
2528 memcpy(m, meta, sizeof(*meta));
2530 q = (MDB_page *)((char *)p + psize);
2533 q->mp_flags = P_META;
2536 memcpy(m, meta, sizeof(*meta));
2541 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2542 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2543 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2546 lseek(env->me_fd, 0, SEEK_SET);
2547 rc = write(env->me_fd, p, psize * 2);
2548 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2554 /** Update the environment info to commit a transaction.
2555 * @param[in] txn the transaction that's being committed
2556 * @return 0 on success, non-zero on failure.
2559 mdb_env_write_meta(MDB_txn *txn)
2562 MDB_meta meta, metab, *mp;
2564 int rc, len, toggle;
2571 assert(txn != NULL);
2572 assert(txn->mt_env != NULL);
2574 toggle = !txn->mt_toggle;
2575 DPRINTF("writing meta page %d for root page %zu",
2576 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2579 mp = env->me_metas[toggle];
2581 if (env->me_flags & MDB_WRITEMAP) {
2582 /* Persist any increases of mapsize config */
2583 if (env->me_mapsize > mp->mm_mapsize)
2584 mp->mm_mapsize = env->me_mapsize;
2585 mp->mm_dbs[0] = txn->mt_dbs[0];
2586 mp->mm_dbs[1] = txn->mt_dbs[1];
2587 mp->mm_last_pg = txn->mt_next_pgno - 1;
2588 mp->mm_txnid = txn->mt_txnid;
2589 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2590 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2593 ptr += env->me_psize;
2594 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2601 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2602 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2604 ptr = (char *)&meta;
2605 if (env->me_mapsize > mp->mm_mapsize) {
2606 /* Persist any increases of mapsize config */
2607 meta.mm_mapsize = env->me_mapsize;
2608 off = offsetof(MDB_meta, mm_mapsize);
2610 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2612 len = sizeof(MDB_meta) - off;
2615 meta.mm_dbs[0] = txn->mt_dbs[0];
2616 meta.mm_dbs[1] = txn->mt_dbs[1];
2617 meta.mm_last_pg = txn->mt_next_pgno - 1;
2618 meta.mm_txnid = txn->mt_txnid;
2621 off += env->me_psize;
2624 /* Write to the SYNC fd */
2625 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2626 env->me_fd : env->me_mfd;
2629 memset(&ov, 0, sizeof(ov));
2631 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2634 rc = pwrite(mfd, ptr, len, off);
2639 DPUTS("write failed, disk error?");
2640 /* On a failure, the pagecache still contains the new data.
2641 * Write some old data back, to prevent it from being used.
2642 * Use the non-SYNC fd; we know it will fail anyway.
2644 meta.mm_last_pg = metab.mm_last_pg;
2645 meta.mm_txnid = metab.mm_txnid;
2647 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2649 r2 = pwrite(env->me_fd, ptr, len, off);
2652 env->me_flags |= MDB_FATAL_ERROR;
2656 /* Memory ordering issues are irrelevant; since the entire writer
2657 * is wrapped by wmutex, all of these changes will become visible
2658 * after the wmutex is unlocked. Since the DB is multi-version,
2659 * readers will get consistent data regardless of how fresh or
2660 * how stale their view of these values is.
2662 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2667 /** Check both meta pages to see which one is newer.
2668 * @param[in] env the environment handle
2669 * @return meta toggle (0 or 1).
2672 mdb_env_pick_meta(const MDB_env *env)
2674 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2678 mdb_env_create(MDB_env **env)
2682 e = calloc(1, sizeof(MDB_env));
2686 e->me_free_pgs = mdb_midl_alloc();
2687 if (!e->me_free_pgs) {
2691 e->me_maxreaders = DEFAULT_READERS;
2693 e->me_fd = INVALID_HANDLE_VALUE;
2694 e->me_lfd = INVALID_HANDLE_VALUE;
2695 e->me_mfd = INVALID_HANDLE_VALUE;
2696 #ifdef MDB_USE_POSIX_SEM
2697 e->me_rmutex = SEM_FAILED;
2698 e->me_wmutex = SEM_FAILED;
2700 e->me_pid = getpid();
2701 VGMEMP_CREATE(e,0,0);
2707 mdb_env_set_mapsize(MDB_env *env, size_t size)
2711 env->me_mapsize = size;
2713 env->me_maxpg = env->me_mapsize / env->me_psize;
2718 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2722 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2727 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2729 if (env->me_map || readers < 1)
2731 env->me_maxreaders = readers;
2736 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2738 if (!env || !readers)
2740 *readers = env->me_maxreaders;
2744 /** Further setup required for opening an MDB environment
2747 mdb_env_open2(MDB_env *env)
2749 unsigned int flags = env->me_flags;
2750 int i, newenv = 0, prot;
2754 memset(&meta, 0, sizeof(meta));
2756 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2759 DPUTS("new mdbenv");
2763 /* Was a mapsize configured? */
2764 if (!env->me_mapsize) {
2765 /* If this is a new environment, take the default,
2766 * else use the size recorded in the existing env.
2768 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2769 } else if (env->me_mapsize < meta.mm_mapsize) {
2770 /* If the configured size is smaller, make sure it's
2771 * still big enough. Silently round up to minimum if not.
2773 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2774 if (env->me_mapsize < minsize)
2775 env->me_mapsize = minsize;
2781 LONG sizelo, sizehi;
2782 sizelo = env->me_mapsize & 0xffffffff;
2783 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2785 /* Windows won't create mappings for zero length files.
2786 * Just allocate the maxsize right now.
2789 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2790 if (!SetEndOfFile(env->me_fd))
2792 SetFilePointer(env->me_fd, 0, NULL, 0);
2794 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2795 PAGE_READWRITE : PAGE_READONLY,
2796 sizehi, sizelo, NULL);
2799 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2800 FILE_MAP_WRITE : FILE_MAP_READ,
2801 0, 0, env->me_mapsize, meta.mm_address);
2809 if (flags & MDB_WRITEMAP) {
2811 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2814 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2816 if (env->me_map == MAP_FAILED) {
2823 if (flags & MDB_FIXEDMAP)
2824 meta.mm_address = env->me_map;
2825 i = mdb_env_init_meta(env, &meta);
2826 if (i != MDB_SUCCESS) {
2829 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2830 /* Can happen because the address argument to mmap() is just a
2831 * hint. mmap() can pick another, e.g. if the range is in use.
2832 * The MAP_FIXED flag would prevent that, but then mmap could
2833 * instead unmap existing pages to make room for the new map.
2835 return EBUSY; /* TODO: Make a new MDB_* error code? */
2837 env->me_psize = meta.mm_psize;
2838 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2840 env->me_maxpg = env->me_mapsize / env->me_psize;
2842 p = (MDB_page *)env->me_map;
2843 env->me_metas[0] = METADATA(p);
2844 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2848 int toggle = mdb_env_pick_meta(env);
2849 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2851 DPRINTF("opened database version %u, pagesize %u",
2852 env->me_metas[0]->mm_version, env->me_psize);
2853 DPRINTF("using meta page %d", toggle);
2854 DPRINTF("depth: %u", db->md_depth);
2855 DPRINTF("entries: %zu", db->md_entries);
2856 DPRINTF("branch pages: %zu", db->md_branch_pages);
2857 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2858 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2859 DPRINTF("root: %zu", db->md_root);
2867 /** Release a reader thread's slot in the reader lock table.
2868 * This function is called automatically when a thread exits.
2869 * @param[in] ptr This points to the slot in the reader lock table.
2872 mdb_env_reader_dest(void *ptr)
2874 MDB_reader *reader = ptr;
2880 /** Junk for arranging thread-specific callbacks on Windows. This is
2881 * necessarily platform and compiler-specific. Windows supports up
2882 * to 1088 keys. Let's assume nobody opens more than 64 environments
2883 * in a single process, for now. They can override this if needed.
2885 #ifndef MAX_TLS_KEYS
2886 #define MAX_TLS_KEYS 64
2888 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2889 static int mdb_tls_nkeys;
2891 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2895 case DLL_PROCESS_ATTACH: break;
2896 case DLL_THREAD_ATTACH: break;
2897 case DLL_THREAD_DETACH:
2898 for (i=0; i<mdb_tls_nkeys; i++) {
2899 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2900 mdb_env_reader_dest(r);
2903 case DLL_PROCESS_DETACH: break;
2908 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2910 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2914 /* Force some symbol references.
2915 * _tls_used forces the linker to create the TLS directory if not already done
2916 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2918 #pragma comment(linker, "/INCLUDE:_tls_used")
2919 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2920 #pragma const_seg(".CRT$XLB")
2921 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2922 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2925 #pragma comment(linker, "/INCLUDE:__tls_used")
2926 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2927 #pragma data_seg(".CRT$XLB")
2928 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2930 #endif /* WIN 32/64 */
2931 #endif /* !__GNUC__ */
2934 /** Downgrade the exclusive lock on the region back to shared */
2936 mdb_env_share_locks(MDB_env *env, int *excl)
2938 int rc = 0, toggle = mdb_env_pick_meta(env);
2940 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2945 /* First acquire a shared lock. The Unlock will
2946 * then release the existing exclusive lock.
2948 memset(&ov, 0, sizeof(ov));
2949 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2952 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2958 struct flock lock_info;
2959 /* The shared lock replaces the existing lock */
2960 memset((void *)&lock_info, 0, sizeof(lock_info));
2961 lock_info.l_type = F_RDLCK;
2962 lock_info.l_whence = SEEK_SET;
2963 lock_info.l_start = 0;
2964 lock_info.l_len = 1;
2965 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2966 (rc = ErrCode()) == EINTR) ;
2967 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
2974 /** Try to get exlusive lock, otherwise shared.
2975 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
2978 mdb_env_excl_lock(MDB_env *env, int *excl)
2982 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2986 memset(&ov, 0, sizeof(ov));
2987 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2994 struct flock lock_info;
2995 memset((void *)&lock_info, 0, sizeof(lock_info));
2996 lock_info.l_type = F_WRLCK;
2997 lock_info.l_whence = SEEK_SET;
2998 lock_info.l_start = 0;
2999 lock_info.l_len = 1;
3000 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3001 (rc = ErrCode()) == EINTR) ;
3005 # ifdef MDB_USE_POSIX_SEM
3006 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3009 lock_info.l_type = F_RDLCK;
3010 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3011 (rc = ErrCode()) == EINTR) ;
3019 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3021 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3023 * @(#) $Revision: 5.1 $
3024 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3025 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3027 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3031 * Please do not copyright this code. This code is in the public domain.
3033 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3034 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3035 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3036 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3037 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3038 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3039 * PERFORMANCE OF THIS SOFTWARE.
3042 * chongo <Landon Curt Noll> /\oo/\
3043 * http://www.isthe.com/chongo/
3045 * Share and Enjoy! :-)
3048 typedef unsigned long long mdb_hash_t;
3049 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3051 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3052 * @param[in] str string to hash
3053 * @param[in] hval initial value for hash
3054 * @return 64 bit hash
3056 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3057 * hval arg on the first call.
3060 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3062 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3063 unsigned char *end = s + val->mv_size;
3065 * FNV-1a hash each octet of the string
3068 /* xor the bottom with the current octet */
3069 hval ^= (mdb_hash_t)*s++;
3071 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3072 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3073 (hval << 7) + (hval << 8) + (hval << 40);
3075 /* return our new hash value */
3079 /** Hash the string and output the hash in hex.
3080 * @param[in] str string to hash
3081 * @param[out] hexbuf an array of 17 chars to hold the hash
3084 mdb_hash_hex(MDB_val *val, char *hexbuf)
3087 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3088 for (i=0; i<8; i++) {
3089 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3095 /** Open and/or initialize the lock region for the environment.
3096 * @param[in] env The MDB environment.
3097 * @param[in] lpath The pathname of the file used for the lock region.
3098 * @param[in] mode The Unix permissions for the file, if we create it.
3099 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3100 * @return 0 on success, non-zero on failure.
