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 /** A set of pages freed by an earlier transaction. */
915 typedef struct MDB_oldpages {
916 /** Usually we only read one record from the FREEDB at a time, but
917 * in case we read more, this will chain them together.
919 struct MDB_oldpages *mo_next;
920 /** The ID of the transaction in which these pages were freed. */
922 /** An #MDB_IDL of the pages */
923 pgno_t mo_pages[1]; /* dynamic */
926 /** The database environment. */
928 HANDLE me_fd; /**< The main data file */
929 HANDLE me_lfd; /**< The lock file */
930 HANDLE me_mfd; /**< just for writing the meta pages */
931 /** Failed to update the meta page. Probably an I/O error. */
932 #define MDB_FATAL_ERROR 0x80000000U
933 /** Read-only Filesystem. Allow read access, no locking. */
934 #define MDB_ROFS 0x40000000U
935 /** Some fields are initialized. */
936 #define MDB_ENV_ACTIVE 0x20000000U
937 uint32_t me_flags; /**< @ref mdb_env */
938 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
939 unsigned int me_maxreaders; /**< size of the reader table */
940 unsigned int me_numreaders; /**< max numreaders set by this env */
941 MDB_dbi me_numdbs; /**< number of DBs opened */
942 MDB_dbi me_maxdbs; /**< size of the DB table */
943 pid_t me_pid; /**< process ID of this env */
944 char *me_path; /**< path to the DB files */
945 char *me_map; /**< the memory map of the data file */
946 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
947 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
948 MDB_txn *me_txn; /**< current write transaction */
949 size_t me_mapsize; /**< size of the data memory map */
950 off_t me_size; /**< current file size */
951 pgno_t me_maxpg; /**< me_mapsize / me_psize */
952 txnid_t me_pgfirst; /**< ID of first old page record we used */
953 txnid_t me_pglast; /**< ID of last old page record we used */
954 MDB_dbx *me_dbxs; /**< array of static DB info */
955 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
956 MDB_oldpages *me_pghead; /**< list of old page records */
957 MDB_oldpages *me_pgfree; /**< list of page records to free */
958 pthread_key_t me_txkey; /**< thread-key for readers */
959 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
960 /** IDL of pages that became unused in a write txn */
962 /** ID2L of pages that were written during a write txn */
963 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
965 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
967 #elif defined(MDB_USE_POSIX_SEM)
968 sem_t *me_rmutex; /* Shared mutexes are not supported */
972 /** max number of pages to commit in one writev() call */
973 #define MDB_COMMIT_PAGES 64
974 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
975 #undef MDB_COMMIT_PAGES
976 #define MDB_COMMIT_PAGES IOV_MAX
979 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
980 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
981 static int mdb_page_touch(MDB_cursor *mc);
983 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
984 static int mdb_page_search_root(MDB_cursor *mc,
985 MDB_val *key, int modify);
986 #define MDB_PS_MODIFY 1
987 #define MDB_PS_ROOTONLY 2
988 static int mdb_page_search(MDB_cursor *mc,
989 MDB_val *key, int flags);
990 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
992 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
993 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
994 pgno_t newpgno, unsigned int nflags);
996 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
997 static int mdb_env_pick_meta(const MDB_env *env);
998 static int mdb_env_write_meta(MDB_txn *txn);
999 static void mdb_env_close0(MDB_env *env, int excl);
1001 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1002 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1003 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1004 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1005 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1006 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1007 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1008 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1009 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1011 static int mdb_rebalance(MDB_cursor *mc);
1012 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1014 static void mdb_cursor_pop(MDB_cursor *mc);
1015 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1017 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1018 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1019 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1020 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1021 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1023 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1024 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1026 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1027 static void mdb_xcursor_init0(MDB_cursor *mc);
1028 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1030 static int mdb_drop0(MDB_cursor *mc, int subs);
1031 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1034 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1038 static SECURITY_DESCRIPTOR mdb_null_sd;
1039 static SECURITY_ATTRIBUTES mdb_all_sa;
1040 static int mdb_sec_inited;
1043 /** Return the library version info. */
1045 mdb_version(int *major, int *minor, int *patch)
1047 if (major) *major = MDB_VERSION_MAJOR;
1048 if (minor) *minor = MDB_VERSION_MINOR;
1049 if (patch) *patch = MDB_VERSION_PATCH;
1050 return MDB_VERSION_STRING;
1053 /** Table of descriptions for MDB @ref errors */
1054 static char *const mdb_errstr[] = {
1055 "MDB_KEYEXIST: Key/data pair already exists",
1056 "MDB_NOTFOUND: No matching key/data pair found",
1057 "MDB_PAGE_NOTFOUND: Requested page not found",
1058 "MDB_CORRUPTED: Located page was wrong type",
1059 "MDB_PANIC: Update of meta page failed",
1060 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1061 "MDB_INVALID: File is not an MDB file",
1062 "MDB_MAP_FULL: Environment mapsize limit reached",
1063 "MDB_DBS_FULL: Environment maxdbs limit reached",
1064 "MDB_READERS_FULL: Environment maxreaders limit reached",
1065 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1066 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1067 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1068 "MDB_PAGE_FULL: Internal error - page has no more space"
1072 mdb_strerror(int err)
1076 return ("Successful return: 0");
1078 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1079 i = err - MDB_KEYEXIST;
1080 return mdb_errstr[i];
1083 return strerror(err);
1087 /** Display a key in hexadecimal and return the address of the result.
1088 * @param[in] key the key to display
1089 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1090 * @return The key in hexadecimal form.
1093 mdb_dkey(MDB_val *key, char *buf)
1096 unsigned char *c = key->mv_data;
1098 if (key->mv_size > MDB_MAXKEYSIZE)
1099 return "MDB_MAXKEYSIZE";
1100 /* may want to make this a dynamic check: if the key is mostly
1101 * printable characters, print it as-is instead of converting to hex.