3103 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3111 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3112 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3113 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
3115 if (rc == ERROR_WRITE_PROTECT && (env->me_flags & MDB_RDONLY)) {
3116 env->me_flags |= MDB_ROFS;
3121 /* Try to get exclusive lock. If we succeed, then
3122 * nobody is using the lock region and we should initialize it.
3124 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3125 size = GetFileSize(env->me_lfd, NULL);
3131 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
3133 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3134 env->me_flags |= MDB_ROFS;
3139 /* Lose record locks when exec*() */
3140 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3141 fcntl(env->me_lfd, F_SETFD, fdflags);
3143 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
3144 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1) {
3146 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3147 env->me_flags |= MDB_ROFS;
3154 /* Try to get exclusive lock. If we succeed, then
3155 * nobody is using the lock region and we should initialize it.
3157 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3159 size = lseek(env->me_lfd, 0, SEEK_END);
3161 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3162 if (size < rsize && *excl > 0) {
3164 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3165 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3167 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3171 size = rsize - sizeof(MDB_txninfo);
3172 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3177 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3179 if (!mh) goto fail_errno;
3180 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3182 if (!env->me_txns) goto fail_errno;
3184 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3186 if (m == MAP_FAILED) goto fail_errno;
3192 BY_HANDLE_FILE_INFORMATION stbuf;
3201 if (!mdb_sec_inited) {
3202 InitializeSecurityDescriptor(&mdb_null_sd,
3203 SECURITY_DESCRIPTOR_REVISION);
3204 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3205 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3206 mdb_all_sa.bInheritHandle = FALSE;
3207 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3210 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3211 idbuf.volume = stbuf.dwVolumeSerialNumber;
3212 idbuf.nhigh = stbuf.nFileIndexHigh;
3213 idbuf.nlow = stbuf.nFileIndexLow;
3214 val.mv_data = &idbuf;
3215 val.mv_size = sizeof(idbuf);
3216 mdb_hash_hex(&val, hexbuf);
3217 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3218 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3219 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3220 if (!env->me_rmutex) goto fail_errno;
3221 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3222 if (!env->me_wmutex) goto fail_errno;
3223 #elif defined(MDB_USE_POSIX_SEM)
3232 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3233 idbuf.dev = stbuf.st_dev;
3234 idbuf.ino = stbuf.st_ino;
3235 val.mv_data = &idbuf;
3236 val.mv_size = sizeof(idbuf);
3237 mdb_hash_hex(&val, hexbuf);
3238 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3239 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3240 /* Clean up after a previous run, if needed: Try to
3241 * remove both semaphores before doing anything else.
3243 sem_unlink(env->me_txns->mti_rmname);
3244 sem_unlink(env->me_txns->mti_wmname);
3245 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3246 O_CREAT|O_EXCL, mode, 1);
3247 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3248 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3249 O_CREAT|O_EXCL, mode, 1);
3250 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3251 #else /* MDB_USE_POSIX_SEM */
3252 pthread_mutexattr_t mattr;
3254 if ((rc = pthread_mutexattr_init(&mattr))
3255 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3256 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3257 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3259 pthread_mutexattr_destroy(&mattr);
3260 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3262 env->me_txns->mti_version = MDB_VERSION;
3263 env->me_txns->mti_magic = MDB_MAGIC;
3264 env->me_txns->mti_txnid = 0;
3265 env->me_txns->mti_numreaders = 0;
3268 if (env->me_txns->mti_magic != MDB_MAGIC) {
3269 DPUTS("lock region has invalid magic");
3273 if (env->me_txns->mti_version != MDB_VERSION) {
3274 DPRINTF("lock region is version %u, expected version %u",
3275 env->me_txns->mti_version, MDB_VERSION);
3276 rc = MDB_VERSION_MISMATCH;
3280 if (rc != EACCES && rc != EAGAIN) {
3284 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3285 if (!env->me_rmutex) goto fail_errno;
3286 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3287 if (!env->me_wmutex) goto fail_errno;
3288 #elif defined(MDB_USE_POSIX_SEM)
3289 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3290 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3291 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3292 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3303 /** The name of the lock file in the DB environment */
3304 #define LOCKNAME "/lock.mdb"
3305 /** The name of the data file in the DB environment */
3306 #define DATANAME "/data.mdb"
3307 /** The suffix of the lock file when no subdir is used */
3308 #define LOCKSUFF "-lock"
3309 /** Only a subset of the @ref mdb_env flags can be changed
3310 * at runtime. Changing other flags requires closing the
3311 * environment and re-opening it with the new flags.
3313 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3314 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP)
3317 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3319 int oflags, rc, len, excl;
3320 char *lpath, *dpath;
3322 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3326 if (flags & MDB_NOSUBDIR) {
3327 rc = len + sizeof(LOCKSUFF) + len + 1;
3329 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3334 if (flags & MDB_NOSUBDIR) {
3335 dpath = lpath + len + sizeof(LOCKSUFF);
3336 sprintf(lpath, "%s" LOCKSUFF, path);
3337 strcpy(dpath, path);
3339 dpath = lpath + len + sizeof(LOCKNAME);
3340 sprintf(lpath, "%s" LOCKNAME, path);
3341 sprintf(dpath, "%s" DATANAME, path);
3344 flags |= env->me_flags;
3345 /* silently ignore WRITEMAP if we're only getting read access */
3346 if (F_ISSET(flags, MDB_RDONLY|MDB_WRITEMAP))
3347 flags ^= MDB_WRITEMAP;
3348 env->me_flags = flags |= MDB_ENV_ACTIVE;
3350 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3355 if (F_ISSET(flags, MDB_RDONLY)) {
3356 oflags = GENERIC_READ;
3357 len = OPEN_EXISTING;
3359 oflags = GENERIC_READ|GENERIC_WRITE;
3362 mode = FILE_ATTRIBUTE_NORMAL;
3363 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3364 NULL, len, mode, NULL);
3366 if (F_ISSET(flags, MDB_RDONLY))
3369 oflags = O_RDWR | O_CREAT;
3371 env->me_fd = open(dpath, oflags, mode);
3373 if (env->me_fd == INVALID_HANDLE_VALUE) {
3378 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3379 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3380 env->me_mfd = env->me_fd;
3382 /* Synchronous fd for meta writes. Needed even with
3383 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3386 env->me_mfd = CreateFile(dpath, oflags,
3387 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3388 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3390 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3392 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3397 DPRINTF("opened dbenv %p", (void *) env);
3398 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3401 env->me_numdbs = 2; /* this notes that me_txkey was set */
3403 /* Windows TLS callbacks need help finding their TLS info. */
3404 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3405 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3412 rc = mdb_env_share_locks(env, &excl);
3416 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3417 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3418 env->me_path = strdup(path);
3419 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3425 mdb_env_close0(env, excl);
3431 /** Destroy resources from mdb_env_open() and clear our readers */
3433 mdb_env_close0(MDB_env *env, int excl)
3437 if (!(env->me_flags & MDB_ENV_ACTIVE))
3440 free(env->me_dbflags);
3444 if (env->me_numdbs) {
3445 pthread_key_delete(env->me_txkey);
3447 /* Delete our key from the global list */
3448 for (i=0; i<mdb_tls_nkeys; i++)
3449 if (mdb_tls_keys[i] == env->me_txkey) {
3450 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3458 munmap(env->me_map, env->me_mapsize);
3460 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3462 if (env->me_fd != INVALID_HANDLE_VALUE)
3465 pid_t pid = env->me_pid;
3466 /* Clearing readers is done in this function because
3467 * me_txkey with its destructor must be disabled first.
3469 for (i = env->me_numreaders; --i >= 0; )
3470 if (env->me_txns->mti_readers[i].mr_pid == pid)
3471 env->me_txns->mti_readers[i].mr_pid = 0;
3473 if (env->me_rmutex) {
3474 CloseHandle(env->me_rmutex);
3475 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3477 /* Windows automatically destroys the mutexes when
3478 * the last handle closes.
3480 #elif defined(MDB_USE_POSIX_SEM)
3481 if (env->me_rmutex != SEM_FAILED) {
3482 sem_close(env->me_rmutex);
3483 if (env->me_wmutex != SEM_FAILED)
3484 sem_close(env->me_wmutex);
3485 /* If we have the filelock: If we are the
3486 * only remaining user, clean up semaphores.
3489 mdb_env_excl_lock(env, &excl);
3491 sem_unlink(env->me_txns->mti_rmname);
3492 sem_unlink(env->me_txns->mti_wmname);
3496 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3498 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3501 /* Unlock the lockfile. Windows would have unlocked it
3502 * after closing anyway, but not necessarily at once.
3504 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3510 env->me_flags &= ~MDB_ENV_ACTIVE;
3514 mdb_env_copy(MDB_env *env, const char *path)
3516 MDB_txn *txn = NULL;
3520 HANDLE newfd = INVALID_HANDLE_VALUE;
3522 if (env->me_flags & MDB_NOSUBDIR) {
3523 lpath = (char *)path;
3526 len += sizeof(DATANAME);
3527 lpath = malloc(len);
3530 sprintf(lpath, "%s" DATANAME, path);
3533 /* The destination path must exist, but the destination file must not.
3534 * We don't want the OS to cache the writes, since the source data is
3535 * already in the OS cache.
3538 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3539 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3541 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3547 if (!(env->me_flags & MDB_NOSUBDIR))
3549 if (newfd == INVALID_HANDLE_VALUE) {
3554 #ifdef F_NOCACHE /* __APPLE__ */
3555 rc = fcntl(newfd, F_NOCACHE, 1);
3562 /* Do the lock/unlock of the reader mutex before starting the
3563 * write txn. Otherwise other read txns could block writers.
3565 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3569 if (!(env->me_flags & MDB_ROFS)) {
3570 /* We must start the actual read txn after blocking writers */
3571 mdb_txn_reset0(txn);
3573 /* Temporarily block writers until we snapshot the meta pages */
3576 rc = mdb_txn_renew0(txn);
3578 UNLOCK_MUTEX_W(env);
3583 wsize = env->me_psize * 2;
3587 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3588 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3591 rc = write(newfd, env->me_map, wsize);
3592 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3594 if (! (env->me_flags & MDB_ROFS))
3595 UNLOCK_MUTEX_W(env);
3600 ptr = env->me_map + wsize;
3601 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3602 #define MAX_WRITE 2147483648U
3606 if (wsize > MAX_WRITE)
3610 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3611 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3620 if (wsize > MAX_WRITE)
3624 wres = write(newfd, ptr, w2);
3625 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3634 if (newfd != INVALID_HANDLE_VALUE)
3641 mdb_env_close(MDB_env *env)
3648 VGMEMP_DESTROY(env);
3649 while ((dp = env->me_dpages) != NULL) {
3650 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3651 env->me_dpages = dp->mp_next;
3655 mdb_env_close0(env, 0);
3656 mdb_midl_free(env->me_free_pgs);
3660 /** Compare two items pointing at aligned size_t's */
3662 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3664 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3665 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3668 /** Compare two items pointing at aligned int's */
3670 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3672 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3673 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3676 /** Compare two items pointing at ints of unknown alignment.
3677 * Nodes and keys are guaranteed to be 2-byte aligned.
3680 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3682 #if BYTE_ORDER == LITTLE_ENDIAN
3683 unsigned short *u, *c;
3686 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3687 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3690 } while(!x && u > (unsigned short *)a->mv_data);
3693 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3697 /** Compare two items lexically */
3699 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3706 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3712 diff = memcmp(a->mv_data, b->mv_data, len);
3713 return diff ? diff : len_diff<0 ? -1 : len_diff;
3716 /** Compare two items in reverse byte order */
3718 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3720 const unsigned char *p1, *p2, *p1_lim;
3724 p1_lim = (const unsigned char *)a->mv_data;
3725 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3726 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3728 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3734 while (p1 > p1_lim) {
3735 diff = *--p1 - *--p2;
3739 return len_diff<0 ? -1 : len_diff;
3742 /** Search for key within a page, using binary search.
3743 * Returns the smallest entry larger or equal to the key.
3744 * If exactp is non-null, stores whether the found entry was an exact match
3745 * in *exactp (1 or 0).
3746 * Updates the cursor index with the index of the found entry.
3747 * If no entry larger or equal to the key is found, returns NULL.
3750 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3752 unsigned int i = 0, nkeys;
3755 MDB_page *mp = mc->mc_pg[mc->mc_top];
3756 MDB_node *node = NULL;
3761 nkeys = NUMKEYS(mp);
3766 COPY_PGNO(pgno, mp->mp_pgno);
3767 DPRINTF("searching %u keys in %s %spage %zu",
3768 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3775 low = IS_LEAF(mp) ? 0 : 1;
3777 cmp = mc->mc_dbx->md_cmp;
3779 /* Branch pages have no data, so if using integer keys,
3780 * alignment is guaranteed. Use faster mdb_cmp_int.
3782 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3783 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3790 nodekey.mv_size = mc->mc_db->md_pad;
3791 node = NODEPTR(mp, 0); /* fake */
3792 while (low <= high) {
3793 i = (low + high) >> 1;
3794 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3795 rc = cmp(key, &nodekey);
3796 DPRINTF("found leaf index %u [%s], rc = %i",
3797 i, DKEY(&nodekey), rc);
3806 while (low <= high) {
3807 i = (low + high) >> 1;
3809 node = NODEPTR(mp, i);
3810 nodekey.mv_size = NODEKSZ(node);
3811 nodekey.mv_data = NODEKEY(node);
3813 rc = cmp(key, &nodekey);
3816 DPRINTF("found leaf index %u [%s], rc = %i",
3817 i, DKEY(&nodekey), rc);
3819 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3820 i, DKEY(&nodekey), NODEPGNO(node), rc);
3831 if (rc > 0) { /* Found entry is less than the key. */
3832 i++; /* Skip to get the smallest entry larger than key. */
3834 node = NODEPTR(mp, i);
3837 *exactp = (rc == 0);
3838 /* store the key index */
3839 mc->mc_ki[mc->mc_top] = i;
3841 /* There is no entry larger or equal to the key. */
3844 /* nodeptr is fake for LEAF2 */
3850 mdb_cursor_adjust(MDB_cursor *mc, func)
3854 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3855 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3862 /** Pop a page off the top of the cursor's stack. */
3864 mdb_cursor_pop(MDB_cursor *mc)
3867 #ifndef MDB_DEBUG_SKIP
3868 MDB_page *top = mc->mc_pg[mc->mc_top];
3874 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3875 mc->mc_dbi, (void *) mc);
3879 /** Push a page onto the top of the cursor's stack. */
3881 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3883 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3884 mc->mc_dbi, (void *) mc);
3886 if (mc->mc_snum >= CURSOR_STACK) {
3887 assert(mc->mc_snum < CURSOR_STACK);
3888 return MDB_CURSOR_FULL;
3891 mc->mc_top = mc->mc_snum++;
3892 mc->mc_pg[mc->mc_top] = mp;
3893 mc->mc_ki[mc->mc_top] = 0;
3898 /** Find the address of the page corresponding to a given page number.
3899 * @param[in] txn the transaction for this access.
3900 * @param[in] pgno the page number for the page to retrieve.
3901 * @param[out] ret address of a pointer where the page's address will be stored.
3902 * @return 0 on success, non-zero on failure.
3905 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3909 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
3910 (txn->mt_env->me_flags & MDB_WRITEMAP)))
3914 MDB_ID2L dl = tx2->mt_u.dirty_list;
3916 unsigned x = mdb_mid2l_search(dl, pgno);
3917 if (x <= dl[0].mid && dl[x].mid == pgno) {
3922 } while ((tx2 = tx2->mt_parent) != NULL);
3925 if (pgno < txn->mt_next_pgno) {
3926 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3928 DPRINTF("page %zu not found", pgno);
3934 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3937 /** Search for the page a given key should be in.
3938 * Pushes parent pages on the cursor stack. This function continues a
3939 * search on a cursor that has already been initialized. (Usually by
3940 * #mdb_page_search() but also by #mdb_node_move().)
3941 * @param[in,out] mc the cursor for this operation.
3942 * @param[in] key the key to search for. If NULL, search for the lowest
3943 * page. (This is used by #mdb_cursor_first().)
3944 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3945 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3946 * @return 0 on success, non-zero on failure.
3949 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3951 MDB_page *mp = mc->mc_pg[mc->mc_top];
3956 while (IS_BRANCH(mp)) {
3960 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3961 assert(NUMKEYS(mp) > 1);
3962 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3964 if (key == NULL) /* Initialize cursor to first page. */
3966 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
3967 /* cursor to last page */
3971 node = mdb_node_search(mc, key, &exact);
3973 i = NUMKEYS(mp) - 1;
3975 i = mc->mc_ki[mc->mc_top];
3984 DPRINTF("following index %u for key [%s]",
3986 assert(i < NUMKEYS(mp));
3987 node = NODEPTR(mp, i);
3989 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3992 mc->mc_ki[mc->mc_top] = i;
3993 if ((rc = mdb_cursor_push(mc, mp)))
3997 if ((rc = mdb_page_touch(mc)) != 0)
3999 mp = mc->mc_pg[mc->mc_top];
4004 DPRINTF("internal error, index points to a %02X page!?",
4006 return MDB_CORRUPTED;
4009 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4010 key ? DKEY(key) : NULL);
4015 /** Search for the page a given key should be in.
4016 * Pushes parent pages on the cursor stack. This function just sets up
4017 * the search; it finds the root page for \b mc's database and sets this
4018 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4019 * called to complete the search.
4020 * @param[in,out] mc the cursor for this operation.
4021 * @param[in] key the key to search for. If NULL, search for the lowest
4022 * page. (This is used by #mdb_cursor_first().)
4023 * @param[in] modify If true, visited pages are updated with new page numbers.
4024 * @return 0 on success, non-zero on failure.
4027 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4032 /* Make sure the txn is still viable, then find the root from
4033 * the txn's db table.
4035 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4036 DPUTS("transaction has failed, must abort");
4039 /* Make sure we're using an up-to-date root */
4040 if (mc->mc_dbi > MAIN_DBI) {
4041 if ((*mc->mc_dbflag & DB_STALE) ||
4042 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4044 unsigned char dbflag = 0;
4045 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4046 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4049 if (*mc->mc_dbflag & DB_STALE) {
4052 MDB_node *leaf = mdb_node_search(&mc2,
4053 &mc->mc_dbx->md_name, &exact);
4055 return MDB_NOTFOUND;
4056 mdb_node_read(mc->mc_txn, leaf, &data);
4057 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4059 if (flags & MDB_PS_MODIFY)
4061 *mc->mc_dbflag = dbflag;
4064 root = mc->mc_db->md_root;
4066 if (root == P_INVALID) { /* Tree is empty. */
4067 DPUTS("tree is empty");
4068 return MDB_NOTFOUND;
4073 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4074 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
4080 DPRINTF("db %u root page %zu has flags 0x%X",
4081 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4083 if (flags & MDB_PS_MODIFY) {
4084 if ((rc = mdb_page_touch(mc)))
4088 if (flags & MDB_PS_ROOTONLY)
4091 return mdb_page_search_root(mc, key, flags);
4094 /** Return the data associated with a given node.
4095 * @param[in] txn The transaction for this operation.
4096 * @param[in] leaf The node being read.
4097 * @param[out] data Updated to point to the node's data.
4098 * @return 0 on success, non-zero on failure.
4101 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4103 MDB_page *omp; /* overflow page */
4107 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4108 data->mv_size = NODEDSZ(leaf);
4109 data->mv_data = NODEDATA(leaf);
4113 /* Read overflow data.
4115 data->mv_size = NODEDSZ(leaf);
4116 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4117 if ((rc = mdb_page_get(txn, pgno, &omp))) {
4118 DPRINTF("read overflow page %zu failed", pgno);
4121 data->mv_data = METADATA(omp);
4127 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4128 MDB_val *key, MDB_val *data)
4137 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4139 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4142 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4146 mdb_cursor_init(&mc, txn, dbi, &mx);
4147 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4150 /** Find a sibling for a page.
4151 * Replaces the page at the top of the cursor's stack with the
4152 * specified sibling, if one exists.
4153 * @param[in] mc The cursor for this operation.
4154 * @param[in] move_right Non-zero if the right sibling is requested,
4155 * otherwise the left sibling.
4156 * @return 0 on success, non-zero on failure.
4159 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4165 if (mc->mc_snum < 2) {
4166 return MDB_NOTFOUND; /* root has no siblings */
4170 DPRINTF("parent page is page %zu, index %u",
4171 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4173 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4174 : (mc->mc_ki[mc->mc_top] == 0)) {
4175 DPRINTF("no more keys left, moving to %s sibling",
4176 move_right ? "right" : "left");
4177 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4178 /* undo cursor_pop before returning */
4185 mc->mc_ki[mc->mc_top]++;
4187 mc->mc_ki[mc->mc_top]--;
4188 DPRINTF("just moving to %s index key %u",
4189 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4191 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4193 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4194 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4197 mdb_cursor_push(mc, mp);
4202 /** Move the cursor to the next data item. */
4204 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4210 if (mc->mc_flags & C_EOF) {
4211 return MDB_NOTFOUND;
4214 assert(mc->mc_flags & C_INITIALIZED);
4216 mp = mc->mc_pg[mc->mc_top];
4218 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4219 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4220 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4221 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4222 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4223 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4227 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4228 if (op == MDB_NEXT_DUP)
4229 return MDB_NOTFOUND;
4233 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4235 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4236 DPUTS("=====> move to next sibling page");
4237 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4238 mc->mc_flags |= C_EOF;
4239 mc->mc_flags &= ~C_INITIALIZED;
4240 return MDB_NOTFOUND;
4242 mp = mc->mc_pg[mc->mc_top];
4243 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4245 mc->mc_ki[mc->mc_top]++;
4247 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4248 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4251 key->mv_size = mc->mc_db->md_pad;
4252 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4256 assert(IS_LEAF(mp));
4257 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4259 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4260 mdb_xcursor_init1(mc, leaf);
4263 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4266 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4267 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4268 if (rc != MDB_SUCCESS)
4273 MDB_GET_KEY(leaf, key);
4277 /** Move the cursor to the previous data item. */
4279 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4285 assert(mc->mc_flags & C_INITIALIZED);
4287 mp = mc->mc_pg[mc->mc_top];
4289 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4290 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4291 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4292 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4293 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4294 if (op != MDB_PREV || rc == MDB_SUCCESS)
4297 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4298 if (op == MDB_PREV_DUP)
4299 return MDB_NOTFOUND;
4304 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4306 if (mc->mc_ki[mc->mc_top] == 0) {
4307 DPUTS("=====> move to prev sibling page");
4308 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4309 mc->mc_flags &= ~C_INITIALIZED;
4310 return MDB_NOTFOUND;
4312 mp = mc->mc_pg[mc->mc_top];
4313 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4314 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4316 mc->mc_ki[mc->mc_top]--;
4318 mc->mc_flags &= ~C_EOF;
4320 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4321 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4324 key->mv_size = mc->mc_db->md_pad;
4325 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4329 assert(IS_LEAF(mp));
4330 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4332 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4333 mdb_xcursor_init1(mc, leaf);
4336 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4339 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4340 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4341 if (rc != MDB_SUCCESS)
4346 MDB_GET_KEY(leaf, key);
4350 /** Set the cursor on a specific data item. */
4352 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4353 MDB_cursor_op op, int *exactp)
4357 MDB_node *leaf = NULL;
4362 assert(key->mv_size > 0);
4364 /* See if we're already on the right page */
4365 if (mc->mc_flags & C_INITIALIZED) {
4368 mp = mc->mc_pg[mc->mc_top];
4370 mc->mc_ki[mc->mc_top] = 0;
4371 return MDB_NOTFOUND;
4373 if (mp->mp_flags & P_LEAF2) {
4374 nodekey.mv_size = mc->mc_db->md_pad;
4375 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4377 leaf = NODEPTR(mp, 0);
4378 MDB_GET_KEY(leaf, &nodekey);
4380 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4382 /* Probably happens rarely, but first node on the page
4383 * was the one we wanted.