1105 for (i=0; i<key->mv_size; i++)
1106 ptr += sprintf(ptr, "%02x", *c++);
1108 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1113 /** Display all the keys in the page. */
1115 mdb_page_list(MDB_page *mp)
1118 unsigned int i, nkeys, nsize;
1122 nkeys = NUMKEYS(mp);
1123 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1124 for (i=0; i<nkeys; i++) {
1125 node = NODEPTR(mp, i);
1126 key.mv_size = node->mn_ksize;
1127 key.mv_data = node->mn_data;
1128 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1129 if (IS_BRANCH(mp)) {
1130 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1133 if (F_ISSET(node->mn_flags, F_BIGDATA))
1134 nsize += sizeof(pgno_t);
1136 nsize += NODEDSZ(node);
1137 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1143 mdb_cursor_chk(MDB_cursor *mc)
1149 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1150 for (i=0; i<mc->mc_top; i++) {
1152 node = NODEPTR(mp, mc->mc_ki[i]);
1153 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1156 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1162 /** Count all the pages in each DB and in the freelist
1163 * and make sure it matches the actual number of pages
1166 static void mdb_audit(MDB_txn *txn)
1170 MDB_ID freecount, count;
1175 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1176 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1177 freecount += *(MDB_ID *)data.mv_data;
1180 for (i = 0; i<txn->mt_numdbs; i++) {
1181 MDB_xcursor mx, *mxp;
1182 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1183 mdb_cursor_init(&mc, txn, i, mxp);
1184 if (txn->mt_dbs[i].md_root == P_INVALID)
1186 count += txn->mt_dbs[i].md_branch_pages +
1187 txn->mt_dbs[i].md_leaf_pages +
1188 txn->mt_dbs[i].md_overflow_pages;
1189 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1190 mdb_page_search(&mc, NULL, 0);
1194 mp = mc.mc_pg[mc.mc_top];
1195 for (j=0; j<NUMKEYS(mp); j++) {
1196 MDB_node *leaf = NODEPTR(mp, j);
1197 if (leaf->mn_flags & F_SUBDATA) {
1199 memcpy(&db, NODEDATA(leaf), sizeof(db));
1200 count += db.md_branch_pages + db.md_leaf_pages +
1201 db.md_overflow_pages;
1205 while (mdb_cursor_sibling(&mc, 1) == 0);
1208 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1209 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1210 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1216 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1218 return txn->mt_dbxs[dbi].md_cmp(a, b);
1222 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1224 if (txn->mt_dbxs[dbi].md_dcmp)
1225 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1227 return EINVAL; /* too bad you can't distinguish this from a valid result */
1230 /** Allocate a single page.
1231 * Re-use old malloc'd pages first, otherwise just malloc.
1234 mdb_page_malloc(MDB_cursor *mc) {
1236 size_t sz = mc->mc_txn->mt_env->me_psize;
1237 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1238 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1239 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1240 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1241 } else if ((ret = malloc(sz)) != NULL) {
1242 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1248 mdb_page_free(MDB_env *env, MDB_page *mp)
1250 mp->mp_next = env->me_dpages;
1251 VGMEMP_FREE(env, mp);
1252 env->me_dpages = mp;
1255 /** Allocate pages for writing.
1256 * If there are free pages available from older transactions, they
1257 * will be re-used first. Otherwise a new page will be allocated.
1258 * @param[in] mc cursor A cursor handle identifying the transaction and
1259 * database for which we are allocating.
1260 * @param[in] num the number of pages to allocate.
1261 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1262 * will always be satisfied by a single contiguous chunk of memory.
1263 * @return 0 on success, non-zero on failure.
1266 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1268 MDB_txn *txn = mc->mc_txn;
1270 pgno_t pgno = P_INVALID;
1272 txnid_t oldest = 0, last;
1277 /* If our dirty list is already full, we can't do anything */
1278 if (txn->mt_u.dirty_list[0].mid >= MDB_IDL_UM_MAX)
1279 return MDB_TXN_FULL;
1281 /* The free list won't have any content at all until txn 2 has
1282 * committed. The pages freed by txn 2 will be unreferenced
1283 * after txn 3 commits, and so will be safe to re-use in txn 4.
1285 if (txn->mt_txnid > 3) {
1287 if (!txn->mt_env->me_pghead &&
1288 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1289 /* See if there's anything in the free DB */
1296 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1297 if (!txn->mt_env->me_pgfirst) {
1298 mdb_page_search(&m2, NULL, 0);
1299 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1300 kptr = (txnid_t *)NODEKEY(leaf);
1307 last = txn->mt_env->me_pglast + 1;
1309 key.mv_data = &last;
1310 key.mv_size = sizeof(last);
1311 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1314 last = *(txnid_t *)key.mv_data;
1320 oldest = txn->mt_txnid - 1;
1321 nr = txn->mt_env->me_txns->mti_numreaders;
1322 r = txn->mt_env->me_txns->mti_readers;
1323 for (i=0; i<nr; i++) {
1324 if (!r[i].mr_pid) continue;
1331 if (oldest > last) {
1332 /* It's usable, grab it.
1337 if (!txn->mt_env->me_pgfirst) {
1338 mdb_node_read(txn, leaf, &data);
1340 txn->mt_env->me_pglast = last;
1341 if (!txn->mt_env->me_pgfirst)
1342 txn->mt_env->me_pgfirst = last;
1343 idl = (MDB_ID *) data.mv_data;
1344 /* We might have a zero-length IDL due to freelist growth
1345 * during a prior commit
1347 if (!idl[0]) goto again;
1348 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1351 mop->mo_next = txn->mt_env->me_pghead;
1352 mop->mo_txnid = last;
1353 txn->mt_env->me_pghead = mop;
1354 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1359 DPRINTF("IDL read txn %zu root %zu num %zu",
1360 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1361 for (i=0; i<idl[0]; i++) {
1362 DPRINTF("IDL %zu", idl[i+1]);
1369 if (txn->mt_env->me_pghead) {
1370 MDB_oldpages *mop = txn->mt_env->me_pghead;
1373 int retry = 500, readit = 0, n2 = num-1;
1374 unsigned int i, j, k;
1376 /* If current list is too short, must fetch more and coalesce */
1377 if (mop->mo_pages[0] < (unsigned)num)
1380 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1382 /* If on freelist, don't try to read more. If what we have
1383 * right now isn't enough just use new pages.
1384 * TODO: get all of this working. Many circular dependencies...