4385 mc->mc_ki[mc->mc_top] = 0;
4392 unsigned int nkeys = NUMKEYS(mp);
4394 if (mp->mp_flags & P_LEAF2) {
4395 nodekey.mv_data = LEAF2KEY(mp,
4396 nkeys-1, nodekey.mv_size);
4398 leaf = NODEPTR(mp, nkeys-1);
4399 MDB_GET_KEY(leaf, &nodekey);
4401 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4403 /* last node was the one we wanted */
4404 mc->mc_ki[mc->mc_top] = nkeys-1;
4410 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4411 /* This is definitely the right page, skip search_page */
4412 if (mp->mp_flags & P_LEAF2) {
4413 nodekey.mv_data = LEAF2KEY(mp,
4414 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4416 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4417 MDB_GET_KEY(leaf, &nodekey);
4419 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4421 /* current node was the one we wanted */
4431 /* If any parents have right-sibs, search.
4432 * Otherwise, there's nothing further.
4434 for (i=0; i<mc->mc_top; i++)
4436 NUMKEYS(mc->mc_pg[i])-1)
4438 if (i == mc->mc_top) {
4439 /* There are no other pages */
4440 mc->mc_ki[mc->mc_top] = nkeys;
4441 return MDB_NOTFOUND;
4445 /* There are no other pages */
4446 mc->mc_ki[mc->mc_top] = 0;
4447 return MDB_NOTFOUND;
4451 rc = mdb_page_search(mc, key, 0);
4452 if (rc != MDB_SUCCESS)
4455 mp = mc->mc_pg[mc->mc_top];
4456 assert(IS_LEAF(mp));
4459 leaf = mdb_node_search(mc, key, exactp);
4460 if (exactp != NULL && !*exactp) {
4461 /* MDB_SET specified and not an exact match. */
4462 return MDB_NOTFOUND;
4466 DPUTS("===> inexact leaf not found, goto sibling");
4467 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4468 return rc; /* no entries matched */
4469 mp = mc->mc_pg[mc->mc_top];
4470 assert(IS_LEAF(mp));
4471 leaf = NODEPTR(mp, 0);
4475 mc->mc_flags |= C_INITIALIZED;
4476 mc->mc_flags &= ~C_EOF;
4479 key->mv_size = mc->mc_db->md_pad;
4480 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4484 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4485 mdb_xcursor_init1(mc, leaf);
4488 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4489 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4490 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4493 if (op == MDB_GET_BOTH) {
4499 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4500 if (rc != MDB_SUCCESS)
4503 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4505 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4507 rc = mc->mc_dbx->md_dcmp(data, &d2);
4509 if (op == MDB_GET_BOTH || rc > 0)
4510 return MDB_NOTFOUND;
4515 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4516 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4521 /* The key already matches in all other cases */
4522 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4523 MDB_GET_KEY(leaf, key);
4524 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4529 /** Move the cursor to the first item in the database. */
4531 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4536 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4537 rc = mdb_page_search(mc, NULL, 0);
4538 if (rc != MDB_SUCCESS)
4541 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4543 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4544 mc->mc_flags |= C_INITIALIZED;
4545 mc->mc_flags &= ~C_EOF;
4547 mc->mc_ki[mc->mc_top] = 0;
4549 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4550 key->mv_size = mc->mc_db->md_pad;
4551 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4556 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4557 mdb_xcursor_init1(mc, leaf);
4558 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4563 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4564 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4568 MDB_GET_KEY(leaf, key);
4572 /** Move the cursor to the last item in the database. */
4574 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4579 if (!(mc->mc_flags & C_EOF)) {
4581 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4584 lkey.mv_size = MDB_MAXKEYSIZE+1;
4585 lkey.mv_data = NULL;
4586 rc = mdb_page_search(mc, &lkey, 0);
4587 if (rc != MDB_SUCCESS)
4590 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4592 mc->mc_flags |= C_INITIALIZED|C_EOF;
4593 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4595 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4597 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4598 key->mv_size = mc->mc_db->md_pad;
4599 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4604 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4605 mdb_xcursor_init1(mc, leaf);
4606 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4611 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4612 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4617 MDB_GET_KEY(leaf, key);
4622 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4631 case MDB_GET_CURRENT:
4632 if (!(mc->mc_flags & C_INITIALIZED)) {
4635 MDB_page *mp = mc->mc_pg[mc->mc_top];
4637 mc->mc_ki[mc->mc_top] = 0;
4643 key->mv_size = mc->mc_db->md_pad;
4644 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4646 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4647 MDB_GET_KEY(leaf, key);
4649 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4650 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4652 rc = mdb_node_read(mc->mc_txn, leaf, data);
4659 case MDB_GET_BOTH_RANGE:
4660 if (data == NULL || mc->mc_xcursor == NULL) {
4668 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4670 } else if (op == MDB_SET_RANGE)
4671 rc = mdb_cursor_set(mc, key, data, op, NULL);
4673 rc = mdb_cursor_set(mc, key, data, op, &exact);
4675 case MDB_GET_MULTIPLE:
4677 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4678 !(mc->mc_flags & C_INITIALIZED)) {
4683 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4684 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4687 case MDB_NEXT_MULTIPLE:
4689 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4693 if (!(mc->mc_flags & C_INITIALIZED))
4694 rc = mdb_cursor_first(mc, key, data);
4696 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4697 if (rc == MDB_SUCCESS) {
4698 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4701 mx = &mc->mc_xcursor->mx_cursor;
4702 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4704 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4705 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4713 case MDB_NEXT_NODUP:
4714 if (!(mc->mc_flags & C_INITIALIZED))
4715 rc = mdb_cursor_first(mc, key, data);
4717 rc = mdb_cursor_next(mc, key, data, op);
4721 case MDB_PREV_NODUP:
4722 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4723 rc = mdb_cursor_last(mc, key, data);
4724 mc->mc_flags |= C_INITIALIZED;
4725 mc->mc_ki[mc->mc_top]++;
4727 rc = mdb_cursor_prev(mc, key, data, op);
4730 rc = mdb_cursor_first(mc, key, data);
4734 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4735 !(mc->mc_flags & C_INITIALIZED) ||
4736 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4740 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4743 rc = mdb_cursor_last(mc, key, data);
4747 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4748 !(mc->mc_flags & C_INITIALIZED) ||
4749 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4753 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4756 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4764 /** Touch all the pages in the cursor stack.
4765 * Makes sure all the pages are writable, before attempting a write operation.
4766 * @param[in] mc The cursor to operate on.
4769 mdb_cursor_touch(MDB_cursor *mc)
4773 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4776 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4777 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4778 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4781 *mc->mc_dbflag = DB_DIRTY;
4783 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4784 rc = mdb_page_touch(mc);
4788 mc->mc_top = mc->mc_snum-1;
4793 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4796 MDB_node *leaf = NULL;
4797 MDB_val xdata, *rdata, dkey;
4800 int do_sub = 0, insert = 0;
4801 unsigned int mcount = 0;
4805 char dbuf[MDB_MAXKEYSIZE+1];
4806 unsigned int nflags;
4809 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4812 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4815 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4818 #if SIZE_MAX > MAXDATASIZE
4819 if (data->mv_size > MAXDATASIZE)
4823 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4824 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4828 if (flags == MDB_CURRENT) {
4829 if (!(mc->mc_flags & C_INITIALIZED))
4832 } else if (mc->mc_db->md_root == P_INVALID) {
4834 /* new database, write a root leaf page */
4835 DPUTS("allocating new root leaf page");
4836 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4840 mdb_cursor_push(mc, np);
4841 mc->mc_db->md_root = np->mp_pgno;
4842 mc->mc_db->md_depth++;
4843 *mc->mc_dbflag = DB_DIRTY;
4844 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4846 np->mp_flags |= P_LEAF2;
4847 mc->mc_flags |= C_INITIALIZED;
4853 if (flags & MDB_APPEND) {
4855 rc = mdb_cursor_last(mc, &k2, &d2);
4857 rc = mc->mc_dbx->md_cmp(key, &k2);
4860 mc->mc_ki[mc->mc_top]++;
4866 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4868 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4869 DPRINTF("duplicate key [%s]", DKEY(key));
4871 return MDB_KEYEXIST;
4873 if (rc && rc != MDB_NOTFOUND)
4877 /* Cursor is positioned, now make sure all pages are writable */
4878 rc2 = mdb_cursor_touch(mc);
4883 /* The key already exists */
4884 if (rc == MDB_SUCCESS) {
4885 /* there's only a key anyway, so this is a no-op */
4886 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4887 unsigned int ksize = mc->mc_db->md_pad;
4888 if (key->mv_size != ksize)
4890 if (flags == MDB_CURRENT) {
4891 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4892 memcpy(ptr, key->mv_data, ksize);
4897 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4900 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4901 /* Was a single item before, must convert now */
4903 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4904 /* Just overwrite the current item */
4905 if (flags == MDB_CURRENT)
4908 dkey.mv_size = NODEDSZ(leaf);
4909 dkey.mv_data = NODEDATA(leaf);
4910 #if UINT_MAX < SIZE_MAX
4911 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4912 #ifdef MISALIGNED_OK
4913 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4915 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4918 /* if data matches, ignore it */
4919 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4920 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4922 /* create a fake page for the dup items */
4923 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4924 dkey.mv_data = dbuf;
4925 fp = (MDB_page *)&pbuf;
4926 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4927 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4928 fp->mp_lower = PAGEHDRSZ;
4929 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4930 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4931 fp->mp_flags |= P_LEAF2;
4932 fp->mp_pad = data->mv_size;
4933 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4935 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4936 (dkey.mv_size & 1) + (data->mv_size & 1);
4938 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4941 xdata.mv_size = fp->mp_upper;
4946 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4947 /* See if we need to convert from fake page to subDB */
4949 unsigned int offset;
4952 fp = NODEDATA(leaf);
4953 if (flags == MDB_CURRENT) {
4955 fp->mp_flags |= P_DIRTY;
4956 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4957 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4961 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4962 offset = fp->mp_pad;
4963 if (SIZELEFT(fp) >= offset)
4965 offset *= 4; /* space for 4 more */
4967 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4969 offset += offset & 1;
4970 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4971 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4973 /* yes, convert it */
4975 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4976 dummy.md_pad = fp->mp_pad;
4977 dummy.md_flags = MDB_DUPFIXED;
4978 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4979 dummy.md_flags |= MDB_INTEGERKEY;
4982 dummy.md_branch_pages = 0;
4983 dummy.md_leaf_pages = 1;
4984 dummy.md_overflow_pages = 0;
4985 dummy.md_entries = NUMKEYS(fp);
4987 xdata.mv_size = sizeof(MDB_db);
4988 xdata.mv_data = &dummy;
4989 if ((rc = mdb_page_alloc(mc, 1, &mp)))
4991 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4992 flags |= F_DUPDATA|F_SUBDATA;
4993 dummy.md_root = mp->mp_pgno;
4995 /* no, just grow it */
4997 xdata.mv_size = NODEDSZ(leaf) + offset;
4998 xdata.mv_data = &pbuf;
4999 mp = (MDB_page *)&pbuf;
5000 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5003 mp->mp_flags = fp->mp_flags | P_DIRTY;
5004 mp->mp_pad = fp->mp_pad;
5005 mp->mp_lower = fp->mp_lower;
5006 mp->mp_upper = fp->mp_upper + offset;
5008 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5010 nsize = NODEDSZ(leaf) - fp->mp_upper;
5011 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5012 for (i=0; i<NUMKEYS(fp); i++)
5013 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5015 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5019 /* data is on sub-DB, just store it */
5020 flags |= F_DUPDATA|F_SUBDATA;
5024 /* overflow page overwrites need special handling */
5025 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5028 int ovpages, dpages;
5030 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5031 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5032 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5033 mdb_page_get(mc->mc_txn, pg, &omp);
5034 /* Is the ov page writable and large enough? */
5035 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5036 /* yes, overwrite it. Note in this case we don't
5037 * bother to try shrinking the node if the new data
5038 * is smaller than the overflow threshold.