1386 if (mc->mc_dbi == FREE_DBI) {
1396 last = mop->mo_txnid + 1;
1398 /* We haven't hit the readers list yet? */
1404 oldest = txn->mt_txnid - 1;
1405 nr = txn->mt_env->me_txns->mti_numreaders;
1406 r = txn->mt_env->me_txns->mti_readers;
1407 for (i=0; i<nr; i++) {
1408 if (!r[i].mr_pid) continue;
1415 /* There's nothing we can use on the freelist */
1416 if (oldest - last < 1)
1420 key.mv_data = &last;
1421 key.mv_size = sizeof(last);
1422 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1425 idl = (MDB_ID *) data.mv_data;
1426 mop2 = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - 2*sizeof(pgno_t) + MDB_IDL_SIZEOF(mop->mo_pages));
1429 /* merge in sorted order */
1430 i = idl[0]; j = mop->mo_pages[0]; mop2->mo_pages[0] = k = i+j;
1431 mop->mo_pages[0] = P_INVALID;
1432 while (i>0 || j>0) {
1433 if (i && idl[i] < mop->mo_pages[j])
1434 mop2->mo_pages[k--] = idl[i--];
1436 mop2->mo_pages[k--] = mop->mo_pages[j--];
1438 txn->mt_env->me_pglast = last;
1439 mop2->mo_txnid = last;
1440 mop2->mo_next = mop->mo_next;
1441 txn->mt_env->me_pghead = mop2;
1444 /* Keep trying to read until we have enough */
1445 if (mop->mo_pages[0] < (unsigned)num) {
1450 /* current list has enough pages, but are they contiguous? */
1451 for (i=mop->mo_pages[0]; i>=(unsigned)num; i--) {
1452 if (mop->mo_pages[i-n2] == mop->mo_pages[i] + n2) {
1453 pgno = mop->mo_pages[i];
1455 /* move any stragglers down */
1456 for (j=i+num; j<=mop->mo_pages[0]; j++)
1457 mop->mo_pages[i++] = mop->mo_pages[j];
1458 mop->mo_pages[0] -= num;
1463 /* Stop if we succeeded, or no more retries */
1464 if (!retry || pgno != P_INVALID)
1471 /* peel pages off tail, so we only have to truncate the list */
1472 pgno = MDB_IDL_LAST(mop->mo_pages);
1475 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1476 txn->mt_env->me_pghead = mop->mo_next;
1477 if (mc->mc_dbi == FREE_DBI) {
1478 mop->mo_next = txn->mt_env->me_pgfree;
1479 txn->mt_env->me_pgfree = mop;
1487 if (pgno == P_INVALID) {
1488 /* DB size is maxed out */
1489 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1490 DPUTS("DB size maxed out");
1491 return MDB_MAP_FULL;
1494 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1495 if (pgno == P_INVALID) {
1496 pgno = txn->mt_next_pgno;
1497 txn->mt_next_pgno += num;
1499 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1502 if (txn->mt_env->me_dpages && num == 1) {
1503 np = txn->mt_env->me_dpages;
1504 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1505 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1506 txn->mt_env->me_dpages = np->mp_next;
1508 size_t sz = txn->mt_env->me_psize * num;
1509 if ((np = malloc(sz)) == NULL)
1511 VGMEMP_ALLOC(txn->mt_env, np, sz);
1513 if (pgno == P_INVALID) {
1514 np->mp_pgno = txn->mt_next_pgno;
1515 txn->mt_next_pgno += num;
1520 mid.mid = np->mp_pgno;
1522 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1523 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1525 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1532 /** Copy a page: avoid copying unused portions of the page.
1533 * @param[in] dst page to copy into
1534 * @param[in] src page to copy from
1537 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1539 dst->mp_flags = src->mp_flags | P_DIRTY;
1540 dst->mp_pages = src->mp_pages;
1542 if (IS_LEAF2(src)) {
1543 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1545 unsigned int i, nkeys = NUMKEYS(src);
1546 for (i=0; i<nkeys; i++)
1547 dst->mp_ptrs[i] = src->mp_ptrs[i];
1548 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1549 psize - src->mp_upper);
1553 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1554 * @param[in] mc cursor pointing to the page to be touched
1555 * @return 0 on success, non-zero on failure.
1558 mdb_page_touch(MDB_cursor *mc)
1560 MDB_page *mp = mc->mc_pg[mc->mc_top];
1564 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1566 if ((rc = mdb_page_alloc(mc, 1, &np)))
1568 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1569 assert(mp->mp_pgno != np->mp_pgno);
1570 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1572 /* If page isn't full, just copy the used portion */
1573 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1576 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1578 np->mp_flags |= P_DIRTY;
1583 /* Adjust other cursors pointing to mp */
1584 if (mc->mc_flags & C_SUB) {
1585 MDB_cursor *m2, *m3;
1586 MDB_dbi dbi = mc->mc_dbi-1;
1588 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1589 if (m2 == mc) continue;
1590 m3 = &m2->mc_xcursor->mx_cursor;
1591 if (m3->mc_snum < mc->mc_snum) continue;
1592 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1593 m3->mc_pg[mc->mc_top] = mp;
1599 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1600 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1601 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1602 m2->mc_pg[mc->mc_top] = mp;
1606 mc->mc_pg[mc->mc_top] = mp;
1607 /** If this page has a parent, update the parent to point to
1611 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1613 mc->mc_db->md_root = mp->mp_pgno;
1614 } else if (mc->mc_txn->mt_parent) {
1617 /* If txn has a parent, make sure the page is in our
1620 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1621 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1622 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1623 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1624 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1625 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1626 mc->mc_pg[mc->mc_top] = mp;
1631 if (mc->mc_txn->mt_u.dirty_list[0].mid >= MDB_IDL_UM_MAX)
1632 return MDB_TXN_FULL;
1634 np = mdb_page_malloc(mc);
1637 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1638 mid.mid = np->mp_pgno;
1640 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1648 mdb_env_sync(MDB_env *env, int force)
1651 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1652 if (env->me_flags & MDB_WRITEMAP) {
1653 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1654 ? MS_ASYNC : MS_SYNC;
1655 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1658 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1662 if (MDB_FDATASYNC(env->me_fd))
1669 /** Make shadow copies of all of parent txn's cursors */
1671 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1673 MDB_cursor *mc, *m2;
1674 unsigned int i, j, size;
1676 for (i=0;i<src->mt_numdbs; i++) {
1677 if (src->mt_cursors[i]) {
1678 size = sizeof(MDB_cursor);
1679 if (src->mt_cursors[i]->mc_xcursor)
1680 size += sizeof(MDB_xcursor);
1681 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1688 mc->mc_db = &dst->mt_dbs[i];
1689 mc->mc_dbx = m2->mc_dbx;
1690 mc->mc_dbflag = &dst->mt_dbflags[i];
1691 mc->mc_snum = m2->mc_snum;
1692 mc->mc_top = m2->mc_top;
1693 mc->mc_flags = m2->mc_flags | C_SHADOW;
1694 for (j=0; j<mc->mc_snum; j++) {
1695 mc->mc_pg[j] = m2->mc_pg[j];
1696 mc->mc_ki[j] = m2->mc_ki[j];
1698 if (m2->mc_xcursor) {
1699 MDB_xcursor *mx, *mx2;
1700 mx = (MDB_xcursor *)(mc+1);
1701 mc->mc_xcursor = mx;
1702 mx2 = m2->mc_xcursor;
1703 mx->mx_db = mx2->mx_db;
1704 mx->mx_dbx = mx2->mx_dbx;
1705 mx->mx_dbflag = mx2->mx_dbflag;
1706 mx->mx_cursor.mc_txn = dst;
1707 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1708 mx->mx_cursor.mc_db = &mx->mx_db;
1709 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1710 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1711 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1712 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1713 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1714 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1715 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1716 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1719 mc->mc_xcursor = NULL;
1721 mc->mc_next = dst->mt_cursors[i];
1722 dst->mt_cursors[i] = mc;
1729 /** Merge shadow cursors back into parent's */
1731 mdb_cursor_merge(MDB_txn *txn)
1734 for (i=0; i<txn->mt_numdbs; i++) {
1735 if (txn->mt_cursors[i]) {
1737 while ((mc = txn->mt_cursors[i])) {
1738 txn->mt_cursors[i] = mc->mc_next;
1739 if (mc->mc_flags & C_SHADOW) {
1740 MDB_cursor *m2 = mc->mc_orig;
1742 m2->mc_snum = mc->mc_snum;
1743 m2->mc_top = mc->mc_top;
1744 for (j=0; j<mc->mc_snum; j++) {
1745 m2->mc_pg[j] = mc->mc_pg[j];
1746 m2->mc_ki[j] = mc->mc_ki[j];
1749 if (mc->mc_flags & C_ALLOCD)
1757 mdb_txn_reset0(MDB_txn *txn);
1759 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1760 * @param[in] txn the transaction handle to initialize
1761 * @return 0 on success, non-zero on failure. This can only
1762 * fail for read-only transactions, and then only if the
1763 * reader table is full.