5040 if (F_ISSET(flags, MDB_RESERVE))
5041 data->mv_data = METADATA(omp);
5043 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5046 /* no, free ovpages */
5048 mc->mc_db->md_overflow_pages -= ovpages;
5049 for (i=0; i<ovpages; i++) {
5050 DPRINTF("freed ov page %zu", pg);
5051 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5055 } else if (NODEDSZ(leaf) == data->mv_size) {
5056 /* same size, just replace it. Note that we could
5057 * also reuse this node if the new data is smaller,
5058 * but instead we opt to shrink the node in that case.
5060 if (F_ISSET(flags, MDB_RESERVE))
5061 data->mv_data = NODEDATA(leaf);
5062 else if (data->mv_size)
5063 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5065 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5068 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5069 mc->mc_db->md_entries--;
5071 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5078 nflags = flags & NODE_ADD_FLAGS;
5079 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5080 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5081 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5082 nflags &= ~MDB_APPEND;
5084 nflags |= MDB_SPLIT_REPLACE;
5085 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5087 /* There is room already in this leaf page. */
5088 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5089 if (rc == 0 && !do_sub && insert) {
5090 /* Adjust other cursors pointing to mp */
5091 MDB_cursor *m2, *m3;
5092 MDB_dbi dbi = mc->mc_dbi;
5093 unsigned i = mc->mc_top;
5094 MDB_page *mp = mc->mc_pg[i];
5096 if (mc->mc_flags & C_SUB)
5099 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5100 if (mc->mc_flags & C_SUB)
5101 m3 = &m2->mc_xcursor->mx_cursor;
5104 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5105 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5112 if (rc != MDB_SUCCESS)
5113 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5115 /* Now store the actual data in the child DB. Note that we're
5116 * storing the user data in the keys field, so there are strict
5117 * size limits on dupdata. The actual data fields of the child
5118 * DB are all zero size.
5125 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5126 if (flags & MDB_CURRENT) {
5127 xflags = MDB_CURRENT;
5129 mdb_xcursor_init1(mc, leaf);
5130 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5132 /* converted, write the original data first */
5134 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5138 /* Adjust other cursors pointing to mp */
5140 unsigned i = mc->mc_top;
5141 MDB_page *mp = mc->mc_pg[i];
5143 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5144 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5145 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5146 mdb_xcursor_init1(m2, leaf);
5151 if (flags & MDB_APPENDDUP)
5152 xflags |= MDB_APPEND;
5153 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5154 if (flags & F_SUBDATA) {
5155 void *db = NODEDATA(leaf);
5156 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5159 /* sub-writes might have failed so check rc again.
5160 * Don't increment count if we just replaced an existing item.
5162 if (!rc && !(flags & MDB_CURRENT))
5163 mc->mc_db->md_entries++;
5164 if (flags & MDB_MULTIPLE) {
5166 if (mcount < data[1].mv_size) {
5167 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5168 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5174 /* If we succeeded and the key didn't exist before, make sure
5175 * the cursor is marked valid.
5178 mc->mc_flags |= C_INITIALIZED;
5183 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5188 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5191 if (!(mc->mc_flags & C_INITIALIZED))
5194 rc = mdb_cursor_touch(mc);
5198 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5200 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5201 if (flags != MDB_NODUPDATA) {
5202 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5203 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5205 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5206 /* If sub-DB still has entries, we're done */
5207 if (mc->mc_xcursor->mx_db.md_entries) {
5208 if (leaf->mn_flags & F_SUBDATA) {
5209 /* update subDB info */
5210 void *db = NODEDATA(leaf);
5211 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5213 /* shrink fake page */
5214 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5216 mc->mc_db->md_entries--;
5219 /* otherwise fall thru and delete the sub-DB */
5222 if (leaf->mn_flags & F_SUBDATA) {
5223 /* add all the child DB's pages to the free list */
5224 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5225 if (rc == MDB_SUCCESS) {
5226 mc->mc_db->md_entries -=
5227 mc->mc_xcursor->mx_db.md_entries;
5232 return mdb_cursor_del0(mc, leaf);
5235 /** Allocate and initialize new pages for a database.
5236 * @param[in] mc a cursor on the database being added to.
5237 * @param[in] flags flags defining what type of page is being allocated.
5238 * @param[in] num the number of pages to allocate. This is usually 1,
5239 * unless allocating overflow pages for a large record.
5240 * @param[out] mp Address of a page, or NULL on failure.
5241 * @return 0 on success, non-zero on failure.
5244 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5249 if ((rc = mdb_page_alloc(mc, num, &np)))
5251 DPRINTF("allocated new mpage %zu, page size %u",
5252 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5253 np->mp_flags = flags | P_DIRTY;
5254 np->mp_lower = PAGEHDRSZ;
5255 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5258 mc->mc_db->md_branch_pages++;
5259 else if (IS_LEAF(np))
5260 mc->mc_db->md_leaf_pages++;
5261 else if (IS_OVERFLOW(np)) {
5262 mc->mc_db->md_overflow_pages += num;
5270 /** Calculate the size of a leaf node.
5271 * The size depends on the environment's page size; if a data item
5272 * is too large it will be put onto an overflow page and the node
5273 * size will only include the key and not the data. Sizes are always
5274 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5275 * of the #MDB_node headers.
5276 * @param[in] env The environment handle.
5277 * @param[in] key The key for the node.
5278 * @param[in] data The data for the node.
5279 * @return The number of bytes needed to store the node.
5282 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5286 sz = LEAFSIZE(key, data);
5287 if (sz >= env->me_psize / MDB_MINKEYS) {
5288 /* put on overflow page */
5289 sz -= data->mv_size - sizeof(pgno_t);
5293 return sz + sizeof(indx_t);
5296 /** Calculate the size of a branch node.
5297 * The size should depend on the environment's page size but since
5298 * we currently don't support spilling large keys onto overflow
5299 * pages, it's simply the size of the #MDB_node header plus the
5300 * size of the key. Sizes are always rounded up to an even number
5301 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5302 * @param[in] env The environment handle.
5303 * @param[in] key The key for the node.
5304 * @return The number of bytes needed to store the node.
5307 mdb_branch_size(MDB_env *env, MDB_val *key)
5312 if (sz >= env->me_psize / MDB_MINKEYS) {
5313 /* put on overflow page */
5314 /* not implemented */
5315 /* sz -= key->size - sizeof(pgno_t); */
5318 return sz + sizeof(indx_t);
5321 /** Add a node to the page pointed to by the cursor.
5322 * @param[in] mc The cursor for this operation.
5323 * @param[in] indx The index on the page where the new node should be added.
5324 * @param[in] key The key for the new node.
5325 * @param[in] data The data for the new node, if any.
5326 * @param[in] pgno The page number, if adding a branch node.
5327 * @param[in] flags Flags for the node.
5328 * @return 0 on success, non-zero on failure. Possible errors are:
5330 * <li>ENOMEM - failed to allocate overflow pages for the node.
5331 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5332 * should never happen since all callers already calculate the
5333 * page's free space before calling this function.
5337 mdb_node_add(MDB_cursor *mc, indx_t indx,
5338 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5341 size_t node_size = NODESIZE;
5344 MDB_page *mp = mc->mc_pg[mc->mc_top];
5345 MDB_page *ofp = NULL; /* overflow page */
5348 assert(mp->mp_upper >= mp->mp_lower);
5350 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5351 IS_LEAF(mp) ? "leaf" : "branch",
5352 IS_SUBP(mp) ? "sub-" : "",
5353 mp->mp_pgno, indx, data ? data->mv_size : 0,
5354 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5357 /* Move higher keys up one slot. */
5358 int ksize = mc->mc_db->md_pad, dif;
5359 char *ptr = LEAF2KEY(mp, indx, ksize);
5360 dif = NUMKEYS(mp) - indx;
5362 memmove(ptr+ksize, ptr, dif*ksize);
5363 /* insert new key */
5364 memcpy(ptr, key->mv_data, ksize);
5366 /* Just using these for counting */
5367 mp->mp_lower += sizeof(indx_t);
5368 mp->mp_upper -= ksize - sizeof(indx_t);
5373 node_size += key->mv_size;
5377 if (F_ISSET(flags, F_BIGDATA)) {
5378 /* Data already on overflow page. */
5379 node_size += sizeof(pgno_t);
5380 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5381 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5383 /* Put data on overflow page. */
5384 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5385 data->mv_size, node_size+data->mv_size);
5386 node_size += sizeof(pgno_t);
5387 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5389 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5392 node_size += data->mv_size;
5395 node_size += node_size & 1;
5397 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5398 DPRINTF("not enough room in page %zu, got %u ptrs",
5399 mp->mp_pgno, NUMKEYS(mp));
5400 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5401 mp->mp_upper - mp->mp_lower);
5402 DPRINTF("node size = %zu", node_size);
5403 return MDB_PAGE_FULL;
5406 /* Move higher pointers up one slot. */
5407 for (i = NUMKEYS(mp); i > indx; i--)
5408 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5410 /* Adjust free space offsets. */
5411 ofs = mp->mp_upper - node_size;
5412 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5413 mp->mp_ptrs[indx] = ofs;
5415 mp->mp_lower += sizeof(indx_t);
5417 /* Write the node data. */
5418 node = NODEPTR(mp, indx);
5419 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5420 node->mn_flags = flags;
5422 SETDSZ(node,data->mv_size);
5427 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5432 if (F_ISSET(flags, F_BIGDATA))
5433 memcpy(node->mn_data + key->mv_size, data->mv_data,
5435 else if (F_ISSET(flags, MDB_RESERVE))
5436 data->mv_data = node->mn_data + key->mv_size;
5438 memcpy(node->mn_data + key->mv_size, data->mv_data,
5441 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5443 if (F_ISSET(flags, MDB_RESERVE))
5444 data->mv_data = METADATA(ofp);
5446 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5453 /** Delete the specified node from a page.
5454 * @param[in] mp The page to operate on.
5455 * @param[in] indx The index of the node to delete.
5456 * @param[in] ksize The size of a node. Only used if the page is
5457 * part of a #MDB_DUPFIXED database.
5460 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5463 indx_t i, j, numkeys, ptr;
5470 COPY_PGNO(pgno, mp->mp_pgno);
5471 DPRINTF("delete node %u on %s page %zu", indx,
5472 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5475 assert(indx < NUMKEYS(mp));
5478 int x = NUMKEYS(mp) - 1 - indx;
5479 base = LEAF2KEY(mp, indx, ksize);
5481 memmove(base, base + ksize, x * ksize);
5482 mp->mp_lower -= sizeof(indx_t);
5483 mp->mp_upper += ksize - sizeof(indx_t);
5487 node = NODEPTR(mp, indx);
5488 sz = NODESIZE + node->mn_ksize;
5490 if (F_ISSET(node->mn_flags, F_BIGDATA))
5491 sz += sizeof(pgno_t);
5493 sz += NODEDSZ(node);
5497 ptr = mp->mp_ptrs[indx];
5498 numkeys = NUMKEYS(mp);
5499 for (i = j = 0; i < numkeys; i++) {
5501 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5502 if (mp->mp_ptrs[i] < ptr)
5503 mp->mp_ptrs[j] += sz;
5508 base = (char *)mp + mp->mp_upper;
5509 memmove(base + sz, base, ptr - mp->mp_upper);
5511 mp->mp_lower -= sizeof(indx_t);
5515 /** Compact the main page after deleting a node on a subpage.
5516 * @param[in] mp The main page to operate on.
5517 * @param[in] indx The index of the subpage on the main page.