1766 mdb_txn_renew0(MDB_txn *txn)
1768 MDB_env *env = txn->mt_env;
1773 txn->mt_numdbs = env->me_numdbs;
1774 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1776 if (txn->mt_flags & MDB_TXN_RDONLY) {
1777 if (env->me_flags & MDB_ROFS) {
1778 i = mdb_env_pick_meta(env);
1779 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1780 txn->mt_u.reader = NULL;
1782 MDB_reader *r = pthread_getspecific(env->me_txkey);
1784 pid_t pid = env->me_pid;
1785 pthread_t tid = pthread_self();
1788 for (i=0; i<env->me_txns->mti_numreaders; i++)
1789 if (env->me_txns->mti_readers[i].mr_pid == 0)
1791 if (i == env->me_maxreaders) {
1792 UNLOCK_MUTEX_R(env);
1793 return MDB_READERS_FULL;
1795 env->me_txns->mti_readers[i].mr_pid = pid;
1796 env->me_txns->mti_readers[i].mr_tid = tid;
1797 if (i >= env->me_txns->mti_numreaders)
1798 env->me_txns->mti_numreaders = i+1;
1799 /* Save numreaders for un-mutexed mdb_env_close() */
1800 env->me_numreaders = env->me_txns->mti_numreaders;
1801 UNLOCK_MUTEX_R(env);
1802 r = &env->me_txns->mti_readers[i];
1803 if ((rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1804 env->me_txns->mti_readers[i].mr_pid = 0;
1808 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1809 txn->mt_u.reader = r;
1811 txn->mt_toggle = txn->mt_txnid & 1;
1812 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1816 txn->mt_txnid = env->me_txns->mti_txnid;
1817 txn->mt_toggle = txn->mt_txnid & 1;
1818 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1821 if (txn->mt_txnid == mdb_debug_start)
1824 txn->mt_u.dirty_list = env->me_dirty_list;
1825 txn->mt_u.dirty_list[0].mid = 0;
1826 txn->mt_free_pgs = env->me_free_pgs;
1827 txn->mt_free_pgs[0] = 0;
1831 /* Copy the DB info and flags */
1832 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1833 for (i=2; i<txn->mt_numdbs; i++)
1834 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1835 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1836 if (txn->mt_numdbs > 2)
1837 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1843 mdb_txn_renew(MDB_txn *txn)
1847 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1850 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1851 DPUTS("environment had fatal error, must shutdown!");
1855 rc = mdb_txn_renew0(txn);
1856 if (rc == MDB_SUCCESS) {
1857 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1858 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1859 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1865 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1870 if (env->me_flags & MDB_FATAL_ERROR) {
1871 DPUTS("environment had fatal error, must shutdown!");
1874 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1877 /* Nested transactions: Max 1 child, write txns only, no writemap */
1878 if (parent->mt_child ||
1879 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1880 (env->me_flags & MDB_WRITEMAP))
1885 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1886 if (!(flags & MDB_RDONLY))
1887 size += env->me_maxdbs * sizeof(MDB_cursor *);
1889 if ((txn = calloc(1, size)) == NULL) {
1890 DPRINTF("calloc: %s", strerror(ErrCode()));
1893 txn->mt_dbs = (MDB_db *)(txn+1);
1894 if (flags & MDB_RDONLY) {
1895 txn->mt_flags |= MDB_TXN_RDONLY;
1896 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1898 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1899 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1904 txn->mt_free_pgs = mdb_midl_alloc();
1905 if (!txn->mt_free_pgs) {
1909 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1910 if (!txn->mt_u.dirty_list) {
1911 free(txn->mt_free_pgs);
1915 txn->mt_txnid = parent->mt_txnid;
1916 txn->mt_toggle = parent->mt_toggle;
1917 txn->mt_u.dirty_list[0].mid = 0;
1918 txn->mt_free_pgs[0] = 0;
1919 txn->mt_next_pgno = parent->mt_next_pgno;
1920 parent->mt_child = txn;
1921 txn->mt_parent = parent;
1922 txn->mt_numdbs = parent->mt_numdbs;
1923 txn->mt_dbxs = parent->mt_dbxs;
1924 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1925 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1926 mdb_cursor_shadow(parent, txn);
1929 rc = mdb_txn_renew0(txn);
1935 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1936 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1937 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1943 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1944 * @param[in] txn the transaction handle to reset
1947 mdb_txn_reset0(MDB_txn *txn)
1949 MDB_env *env = txn->mt_env;
1951 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1952 if (!(env->me_flags & MDB_ROFS))
1953 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1959 /* close(free) all cursors */
1960 for (i=0; i<txn->mt_numdbs; i++) {
1961 if (txn->mt_cursors[i]) {
1963 while ((mc = txn->mt_cursors[i])) {
1964 txn->mt_cursors[i] = mc->mc_next;
1965 if (mc->mc_flags & C_ALLOCD)
1971 if (!(env->me_flags & MDB_WRITEMAP)) {
1972 /* return all dirty pages to dpage list */
1973 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1974 dp = txn->mt_u.dirty_list[i].mptr;
1975 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1976 mdb_page_free(txn->mt_env, dp);
1978 /* large pages just get freed directly */
1979 VGMEMP_FREE(txn->mt_env, dp);
1985 if (txn->mt_parent) {
1986 txn->mt_parent->mt_child = NULL;
1987 mdb_midl_free(txn->mt_free_pgs);
1988 free(txn->mt_u.dirty_list);
1991 if (mdb_midl_shrink(&txn->mt_free_pgs))
1992 env->me_free_pgs = txn->mt_free_pgs;
1995 while ((mop = txn->mt_env->me_pghead)) {
1996 txn->mt_env->me_pghead = mop->mo_next;
1999 txn->mt_env->me_pgfirst = 0;
2000 txn->mt_env->me_pglast = 0;
2003 /* The writer mutex was locked in mdb_txn_begin. */
2004 UNLOCK_MUTEX_W(env);
2009 mdb_txn_reset(MDB_txn *txn)
2014 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2015 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2016 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2018 mdb_txn_reset0(txn);
2022 mdb_txn_abort(MDB_txn *txn)
2027 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2028 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2029 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2032 mdb_txn_abort(txn->mt_child);
2034 mdb_txn_reset0(txn);
2039 mdb_txn_commit(MDB_txn *txn)
2047 pgno_t next, freecnt;
2050 assert(txn != NULL);
2051 assert(txn->mt_env != NULL);
2053 if (txn->mt_child) {
2054 mdb_txn_commit(txn->mt_child);
2055 txn->mt_child = NULL;
2060 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2061 if (txn->mt_numdbs > env->me_numdbs) {
2062 /* update the DB flags */
2064 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2065 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2072 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2073 DPUTS("error flag is set, can't commit");
2075 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2080 if (txn->mt_parent) {
2086 /* Merge (and close) our cursors with parent's */
2087 mdb_cursor_merge(txn);
2089 /* Update parent's DB table */
2090 ip = &txn->mt_parent->mt_dbs[2];
2091 jp = &txn->mt_dbs[2];
2092 for (i = 2; i < txn->mt_numdbs; i++) {
2093 if (ip->md_root != jp->md_root)
2097 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2099 /* Append our free list to parent's */
2100 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
2102 mdb_midl_free(txn->mt_free_pgs);
2104 /* Merge our dirty list with parent's */
2105 dst = txn->mt_parent->mt_u.dirty_list;
2106 src = txn->mt_u.dirty_list;
2107 x = mdb_mid2l_search(dst, src[1].mid);
2108 for (y=1; y<=src[0].mid; y++) {
2109 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
2113 dst[x].mptr = src[y].mptr;
2116 for (; y<=src[0].mid; y++) {
2117 if (++x >= MDB_IDL_UM_MAX) {
2119 return MDB_TXN_FULL;
2124 free(txn->mt_u.dirty_list);
2125 txn->mt_parent->mt_child = NULL;
2130 if (txn != env->me_txn) {
2131 DPUTS("attempt to commit unknown transaction");
2136 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2139 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2140 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2142 /* Update DB root pointers. Their pages have already been
2143 * touched so this is all in-place and cannot fail.
2145 if (txn->mt_numdbs > 2) {
2148 data.mv_size = sizeof(MDB_db);
2150 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2151 for (i = 2; i < txn->mt_numdbs; i++) {
2152 if (txn->mt_dbflags[i] & DB_DIRTY) {
2153 data.mv_data = &txn->mt_dbs[i];
2154 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2159 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2161 /* should only be one record now */
2162 if (env->me_pghead || env->me_pgfirst) {
2163 /* make sure first page of freeDB is touched and on freelist */
2164 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2165 if (rc && rc != MDB_NOTFOUND) {
2172 /* Delete IDLs we used from the free list */
2173 if (env->me_pgfirst) {
2178 key.mv_size = sizeof(cur);
2179 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
2182 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
2183 rc = mdb_cursor_del(&mc, 0);
2187 env->me_pgfirst = 0;
2191 /* save to free list */
2193 freecnt = txn->mt_free_pgs[0];
2194 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2197 /* make sure last page of freeDB is touched and on freelist */
2198 key.mv_size = MDB_MAXKEYSIZE+1;
2200 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2201 if (rc && rc != MDB_NOTFOUND)
2207 MDB_IDL idl = txn->mt_free_pgs;
2208 mdb_midl_sort(txn->mt_free_pgs);
2209 DPRINTF("IDL write txn %zu root %zu num %zu",
2210 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2211 for (i=1; i<=idl[0]; i++) {
2212 DPRINTF("IDL %zu", idl[i]);
2216 /* write to last page of freeDB */
2217 key.mv_size = sizeof(pgno_t);
2218 key.mv_data = &txn->mt_txnid;
2219 data.mv_data = txn->mt_free_pgs;
2220 /* The free list can still grow during this call,
2221 * despite the pre-emptive touches above. So check
2222 * and make sure the entire thing got written.
2225 freecnt = txn->mt_free_pgs[0];
2226 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2227 mdb_midl_sort(txn->mt_free_pgs);
2228 rc = mdb_cursor_put(&mc, &key, &data, 0);
2231 } while (freecnt != txn->mt_free_pgs[0]);
2233 /* should only be one record now */
2235 if (env->me_pghead) {
2241 mop = env->me_pghead;
2243 key.mv_size = sizeof(id);
2245 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2246 data.mv_data = mop->mo_pages;
2247 orig = mop->mo_pages[0];
2248 /* These steps may grow the freelist again
2249 * due to freed overflow pages...
2251 rc = mdb_cursor_put(&mc, &key, &data, 0);
2254 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2255 /* could have been used again here */
2256 if (mop->mo_pages[0] != orig) {
2257 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2258 data.mv_data = mop->mo_pages;
2260 rc = mdb_cursor_put(&mc, &key, &data, 0);
2265 /* was completely used up */
2266 rc = mdb_cursor_del(&mc, 0);
2272 env->me_pgfirst = 0;
2276 /* Check for growth of freelist again */
2277 if (freecnt != txn->mt_free_pgs[0])
2280 if (env->me_pghead) {
2281 free(env->me_pghead);
2282 env->me_pghead = NULL;
2285 while (env->me_pgfree) {
2286 MDB_oldpages *mop = env->me_pgfree;
2287 env->me_pgfree = mop->mo_next;
2291 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2292 if (mdb_midl_shrink(&txn->mt_free_pgs))
2293 env->me_free_pgs = txn->mt_free_pgs;
2300 if (env->me_flags & MDB_WRITEMAP) {
2301 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2302 dp = txn->mt_u.dirty_list[i].mptr;
2303 /* clear dirty flag */
2304 dp->mp_flags &= ~P_DIRTY;
2305 txn->mt_u.dirty_list[i].mid = 0;
2307 txn->mt_u.dirty_list[0].mid = 0;
2311 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2317 /* Windows actually supports scatter/gather I/O, but only on
2318 * unbuffered file handles. Since we're relying on the OS page
2319 * cache for all our data, that's self-defeating. So we just
2320 * write pages one at a time. We use the ov structure to set
2321 * the write offset, to at least save the overhead of a Seek
2325 memset(&ov, 0, sizeof(ov));
2326 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2328 dp = txn->mt_u.dirty_list[i].mptr;
2329 DPRINTF("committing page %zu", dp->mp_pgno);
2330 size = dp->mp_pgno * env->me_psize;
2331 ov.Offset = size & 0xffffffff;
2332 ov.OffsetHigh = size >> 16;
2333 ov.OffsetHigh >>= 16;
2334 /* clear dirty flag */
2335 dp->mp_flags &= ~P_DIRTY;
2336 wsize = env->me_psize;
2337 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2338 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2341 DPRINTF("WriteFile: %d", n);
2348 struct iovec iov[MDB_COMMIT_PAGES];
2352 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2353 dp = txn->mt_u.dirty_list[i].mptr;
2354 if (dp->mp_pgno != next) {
2356 rc = writev(env->me_fd, iov, n);
2360 DPUTS("short write, filesystem full?");
2362 DPRINTF("writev: %s", strerror(n));
2369 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2372 DPRINTF("committing page %zu", dp->mp_pgno);
2373 iov[n].iov_len = env->me_psize;
2374 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2375 iov[n].iov_base = (char *)dp;
2376 size += iov[n].iov_len;
2377 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2378 /* clear dirty flag */
2379 dp->mp_flags &= ~P_DIRTY;
2380 if (++n >= MDB_COMMIT_PAGES) {
2390 rc = writev(env->me_fd, iov, n);
2394 DPUTS("short write, filesystem full?");
2396 DPRINTF("writev: %s", strerror(n));
2403 /* Drop the dirty pages.