5520 mdb_node_shrink(MDB_page *mp, indx_t indx)
5527 indx_t i, numkeys, ptr;
5529 node = NODEPTR(mp, indx);
5530 sp = (MDB_page *)NODEDATA(node);
5531 osize = NODEDSZ(node);
5533 delta = sp->mp_upper - sp->mp_lower;
5534 SETDSZ(node, osize - delta);
5535 xp = (MDB_page *)((char *)sp + delta);
5537 /* shift subpage upward */
5539 nsize = NUMKEYS(sp) * sp->mp_pad;
5540 memmove(METADATA(xp), METADATA(sp), nsize);
5543 nsize = osize - sp->mp_upper;
5544 numkeys = NUMKEYS(sp);
5545 for (i=numkeys-1; i>=0; i--)
5546 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5548 xp->mp_upper = sp->mp_lower;
5549 xp->mp_lower = sp->mp_lower;
5550 xp->mp_flags = sp->mp_flags;
5551 xp->mp_pad = sp->mp_pad;
5552 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5554 /* shift lower nodes upward */
5555 ptr = mp->mp_ptrs[indx];
5556 numkeys = NUMKEYS(mp);
5557 for (i = 0; i < numkeys; i++) {
5558 if (mp->mp_ptrs[i] <= ptr)
5559 mp->mp_ptrs[i] += delta;
5562 base = (char *)mp + mp->mp_upper;
5563 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5564 mp->mp_upper += delta;
5567 /** Initial setup of a sorted-dups cursor.
5568 * Sorted duplicates are implemented as a sub-database for the given key.
5569 * The duplicate data items are actually keys of the sub-database.
5570 * Operations on the duplicate data items are performed using a sub-cursor
5571 * initialized when the sub-database is first accessed. This function does
5572 * the preliminary setup of the sub-cursor, filling in the fields that
5573 * depend only on the parent DB.
5574 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5577 mdb_xcursor_init0(MDB_cursor *mc)
5579 MDB_xcursor *mx = mc->mc_xcursor;
5581 mx->mx_cursor.mc_xcursor = NULL;
5582 mx->mx_cursor.mc_txn = mc->mc_txn;
5583 mx->mx_cursor.mc_db = &mx->mx_db;
5584 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5585 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5586 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5587 mx->mx_cursor.mc_snum = 0;
5588 mx->mx_cursor.mc_top = 0;
5589 mx->mx_cursor.mc_flags = C_SUB;
5590 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5591 mx->mx_dbx.md_dcmp = NULL;
5592 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5595 /** Final setup of a sorted-dups cursor.
5596 * Sets up the fields that depend on the data from the main cursor.
5597 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5598 * @param[in] node The data containing the #MDB_db record for the
5599 * sorted-dup database.
5602 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5604 MDB_xcursor *mx = mc->mc_xcursor;
5606 if (node->mn_flags & F_SUBDATA) {
5607 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5608 mx->mx_cursor.mc_pg[0] = 0;
5609 mx->mx_cursor.mc_snum = 0;
5610 mx->mx_cursor.mc_flags = C_SUB;
5612 MDB_page *fp = NODEDATA(node);
5613 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5614 mx->mx_db.md_flags = 0;
5615 mx->mx_db.md_depth = 1;
5616 mx->mx_db.md_branch_pages = 0;
5617 mx->mx_db.md_leaf_pages = 1;
5618 mx->mx_db.md_overflow_pages = 0;
5619 mx->mx_db.md_entries = NUMKEYS(fp);
5620 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5621 mx->mx_cursor.mc_snum = 1;
5622 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5623 mx->mx_cursor.mc_top = 0;
5624 mx->mx_cursor.mc_pg[0] = fp;
5625 mx->mx_cursor.mc_ki[0] = 0;
5626 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5627 mx->mx_db.md_flags = MDB_DUPFIXED;
5628 mx->mx_db.md_pad = fp->mp_pad;
5629 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5630 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5633 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5635 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5637 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5638 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5639 #if UINT_MAX < SIZE_MAX
5640 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5641 #ifdef MISALIGNED_OK
5642 mx->mx_dbx.md_cmp = mdb_cmp_long;
5644 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5649 /** Initialize a cursor for a given transaction and database. */
5651 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5656 mc->mc_db = &txn->mt_dbs[dbi];
5657 mc->mc_dbx = &txn->mt_dbxs[dbi];
5658 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5663 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5665 mc->mc_xcursor = mx;
5666 mdb_xcursor_init0(mc);
5668 mc->mc_xcursor = NULL;
5670 if (*mc->mc_dbflag & DB_STALE) {
5671 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5676 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5679 MDB_xcursor *mx = NULL;
5680 size_t size = sizeof(MDB_cursor);
5682 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5685 /* Allow read access to the freelist */
5686 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5689 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5690 size += sizeof(MDB_xcursor);
5692 if ((mc = malloc(size)) != NULL) {
5693 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5694 mx = (MDB_xcursor *)(mc + 1);
5696 mdb_cursor_init(mc, txn, dbi, mx);
5697 if (txn->mt_cursors) {
5698 mc->mc_next = txn->mt_cursors[dbi];
5699 txn->mt_cursors[dbi] = mc;
5701 mc->mc_flags |= C_ALLOCD;
5712 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5714 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5717 if (txn->mt_cursors)
5720 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5724 /* Return the count of duplicate data items for the current key */
5726 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5730 if (mc == NULL || countp == NULL)
5733 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5736 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5737 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5740 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5743 *countp = mc->mc_xcursor->mx_db.md_entries;
5749 mdb_cursor_close(MDB_cursor *mc)
5752 /* remove from txn, if tracked */
5753 if (mc->mc_txn->mt_cursors) {
5754 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5755 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5757 *prev = mc->mc_next;
5759 if (mc->mc_flags & C_ALLOCD)
5765 mdb_cursor_txn(MDB_cursor *mc)
5767 if (!mc) return NULL;
5772 mdb_cursor_dbi(MDB_cursor *mc)
5778 /** Replace the key for a node with a new key.
5779 * @param[in] mp The page containing the node to operate on.
5780 * @param[in] indx The index of the node to operate on.
5781 * @param[in] key The new key to use.
5782 * @return 0 on success, non-zero on failure.
5785 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5792 indx_t ptr, i, numkeys, indx;
5795 indx = mc->mc_ki[mc->mc_top];
5796 mp = mc->mc_pg[mc->mc_top];
5797 node = NODEPTR(mp, indx);
5798 ptr = mp->mp_ptrs[indx];
5802 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5803 k2.mv_data = NODEKEY(node);
5804 k2.mv_size = node->mn_ksize;
5805 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5807 mdb_dkey(&k2, kbuf2),
5813 delta0 = delta = key->mv_size - node->mn_ksize;
5815 /* Must be 2-byte aligned. If new key is
5816 * shorter by 1, the shift will be skipped.
5818 delta += (delta & 1);
5820 if (delta > 0 && SIZELEFT(mp) < delta) {
5822 /* not enough space left, do a delete and split */
5823 DPRINTF("Not enough room, delta = %d, splitting...", delta);
5824 pgno = NODEPGNO(node);
5825 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5826 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
5829 numkeys = NUMKEYS(mp);
5830 for (i = 0; i < numkeys; i++) {
5831 if (mp->mp_ptrs[i] <= ptr)
5832 mp->mp_ptrs[i] -= delta;
5835 base = (char *)mp + mp->mp_upper;
5836 len = ptr - mp->mp_upper + NODESIZE;
5837 memmove(base - delta, base, len);
5838 mp->mp_upper -= delta;
5840 node = NODEPTR(mp, indx);
5843 /* But even if no shift was needed, update ksize */
5845 node->mn_ksize = key->mv_size;
5848 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5854 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
5856 /** Move a node from csrc to cdst.
5859 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5866 unsigned short flags;
5870 /* Mark src and dst as dirty. */
5871 if ((rc = mdb_page_touch(csrc)) ||
5872 (rc = mdb_page_touch(cdst)))
5875 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5876 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5877 key.mv_size = csrc->mc_db->md_pad;
5878 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5880 data.mv_data = NULL;
5884 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5885 assert(!((long)srcnode&1));
5886 srcpg = NODEPGNO(srcnode);
5887 flags = srcnode->mn_flags;
5888 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5889 unsigned int snum = csrc->mc_snum;
5891 /* must find the lowest key below src */
5892 mdb_page_search_root(csrc, NULL, 0);
5893 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5894 key.mv_size = csrc->mc_db->md_pad;
5895 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5897 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5898 key.mv_size = NODEKSZ(s2);
5899 key.mv_data = NODEKEY(s2);
5901 csrc->mc_snum = snum--;
5902 csrc->mc_top = snum;
5904 key.mv_size = NODEKSZ(srcnode);
5905 key.mv_data = NODEKEY(srcnode);
5907 data.mv_size = NODEDSZ(srcnode);
5908 data.mv_data = NODEDATA(srcnode);
5910 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5911 unsigned int snum = cdst->mc_snum;
5914 /* must find the lowest key below dst */
5915 mdb_page_search_root(cdst, NULL, 0);
5916 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5917 bkey.mv_size = cdst->mc_db->md_pad;
5918 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5920 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5921 bkey.mv_size = NODEKSZ(s2);
5922 bkey.mv_data = NODEKEY(s2);
5924 cdst->mc_snum = snum--;
5925 cdst->mc_top = snum;
5926 mdb_cursor_copy(cdst, &mn);
5928 rc = mdb_update_key(&mn, &bkey);
5933 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5934 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5935 csrc->mc_ki[csrc->mc_top],
5937 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5938 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5940 /* Add the node to the destination page.
5942 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5943 if (rc != MDB_SUCCESS)
5946 /* Delete the node from the source page.
5948 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5951 /* Adjust other cursors pointing to mp */
5952 MDB_cursor *m2, *m3;
5953 MDB_dbi dbi = csrc->mc_dbi;
5954 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5956 if (csrc->mc_flags & C_SUB)
5959 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5960 if (m2 == csrc) continue;
5961 if (csrc->mc_flags & C_SUB)
5962 m3 = &m2->mc_xcursor->mx_cursor;
5965 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5966 csrc->mc_ki[csrc->mc_top]) {
5967 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5968 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5973 /* Update the parent separators.
5975 if (csrc->mc_ki[csrc->mc_top] == 0) {
5976 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5977 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5978 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5980 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5981 key.mv_size = NODEKSZ(srcnode);
5982 key.mv_data = NODEKEY(srcnode);
5984 DPRINTF("update separator for source page %zu to [%s]",
5985 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5986 mdb_cursor_copy(csrc, &mn);
5989 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
5992 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5994 indx_t ix = csrc->mc_ki[csrc->mc_top];
5995 nullkey.mv_size = 0;
5996 csrc->mc_ki[csrc->mc_top] = 0;
5997 rc = mdb_update_key(csrc, &nullkey);
5998 csrc->mc_ki[csrc->mc_top] = ix;
5999 assert(rc == MDB_SUCCESS);
6003 if (cdst->mc_ki[cdst->mc_top] == 0) {
6004 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6005 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6006 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6008 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6009 key.mv_size = NODEKSZ(srcnode);
6010 key.mv_data = NODEKEY(srcnode);
6012 DPRINTF("update separator for destination page %zu to [%s]",
6013 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6014 mdb_cursor_copy(cdst, &mn);
6017 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6020 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6022 indx_t ix = cdst->mc_ki[cdst->mc_top];
6023 nullkey.mv_size = 0;
6024 cdst->mc_ki[cdst->mc_top] = 0;
6025 rc = mdb_update_key(cdst, &nullkey);
6026 cdst->mc_ki[cdst->mc_top] = ix;
6027 assert(rc == MDB_SUCCESS);
6034 /** Merge one page into another.
6035 * The nodes from the page pointed to by \b csrc will
6036 * be copied to the page pointed to by \b cdst and then
6037 * the \b csrc page will be freed.
6038 * @param[in] csrc Cursor pointing to the source page.
6039 * @param[in] cdst Cursor pointing to the destination page.
6042 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6050 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6051 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6053 assert(csrc->mc_snum > 1); /* can't merge root page */
6054 assert(cdst->mc_snum > 1);
6056 /* Mark dst as dirty. */
6057 if ((rc = mdb_page_touch(cdst)))
6060 /* Move all nodes from src to dst.