2405 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2406 dp = txn->mt_u.dirty_list[i].mptr;
2407 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2408 mdb_page_free(txn->mt_env, dp);
2410 VGMEMP_FREE(txn->mt_env, dp);
2413 txn->mt_u.dirty_list[i].mid = 0;
2415 txn->mt_u.dirty_list[0].mid = 0;
2418 if ((n = mdb_env_sync(env, 0)) != 0 ||
2419 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2426 if (txn->mt_numdbs > env->me_numdbs) {
2427 /* update the DB flags */
2429 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2430 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2434 UNLOCK_MUTEX_W(env);
2440 /** Read the environment parameters of a DB environment before
2441 * mapping it into memory.
2442 * @param[in] env the environment handle
2443 * @param[out] meta address of where to store the meta information
2444 * @return 0 on success, non-zero on failure.
2447 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2454 /* We don't know the page size yet, so use a minimum value.
2455 * Read both meta pages so we can use the latest one.
2458 for (i=0; i<2; i++) {
2460 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2462 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2467 else if (rc != MDB_PAGESIZE) {
2471 DPRINTF("read: %s", strerror(err));
2475 p = (MDB_page *)&pbuf;
2477 if (!F_ISSET(p->mp_flags, P_META)) {
2478 DPRINTF("page %zu not a meta page", p->mp_pgno);
2483 if (m->mm_magic != MDB_MAGIC) {
2484 DPUTS("meta has invalid magic");
2488 if (m->mm_version != MDB_VERSION) {
2489 DPRINTF("database is version %u, expected version %u",
2490 m->mm_version, MDB_VERSION);
2491 return MDB_VERSION_MISMATCH;
2495 if (m->mm_txnid > meta->mm_txnid)
2496 memcpy(meta, m, sizeof(*m));
2498 memcpy(meta, m, sizeof(*m));
2500 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2502 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2510 /** Write the environment parameters of a freshly created DB environment.
2511 * @param[in] env the environment handle
2512 * @param[out] meta address of where to store the meta information
2513 * @return 0 on success, non-zero on failure.
2516 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2523 DPUTS("writing new meta page");
2525 GET_PAGESIZE(psize);
2527 meta->mm_magic = MDB_MAGIC;
2528 meta->mm_version = MDB_VERSION;
2529 meta->mm_mapsize = env->me_mapsize;
2530 meta->mm_psize = psize;
2531 meta->mm_last_pg = 1;
2532 meta->mm_flags = env->me_flags & 0xffff;
2533 meta->mm_flags |= MDB_INTEGERKEY;
2534 meta->mm_dbs[0].md_root = P_INVALID;
2535 meta->mm_dbs[1].md_root = P_INVALID;
2537 p = calloc(2, psize);
2539 p->mp_flags = P_META;
2542 memcpy(m, meta, sizeof(*meta));
2544 q = (MDB_page *)((char *)p + psize);
2547 q->mp_flags = P_META;
2550 memcpy(m, meta, sizeof(*meta));
2555 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2556 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2557 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2560 lseek(env->me_fd, 0, SEEK_SET);
2561 rc = write(env->me_fd, p, psize * 2);
2562 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2568 /** Update the environment info to commit a transaction.
2569 * @param[in] txn the transaction that's being committed
2570 * @return 0 on success, non-zero on failure.
2573 mdb_env_write_meta(MDB_txn *txn)
2576 MDB_meta meta, metab, *mp;
2578 int rc, len, toggle;
2585 assert(txn != NULL);
2586 assert(txn->mt_env != NULL);
2588 toggle = !txn->mt_toggle;
2589 DPRINTF("writing meta page %d for root page %zu",
2590 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2593 mp = env->me_metas[toggle];
2595 if (env->me_flags & MDB_WRITEMAP) {
2596 /* Persist any increases of mapsize config */
2597 if (env->me_mapsize > mp->mm_mapsize)
2598 mp->mm_mapsize = env->me_mapsize;
2599 mp->mm_dbs[0] = txn->mt_dbs[0];
2600 mp->mm_dbs[1] = txn->mt_dbs[1];
2601 mp->mm_last_pg = txn->mt_next_pgno - 1;
2602 mp->mm_txnid = txn->mt_txnid;
2603 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2604 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2607 ptr += env->me_psize;
2608 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2615 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2616 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2618 ptr = (char *)&meta;
2619 if (env->me_mapsize > mp->mm_mapsize) {
2620 /* Persist any increases of mapsize config */
2621 meta.mm_mapsize = env->me_mapsize;
2622 off = offsetof(MDB_meta, mm_mapsize);
2624 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2626 len = sizeof(MDB_meta) - off;
2629 meta.mm_dbs[0] = txn->mt_dbs[0];
2630 meta.mm_dbs[1] = txn->mt_dbs[1];
2631 meta.mm_last_pg = txn->mt_next_pgno - 1;
2632 meta.mm_txnid = txn->mt_txnid;
2635 off += env->me_psize;
2638 /* Write to the SYNC fd */
2639 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2640 env->me_fd : env->me_mfd;
2643 memset(&ov, 0, sizeof(ov));
2645 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2648 rc = pwrite(mfd, ptr, len, off);
2653 DPUTS("write failed, disk error?");
2654 /* On a failure, the pagecache still contains the new data.
2655 * Write some old data back, to prevent it from being used.
2656 * Use the non-SYNC fd; we know it will fail anyway.
2658 meta.mm_last_pg = metab.mm_last_pg;
2659 meta.mm_txnid = metab.mm_txnid;
2661 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2663 r2 = pwrite(env->me_fd, ptr, len, off);
2666 env->me_flags |= MDB_FATAL_ERROR;
2670 /* Memory ordering issues are irrelevant; since the entire writer
2671 * is wrapped by wmutex, all of these changes will become visible
2672 * after the wmutex is unlocked. Since the DB is multi-version,
2673 * readers will get consistent data regardless of how fresh or
2674 * how stale their view of these values is.
2676 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2681 /** Check both meta pages to see which one is newer.
2682 * @param[in] env the environment handle
2683 * @return meta toggle (0 or 1).
2686 mdb_env_pick_meta(const MDB_env *env)
2688 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2692 mdb_env_create(MDB_env **env)
2696 e = calloc(1, sizeof(MDB_env));
2700 e->me_free_pgs = mdb_midl_alloc();
2701 if (!e->me_free_pgs) {
2705 e->me_maxreaders = DEFAULT_READERS;
2707 e->me_fd = INVALID_HANDLE_VALUE;
2708 e->me_lfd = INVALID_HANDLE_VALUE;
2709 e->me_mfd = INVALID_HANDLE_VALUE;
2710 #ifdef MDB_USE_POSIX_SEM
2711 e->me_rmutex = SEM_FAILED;
2712 e->me_wmutex = SEM_FAILED;
2714 e->me_pid = getpid();
2715 VGMEMP_CREATE(e,0,0);
2721 mdb_env_set_mapsize(MDB_env *env, size_t size)
2725 env->me_mapsize = size;
2727 env->me_maxpg = env->me_mapsize / env->me_psize;
2732 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2736 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2741 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2743 if (env->me_map || readers < 1)
2745 env->me_maxreaders = readers;
2750 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2752 if (!env || !readers)
2754 *readers = env->me_maxreaders;
2758 /** Further setup required for opening an MDB environment
2761 mdb_env_open2(MDB_env *env)
2763 unsigned int flags = env->me_flags;
2764 int i, newenv = 0, prot;
2768 memset(&meta, 0, sizeof(meta));
2770 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2773 DPUTS("new mdbenv");
2777 /* Was a mapsize configured? */
2778 if (!env->me_mapsize) {
2779 /* If this is a new environment, take the default,
2780 * else use the size recorded in the existing env.
2782 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2783 } else if (env->me_mapsize < meta.mm_mapsize) {
2784 /* If the configured size is smaller, make sure it's
2785 * still big enough. Silently round up to minimum if not.
2787 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2788 if (env->me_mapsize < minsize)
2789 env->me_mapsize = minsize;
2795 LONG sizelo, sizehi;
2796 sizelo = env->me_mapsize & 0xffffffff;
2797 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2799 /* Windows won't create mappings for zero length files.
2800 * Just allocate the maxsize right now.
2803 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2804 if (!SetEndOfFile(env->me_fd))
2806 SetFilePointer(env->me_fd, 0, NULL, 0);
2808 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2809 PAGE_READWRITE : PAGE_READONLY,
2810 sizehi, sizelo, NULL);
2813 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2814 FILE_MAP_WRITE : FILE_MAP_READ,
2815 0, 0, env->me_mapsize, meta.mm_address);
2823 if (flags & MDB_WRITEMAP) {
2825 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2828 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2830 if (env->me_map == MAP_FAILED) {
2837 if (flags & MDB_FIXEDMAP)
2838 meta.mm_address = env->me_map;
2839 i = mdb_env_init_meta(env, &meta);
2840 if (i != MDB_SUCCESS) {
2843 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2844 /* Can happen because the address argument to mmap() is just a
2845 * hint. mmap() can pick another, e.g. if the range is in use.
2846 * The MAP_FIXED flag would prevent that, but then mmap could
2847 * instead unmap existing pages to make room for the new map.
2849 return EBUSY; /* TODO: Make a new MDB_* error code? */
2851 env->me_psize = meta.mm_psize;
2853 env->me_maxpg = env->me_mapsize / env->me_psize;
2855 p = (MDB_page *)env->me_map;
2856 env->me_metas[0] = METADATA(p);
2857 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2861 int toggle = mdb_env_pick_meta(env);
2862 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2864 DPRINTF("opened database version %u, pagesize %u",
2865 env->me_metas[0]->mm_version, env->me_psize);
2866 DPRINTF("using meta page %d", toggle);
2867 DPRINTF("depth: %u", db->md_depth);
2868 DPRINTF("entries: %zu", db->md_entries);
2869 DPRINTF("branch pages: %zu", db->md_branch_pages);
2870 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2871 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2872 DPRINTF("root: %zu", db->md_root);
2880 /** Release a reader thread's slot in the reader lock table.
2881 * This function is called automatically when a thread exits.
2882 * @param[in] ptr This points to the slot in the reader lock table.
2885 mdb_env_reader_dest(void *ptr)
2887 MDB_reader *reader = ptr;
2893 /** Junk for arranging thread-specific callbacks on Windows. This is
2894 * necessarily platform and compiler-specific. Windows supports up
2895 * to 1088 keys. Let's assume nobody opens more than 64 environments
2896 * in a single process, for now. They can override this if needed.
2898 #ifndef MAX_TLS_KEYS
2899 #define MAX_TLS_KEYS 64
2901 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2902 static int mdb_tls_nkeys;
2904 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2908 case DLL_PROCESS_ATTACH: break;
2909 case DLL_THREAD_ATTACH: break;
2910 case DLL_THREAD_DETACH:
2911 for (i=0; i<mdb_tls_nkeys; i++) {
2912 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2913 mdb_env_reader_dest(r);
2916 case DLL_PROCESS_DETACH: break;
2921 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2923 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2927 /* Force some symbol references.
2928 * _tls_used forces the linker to create the TLS directory if not already done
2929 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2931 #pragma comment(linker, "/INCLUDE:_tls_used")
2932 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2933 #pragma const_seg(".CRT$XLB")
2934 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2935 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2938 #pragma comment(linker, "/INCLUDE:__tls_used")
2939 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2940 #pragma data_seg(".CRT$XLB")
2941 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2943 #endif /* WIN 32/64 */
2944 #endif /* !__GNUC__ */
2947 /** Downgrade the exclusive lock on the region back to shared */
2949 mdb_env_share_locks(MDB_env *env, int *excl)
2951 int rc = 0, toggle = mdb_env_pick_meta(env);
2953 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2958 /* First acquire a shared lock. The Unlock will
2959 * then release the existing exclusive lock.
2961 memset(&ov, 0, sizeof(ov));
2962 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2965 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2971 struct flock lock_info;
2972 /* The shared lock replaces the existing lock */
2973 memset((void *)&lock_info, 0, sizeof(lock_info));
2974 lock_info.l_type = F_RDLCK;
2975 lock_info.l_whence = SEEK_SET;
2976 lock_info.l_start = 0;
2977 lock_info.l_len = 1;
2978 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2979 (rc = ErrCode()) == EINTR) ;
2980 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
2987 /** Try to get exlusive lock, otherwise shared.
2988 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
2991 mdb_env_excl_lock(MDB_env *env, int *excl)
2995 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2999 memset(&ov, 0, sizeof(ov));
3000 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3007 struct flock lock_info;
3008 memset((void *)&lock_info, 0, sizeof(lock_info));
3009 lock_info.l_type = F_WRLCK;
3010 lock_info.l_whence = SEEK_SET;
3011 lock_info.l_start = 0;
3012 lock_info.l_len = 1;
3013 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3014 (rc = ErrCode()) == EINTR) ;
3018 # ifdef MDB_USE_POSIX_SEM
3019 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3022 lock_info.l_type = F_RDLCK;
3023 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3024 (rc = ErrCode()) == EINTR) ;
3032 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3034 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3036 * @(#) $Revision: 5.1 $
3037 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3038 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3040 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3044 * Please do not copyright this code. This code is in the public domain.
3046 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3047 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3048 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3049 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3050 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3051 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3052 * PERFORMANCE OF THIS SOFTWARE.