6062 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6063 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6064 key.mv_size = csrc->mc_db->md_pad;
6065 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6066 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6067 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6068 if (rc != MDB_SUCCESS)
6070 key.mv_data = (char *)key.mv_data + key.mv_size;
6073 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6074 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6075 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6076 unsigned int snum = csrc->mc_snum;
6078 /* must find the lowest key below src */
6079 mdb_page_search_root(csrc, NULL, 0);
6080 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6081 key.mv_size = csrc->mc_db->md_pad;
6082 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6084 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6085 key.mv_size = NODEKSZ(s2);
6086 key.mv_data = NODEKEY(s2);
6088 csrc->mc_snum = snum--;
6089 csrc->mc_top = snum;
6091 key.mv_size = srcnode->mn_ksize;
6092 key.mv_data = NODEKEY(srcnode);
6095 data.mv_size = NODEDSZ(srcnode);
6096 data.mv_data = NODEDATA(srcnode);
6097 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6098 if (rc != MDB_SUCCESS)
6103 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6104 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);
6106 /* Unlink the src page from parent and add to free list.
6108 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6109 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6112 rc = mdb_update_key(csrc, &key);
6118 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6119 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6120 csrc->mc_db->md_leaf_pages--;
6122 csrc->mc_db->md_branch_pages--;
6124 /* Adjust other cursors pointing to mp */
6125 MDB_cursor *m2, *m3;
6126 MDB_dbi dbi = csrc->mc_dbi;
6127 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6129 if (csrc->mc_flags & C_SUB)
6132 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6133 if (csrc->mc_flags & C_SUB)
6134 m3 = &m2->mc_xcursor->mx_cursor;
6137 if (m3 == csrc) continue;
6138 if (m3->mc_snum < csrc->mc_snum) continue;
6139 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6140 m3->mc_pg[csrc->mc_top] = mp;
6141 m3->mc_ki[csrc->mc_top] += nkeys;
6145 mdb_cursor_pop(csrc);
6147 return mdb_rebalance(csrc);
6150 /** Copy the contents of a cursor.
6151 * @param[in] csrc The cursor to copy from.
6152 * @param[out] cdst The cursor to copy to.
6155 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6159 cdst->mc_txn = csrc->mc_txn;
6160 cdst->mc_dbi = csrc->mc_dbi;
6161 cdst->mc_db = csrc->mc_db;
6162 cdst->mc_dbx = csrc->mc_dbx;
6163 cdst->mc_snum = csrc->mc_snum;
6164 cdst->mc_top = csrc->mc_top;
6165 cdst->mc_flags = csrc->mc_flags;
6167 for (i=0; i<csrc->mc_snum; i++) {
6168 cdst->mc_pg[i] = csrc->mc_pg[i];
6169 cdst->mc_ki[i] = csrc->mc_ki[i];
6173 /** Rebalance the tree after a delete operation.
6174 * @param[in] mc Cursor pointing to the page where rebalancing
6176 * @return 0 on success, non-zero on failure.
6179 mdb_rebalance(MDB_cursor *mc)
6189 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6190 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6191 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6192 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6196 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
6199 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6200 DPRINTF("no need to rebalance page %zu, above fill threshold",
6206 if (mc->mc_snum < 2) {
6207 MDB_page *mp = mc->mc_pg[0];
6208 if (NUMKEYS(mp) == 0) {
6209 DPUTS("tree is completely empty");
6210 mc->mc_db->md_root = P_INVALID;
6211 mc->mc_db->md_depth = 0;
6212 mc->mc_db->md_leaf_pages = 0;
6213 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6217 /* Adjust other cursors pointing to mp */
6218 MDB_cursor *m2, *m3;
6219 MDB_dbi dbi = mc->mc_dbi;
6221 if (mc->mc_flags & C_SUB)
6224 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6225 if (m2 == mc) continue;
6226 if (mc->mc_flags & C_SUB)
6227 m3 = &m2->mc_xcursor->mx_cursor;
6230 if (m3->mc_snum < mc->mc_snum) continue;
6231 if (m3->mc_pg[0] == mp) {
6237 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6238 DPUTS("collapsing root page!");
6239 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6240 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6241 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6244 mc->mc_db->md_depth--;
6245 mc->mc_db->md_branch_pages--;
6247 /* Adjust other cursors pointing to mp */
6248 MDB_cursor *m2, *m3;
6249 MDB_dbi dbi = mc->mc_dbi;
6251 if (mc->mc_flags & C_SUB)
6254 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6255 if (m2 == mc) continue;
6256 if (mc->mc_flags & C_SUB)
6257 m3 = &m2->mc_xcursor->mx_cursor;
6260 if (m3->mc_snum < mc->mc_snum) continue;
6261 if (m3->mc_pg[0] == mp) {
6262 m3->mc_pg[0] = mc->mc_pg[0];
6267 DPUTS("root page doesn't need rebalancing");
6271 /* The parent (branch page) must have at least 2 pointers,
6272 * otherwise the tree is invalid.
6274 ptop = mc->mc_top-1;
6275 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6277 /* Leaf page fill factor is below the threshold.
6278 * Try to move keys from left or right neighbor, or
6279 * merge with a neighbor page.
6284 mdb_cursor_copy(mc, &mn);
6285 mn.mc_xcursor = NULL;
6287 if (mc->mc_ki[ptop] == 0) {
6288 /* We're the leftmost leaf in our parent.
6290 DPUTS("reading right neighbor");
6292 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6293 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6295 mn.mc_ki[mn.mc_top] = 0;
6296 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6298 /* There is at least one neighbor to the left.
6300 DPUTS("reading left neighbor");
6302 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6303 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6305 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6306 mc->mc_ki[mc->mc_top] = 0;
6309 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6310 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);
6312 /* If the neighbor page is above threshold and has at least two
6313 * keys, move one key from it.
6315 * Otherwise we should try to merge them.
6317 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6318 return mdb_node_move(&mn, mc);
6320 if (mc->mc_ki[ptop] == 0)
6321 rc = mdb_page_merge(&mn, mc);
6323 rc = mdb_page_merge(mc, &mn);
6324 mc->mc_flags &= ~C_INITIALIZED;
6329 /** Complete a delete operation started by #mdb_cursor_del(). */
6331 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6335 /* add overflow pages to free list */
6336 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6340 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6341 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6342 mc->mc_db->md_overflow_pages -= ovpages;
6343 for (i=0; i<ovpages; i++) {
6344 DPRINTF("freed ov page %zu", pg);
6345 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6349 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6350 mc->mc_db->md_entries--;
6351 rc = mdb_rebalance(mc);
6352 if (rc != MDB_SUCCESS)
6353 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6359 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6360 MDB_val *key, MDB_val *data)
6365 MDB_val rdata, *xdata;
6369 assert(key != NULL);
6371 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6373 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6376 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6380 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6384 mdb_cursor_init(&mc, txn, dbi, &mx);
6395 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6397 /* let mdb_page_split know about this cursor if needed:
6398 * delete will trigger a rebalance; if it needs to move
6399 * a node from one page to another, it will have to
6400 * update the parent's separator key(s). If the new sepkey
6401 * is larger than the current one, the parent page may
6402 * run out of space, triggering a split. We need this
6403 * cursor to be consistent until the end of the rebalance.
6405 mc.mc_next = txn->mt_cursors[dbi];
6406 txn->mt_cursors[dbi] = &mc;
6407 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6408 txn->mt_cursors[dbi] = mc.mc_next;
6413 /** Split a page and insert a new node.
6414 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6415 * The cursor will be updated to point to the actual page and index where
6416 * the node got inserted after the split.
6417 * @param[in] newkey The key for the newly inserted node.
6418 * @param[in] newdata The data for the newly inserted node.
6419 * @param[in] newpgno The page number, if the new node is a branch node.
6420 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6421 * @return 0 on success, non-zero on failure.
6424 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6425 unsigned int nflags)
6428 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6431 unsigned int i, j, split_indx, nkeys, pmax;
6433 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6435 MDB_page *mp, *rp, *pp;
6440 mp = mc->mc_pg[mc->mc_top];
6441 newindx = mc->mc_ki[mc->mc_top];
6443 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6444 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6445 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6447 /* Create a right sibling. */
6448 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6450 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6452 if (mc->mc_snum < 2) {
6453 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6455 /* shift current top to make room for new parent */
6456 mc->mc_pg[1] = mc->mc_pg[0];
6457 mc->mc_ki[1] = mc->mc_ki[0];
6460 mc->mc_db->md_root = pp->mp_pgno;
6461 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6462 mc->mc_db->md_depth++;
6465 /* Add left (implicit) pointer. */
6466 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6467 /* undo the pre-push */
6468 mc->mc_pg[0] = mc->mc_pg[1];
6469 mc->mc_ki[0] = mc->mc_ki[1];
6470 mc->mc_db->md_root = mp->mp_pgno;
6471 mc->mc_db->md_depth--;
6478 ptop = mc->mc_top-1;
6479 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6482 mc->mc_flags |= C_SPLITTING;
6483 mdb_cursor_copy(mc, &mn);
6484 mn.mc_pg[mn.mc_top] = rp;
6485 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6487 if (nflags & MDB_APPEND) {
6488 mn.mc_ki[mn.mc_top] = 0;
6490 split_indx = newindx;
6495 nkeys = NUMKEYS(mp);
6496 split_indx = nkeys / 2;
6497 if (newindx < split_indx)
6503 unsigned int lsize, rsize, ksize;
6504 /* Move half of the keys to the right sibling */
6506 x = mc->mc_ki[mc->mc_top] - split_indx;
6507 ksize = mc->mc_db->md_pad;
6508 split = LEAF2KEY(mp, split_indx, ksize);
6509 rsize = (nkeys - split_indx) * ksize;
6510 lsize = (nkeys - split_indx) * sizeof(indx_t);
6511 mp->mp_lower -= lsize;
6512 rp->mp_lower += lsize;
6513 mp->mp_upper += rsize - lsize;
6514 rp->mp_upper -= rsize - lsize;
6515 sepkey.mv_size = ksize;
6516 if (newindx == split_indx) {
6517 sepkey.mv_data = newkey->mv_data;
6519 sepkey.mv_data = split;
6522 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6523 memcpy(rp->mp_ptrs, split, rsize);
6524 sepkey.mv_data = rp->mp_ptrs;
6525 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6526 memcpy(ins, newkey->mv_data, ksize);
6527 mp->mp_lower += sizeof(indx_t);
6528 mp->mp_upper -= ksize - sizeof(indx_t);
6531 memcpy(rp->mp_ptrs, split, x * ksize);
6532 ins = LEAF2KEY(rp, x, ksize);
6533 memcpy(ins, newkey->mv_data, ksize);
6534 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6535 rp->mp_lower += sizeof(indx_t);
6536 rp->mp_upper -= ksize - sizeof(indx_t);
6537 mc->mc_ki[mc->mc_top] = x;
6538 mc->mc_pg[mc->mc_top] = rp;
6543 /* For leaf pages, check the split point based on what
6544 * fits where, since otherwise mdb_node_add can fail.
6546 * This check is only needed when the data items are
6547 * relatively large, such that being off by one will
6548 * make the difference between success or failure.
6550 * It's also relevant if a page happens to be laid out
6551 * such that one half of its nodes are all "small" and
6552 * the other half of its nodes are "large." If the new
6553 * item is also "large" and falls on the half with
6554 * "large" nodes, it also may not fit.
6557 unsigned int psize, nsize;
6558 /* Maximum free space in an empty page */
6559 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6560 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6561 if ((nkeys < 20) || (nsize > pmax/16)) {
6562 if (newindx <= split_indx) {
6565 for (i=0; i<split_indx; i++) {
6566 node = NODEPTR(mp, i);
6567 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6568 if (F_ISSET(node->mn_flags, F_BIGDATA))
6569 psize += sizeof(pgno_t);
6571 psize += NODEDSZ(node);
6575 split_indx = newindx;
6586 for (i=nkeys-1; i>=split_indx; i--) {
6587 node = NODEPTR(mp, i);
6588 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6589 if (F_ISSET(node->mn_flags, F_BIGDATA))
6590 psize += sizeof(pgno_t);
6592 psize += NODEDSZ(node);
6596 split_indx = newindx;
6607 /* First find the separating key between the split pages.
6608 * The case where newindx == split_indx is ambiguous; the
6609 * new item could go to the new page or stay on the original
6610 * page. If newpos == 1 it goes to the new page.
6612 if (newindx == split_indx && newpos) {
6613 sepkey.mv_size = newkey->mv_size;
6614 sepkey.mv_data = newkey->mv_data;
6616 node = NODEPTR(mp, split_indx);
6617 sepkey.mv_size = node->mn_ksize;
6618 sepkey.mv_data = NODEKEY(node);
6622 DPRINTF("separator is [%s]", DKEY(&sepkey));
6624 /* Copy separator key to the parent.
6626 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6630 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6633 if (mn.mc_snum == mc->mc_snum) {
6634 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6635 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6636 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6637 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6642 /* Right page might now have changed parent.
6643 * Check if left page also changed parent.
6645 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6646 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6647 for (i=0; i<ptop; i++) {
6648 mc->mc_pg[i] = mn.mc_pg[i];
6649 mc->mc_ki[i] = mn.mc_ki[i];
6651 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6652 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6656 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6659 mc->mc_flags ^= C_SPLITTING;
6660 if (rc != MDB_SUCCESS) {
6663 if (nflags & MDB_APPEND) {
6664 mc->mc_pg[mc->mc_top] = rp;
6665 mc->mc_ki[mc->mc_top] = 0;
6666 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6669 for (i=0; i<mc->mc_top; i++)
6670 mc->mc_ki[i] = mn.mc_ki[i];
6677 /* Move half of the keys to the right sibling. */
6679 /* grab a page to hold a temporary copy */
6680 copy = mdb_page_malloc(mc);
6684 copy->mp_pgno = mp->mp_pgno;
6685 copy->mp_flags = mp->mp_flags;
6686 copy->mp_lower = PAGEHDRSZ;
6687 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6688 mc->mc_pg[mc->mc_top] = copy;
6689 for (i = j = 0; i <= nkeys; j++) {
6690 if (i == split_indx) {
6691 /* Insert in right sibling. */
6692 /* Reset insert index for right sibling. */
6693 if (i != newindx || (newpos ^ ins_new)) {
6695 mc->mc_pg[mc->mc_top] = rp;
6699 if (i == newindx && !ins_new) {
6700 /* Insert the original entry that caused the split. */
6701 rkey.mv_data = newkey->mv_data;
6702 rkey.mv_size = newkey->mv_size;
6711 /* Update index for the new key. */
6712 mc->mc_ki[mc->mc_top] = j;
6713 } else if (i == nkeys) {
6716 node = NODEPTR(mp, i);
6717 rkey.mv_data = NODEKEY(node);
6718 rkey.mv_size = node->mn_ksize;
6720 xdata.mv_data = NODEDATA(node);
6721 xdata.mv_size = NODEDSZ(node);
6724 pgno = NODEPGNO(node);
6725 flags = node->mn_flags;
6730 if (!IS_LEAF(mp) && j == 0) {
6731 /* First branch index doesn't need key data. */
6735 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6739 nkeys = NUMKEYS(copy);
6740 for (i=0; i<nkeys; i++)
6741 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6742 mp->mp_lower = copy->mp_lower;
6743 mp->mp_upper = copy->mp_upper;
6744 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6745 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6747 /* reset back to original page */
6748 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6749 mc->mc_pg[mc->mc_top] = mp;
6750 if (nflags & MDB_RESERVE) {
6751 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6752 if (!(node->mn_flags & F_BIGDATA))
6753 newdata->mv_data = NODEDATA(node);
6759 /* return tmp page to freelist */
6760 mdb_page_free(mc->mc_txn->mt_env, copy);
6763 /* Adjust other cursors pointing to mp */
6764 MDB_cursor *m2, *m3;
6765 MDB_dbi dbi = mc->mc_dbi;
6766 int fixup = NUMKEYS(mp);
6768 if (mc->mc_flags & C_SUB)
6771 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6772 if (m2 == mc) continue;
6773 if (mc->mc_flags & C_SUB)
6774 m3 = &m2->mc_xcursor->mx_cursor;
6777 if (!(m3->mc_flags & C_INITIALIZED))
6779 if (m3->mc_flags & C_SPLITTING)
6784 for (k=m3->mc_top; k>=0; k--) {
6785 m3->mc_ki[k+1] = m3->mc_ki[k];
6786 m3->mc_pg[k+1] = m3->mc_pg[k];
6788 if (m3->mc_ki[0] >= split_indx) {
6793 m3->mc_pg[0] = mc->mc_pg[0];
6797 if (m3->mc_pg[mc->mc_top] == mp) {
6798 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6799 m3->mc_ki[mc->mc_top]++;
6800 if (m3->mc_ki[mc->mc_top] >= fixup) {
6801 m3->mc_pg[mc->mc_top] = rp;
6802 m3->mc_ki[mc->mc_top] -= fixup;
6803 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6805 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6806 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6815 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6816 MDB_val *key, MDB_val *data, unsigned int flags)
6821 assert(key != NULL);
6822 assert(data != NULL);
6824 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6827 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6831 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6835 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6838 mdb_cursor_init(&mc, txn, dbi, &mx);
6839 return mdb_cursor_put(&mc, key, data, flags);
6843 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6845 if ((flag & CHANGEABLE) != flag)
6848 env->me_flags |= flag;
6850 env->me_flags &= ~flag;
6855 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6860 *arg = env->me_flags;
6865 mdb_env_get_path(MDB_env *env, const char **arg)
6870 *arg = env->me_path;
6874 /** Common code for #mdb_stat() and #mdb_env_stat().
6875 * @param[in] env the environment to operate in.
6876 * @param[in] db the #MDB_db record containing the stats to return.
6877 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6878 * @return 0, this function always succeeds.
6881 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6883 arg->ms_psize = env->me_psize;
6884 arg->ms_depth = db->md_depth;
6885 arg->ms_branch_pages = db->md_branch_pages;
6886 arg->ms_leaf_pages = db->md_leaf_pages;
6887 arg->ms_overflow_pages = db->md_overflow_pages;
6888 arg->ms_entries = db->md_entries;
6893 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6897 if (env == NULL || arg == NULL)
6900 toggle = mdb_env_pick_meta(env);
6902 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6906 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
6910 if (env == NULL || arg == NULL)
6913 toggle = mdb_env_pick_meta(env);
6914 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
6915 arg->me_mapsize = env->me_mapsize;
6916 arg->me_maxreaders = env->me_maxreaders;
6917 arg->me_numreaders = env->me_numreaders;
6918 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
6919 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
6923 /** Set the default comparison functions for a database.
6924 * Called immediately after a database is opened to set the defaults.
6925 * The user can then override them with #mdb_set_compare() or
6926 * #mdb_set_dupsort().
6927 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6928 * @param[in] dbi A database handle returned by #mdb_dbi_open()
6931 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6933 uint16_t f = txn->mt_dbs[dbi].md_flags;
6935 txn->mt_dbxs[dbi].md_cmp =
6936 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6937 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6939 txn->mt_dbxs[dbi].md_dcmp =
6940 !(f & MDB_DUPSORT) ? 0 :
6941 ((f & MDB_INTEGERDUP)
6942 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6943 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6946 #define PERSISTENT_FLAGS 0xffff
6947 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
6948 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
6949 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6954 int rc, dbflag, exact;
6955 unsigned int unused = 0;
6958 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6959 mdb_default_cmp(txn, FREE_DBI);
6962 if ((flags & VALID_FLAGS) != flags)
6968 if (flags & PERSISTENT_FLAGS) {
6969 uint16_t f2 = flags & PERSISTENT_FLAGS;
6970 /* make sure flag changes get committed */
6971 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
6972 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
6973 txn->mt_flags |= MDB_TXN_DIRTY;
6976 mdb_default_cmp(txn, MAIN_DBI);
6980 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6981 mdb_default_cmp(txn, MAIN_DBI);
6984 /* Is the DB already open? */
6986 for (i=2; i<txn->mt_numdbs; i++) {
6987 if (!txn->mt_dbxs[i].md_name.mv_size) {
6988 /* Remember this free slot */
6989 if (!unused) unused = i;
6992 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6993 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6999 /* If no free slot and max hit, fail */
7000 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7001 return MDB_DBS_FULL;
7003 /* Find the DB info */
7007 key.mv_data = (void *)name;
7008 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7009 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7010 if (rc == MDB_SUCCESS) {
7011 /* make sure this is actually a DB */
7012 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7013 if (!(node->mn_flags & F_SUBDATA))
7015 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7016 /* Create if requested */
7018 data.mv_size = sizeof(MDB_db);
7019 data.mv_data = &dummy;
7020 memset(&dummy, 0, sizeof(dummy));
7021 dummy.md_root = P_INVALID;
7022 dummy.md_flags = flags & PERSISTENT_FLAGS;
7023 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7027 /* OK, got info, add to table */
7028 if (rc == MDB_SUCCESS) {
7029 unsigned int slot = unused ? unused : txn->mt_numdbs;
7030 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7031 txn->mt_dbxs[slot].md_name.mv_size = len;
7032 txn->mt_dbxs[slot].md_rel = NULL;
7033 txn->mt_dbflags[slot] = dbflag;
7034 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7036 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7037 mdb_default_cmp(txn, slot);
7040 txn->mt_env->me_numdbs++;
7047 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7049 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7052 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7055 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7058 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
7060 ptr = env->me_dbxs[dbi].md_name.mv_data;
7061 env->me_dbxs[dbi].md_name.mv_data = NULL;
7062 env->me_dbxs[dbi].md_name.mv_size = 0;
7066 /** Add all the DB's pages to the free list.
7067 * @param[in] mc Cursor on the DB to free.
7068 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7069 * @return 0 on success, non-zero on failure.
7072 mdb_drop0(MDB_cursor *mc, int subs)
7076 rc = mdb_page_search(mc, NULL, 0);
7077 if (rc == MDB_SUCCESS) {
7082 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7083 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
7086 mdb_cursor_copy(mc, &mx);
7087 while (mc->mc_snum > 0) {
7088 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7089 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7090 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7091 if (ni->mn_flags & F_SUBDATA) {
7092 mdb_xcursor_init1(mc, ni);
7093 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7099 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7101 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7104 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7109 rc = mdb_cursor_sibling(mc, 1);
7111 /* no more siblings, go back to beginning
7112 * of previous level.
7115 for (i=1; i<mc->mc_top; i++)
7116 mc->mc_pg[i] = mx.mc_pg[i];
7120 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7121 mc->mc_db->md_root);
7126 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7131 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1)
7134 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7137 rc = mdb_cursor_open(txn, dbi, &mc);
7141 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7145 /* Can't delete the main DB */
7146 if (del && dbi > MAIN_DBI) {
7147 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7149 txn->mt_dbflags[dbi] = DB_STALE;
7150 mdb_dbi_close(txn->mt_env, dbi);
7153 /* reset the DB record, mark it dirty */
7154 txn->mt_dbflags[dbi] |= DB_DIRTY;
7155 txn->mt_dbs[dbi].md_depth = 0;
7156 txn->mt_dbs[dbi].md_branch_pages = 0;
7157 txn->mt_dbs[dbi].md_leaf_pages = 0;
7158 txn->mt_dbs[dbi].md_overflow_pages = 0;
7159 txn->mt_dbs[dbi].md_entries = 0;
7160 txn->mt_dbs[dbi].md_root = P_INVALID;
7162 if (!txn->mt_u.dirty_list[0].mid) {
7165 /* make sure we have at least one dirty page in this txn
7166 * otherwise these changes will be ignored.
7168 key.mv_size = sizeof(txnid_t);
7169 key.mv_data = &txn->mt_txnid;
7170 data.mv_size = sizeof(MDB_ID);
7171 data.mv_data = txn->mt_free_pgs;
7172 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
7173 rc = mdb_cursor_put(&m2, &key, &data, 0);
7177 mdb_cursor_close(mc);
7181 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7183 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7186 txn->mt_dbxs[dbi].md_cmp = cmp;
7190 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7192 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7195 txn->mt_dbxs[dbi].md_dcmp = cmp;
7199 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7201 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7204 txn->mt_dbxs[dbi].md_rel = rel;
7208 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7210 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7213 txn->mt_dbxs[dbi].md_relctx = ctx;