3055 * chongo <Landon Curt Noll> /\oo/\
3056 * http://www.isthe.com/chongo/
3058 * Share and Enjoy! :-)
3061 typedef unsigned long long mdb_hash_t;
3062 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3064 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3065 * @param[in] str string to hash
3066 * @param[in] hval initial value for hash
3067 * @return 64 bit hash
3069 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3070 * hval arg on the first call.
3073 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3075 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3076 unsigned char *end = s + val->mv_size;
3078 * FNV-1a hash each octet of the string
3081 /* xor the bottom with the current octet */
3082 hval ^= (mdb_hash_t)*s++;
3084 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3085 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3086 (hval << 7) + (hval << 8) + (hval << 40);
3088 /* return our new hash value */
3092 /** Hash the string and output the hash in hex.
3093 * @param[in] str string to hash
3094 * @param[out] hexbuf an array of 17 chars to hold the hash
3097 mdb_hash_hex(MDB_val *val, char *hexbuf)
3100 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3101 for (i=0; i<8; i++) {
3102 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3108 /** Open and/or initialize the lock region for the environment.
3109 * @param[in] env The MDB environment.
3110 * @param[in] lpath The pathname of the file used for the lock region.
3111 * @param[in] mode The Unix permissions for the file, if we create it.
3112 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3113 * @return 0 on success, non-zero on failure.
3116 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3124 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3125 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3126 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
3128 if (rc == ERROR_WRITE_PROTECT && (env->me_flags & MDB_RDONLY)) {
3129 env->me_flags |= MDB_ROFS;
3134 /* Try to get exclusive lock. If we succeed, then
3135 * nobody is using the lock region and we should initialize it.
3137 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3138 size = GetFileSize(env->me_lfd, NULL);
3144 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
3146 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3147 env->me_flags |= MDB_ROFS;
3152 /* Lose record locks when exec*() */
3153 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3154 fcntl(env->me_lfd, F_SETFD, fdflags);
3156 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
3157 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1) {
3159 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3160 env->me_flags |= MDB_ROFS;
3167 /* Try to get exclusive lock. If we succeed, then
3168 * nobody is using the lock region and we should initialize it.
3170 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3172 size = lseek(env->me_lfd, 0, SEEK_END);
3174 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3175 if (size < rsize && *excl > 0) {
3177 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3178 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3180 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3184 size = rsize - sizeof(MDB_txninfo);
3185 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3190 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3192 if (!mh) goto fail_errno;
3193 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3195 if (!env->me_txns) goto fail_errno;
3197 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3199 if (m == MAP_FAILED) goto fail_errno;
3205 BY_HANDLE_FILE_INFORMATION stbuf;
3214 if (!mdb_sec_inited) {
3215 InitializeSecurityDescriptor(&mdb_null_sd,
3216 SECURITY_DESCRIPTOR_REVISION);
3217 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3218 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3219 mdb_all_sa.bInheritHandle = FALSE;
3220 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3223 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3224 idbuf.volume = stbuf.dwVolumeSerialNumber;
3225 idbuf.nhigh = stbuf.nFileIndexHigh;
3226 idbuf.nlow = stbuf.nFileIndexLow;
3227 val.mv_data = &idbuf;
3228 val.mv_size = sizeof(idbuf);
3229 mdb_hash_hex(&val, hexbuf);
3230 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3231 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3232 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3233 if (!env->me_rmutex) goto fail_errno;
3234 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3235 if (!env->me_wmutex) goto fail_errno;
3236 #elif defined(MDB_USE_POSIX_SEM)
3245 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3246 idbuf.dev = stbuf.st_dev;
3247 idbuf.ino = stbuf.st_ino;
3248 val.mv_data = &idbuf;
3249 val.mv_size = sizeof(idbuf);
3250 mdb_hash_hex(&val, hexbuf);
3251 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3252 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3253 /* Clean up after a previous run, if needed: Try to
3254 * remove both semaphores before doing anything else.
3256 sem_unlink(env->me_txns->mti_rmname);
3257 sem_unlink(env->me_txns->mti_wmname);
3258 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3259 O_CREAT|O_EXCL, mode, 1);
3260 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3261 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3262 O_CREAT|O_EXCL, mode, 1);
3263 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3264 #else /* MDB_USE_POSIX_SEM */
3265 pthread_mutexattr_t mattr;
3267 if ((rc = pthread_mutexattr_init(&mattr))
3268 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3269 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3270 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3272 pthread_mutexattr_destroy(&mattr);
3273 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3275 env->me_txns->mti_version = MDB_VERSION;
3276 env->me_txns->mti_magic = MDB_MAGIC;
3277 env->me_txns->mti_txnid = 0;
3278 env->me_txns->mti_numreaders = 0;
3281 if (env->me_txns->mti_magic != MDB_MAGIC) {
3282 DPUTS("lock region has invalid magic");
3286 if (env->me_txns->mti_version != MDB_VERSION) {
3287 DPRINTF("lock region is version %u, expected version %u",
3288 env->me_txns->mti_version, MDB_VERSION);
3289 rc = MDB_VERSION_MISMATCH;
3293 if (rc != EACCES && rc != EAGAIN) {
3297 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3298 if (!env->me_rmutex) goto fail_errno;
3299 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3300 if (!env->me_wmutex) goto fail_errno;
3301 #elif defined(MDB_USE_POSIX_SEM)
3302 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3303 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3304 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3305 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3316 /** The name of the lock file in the DB environment */
3317 #define LOCKNAME "/lock.mdb"
3318 /** The name of the data file in the DB environment */
3319 #define DATANAME "/data.mdb"
3320 /** The suffix of the lock file when no subdir is used */
3321 #define LOCKSUFF "-lock"
3322 /** Only a subset of the @ref mdb_env flags can be changed
3323 * at runtime. Changing other flags requires closing the
3324 * environment and re-opening it with the new flags.
3326 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3327 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP)
3330 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3332 int oflags, rc, len, excl;
3333 char *lpath, *dpath;
3335 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3339 if (flags & MDB_NOSUBDIR) {
3340 rc = len + sizeof(LOCKSUFF) + len + 1;
3342 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3347 if (flags & MDB_NOSUBDIR) {
3348 dpath = lpath + len + sizeof(LOCKSUFF);
3349 sprintf(lpath, "%s" LOCKSUFF, path);
3350 strcpy(dpath, path);
3352 dpath = lpath + len + sizeof(LOCKNAME);
3353 sprintf(lpath, "%s" LOCKNAME, path);
3354 sprintf(dpath, "%s" DATANAME, path);
3357 flags |= env->me_flags;
3358 /* silently ignore WRITEMAP if we're only getting read access */
3359 if (F_ISSET(flags, MDB_RDONLY|MDB_WRITEMAP))
3360 flags ^= MDB_WRITEMAP;
3361 env->me_flags = flags |= MDB_ENV_ACTIVE;
3363 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3368 if (F_ISSET(flags, MDB_RDONLY)) {
3369 oflags = GENERIC_READ;
3370 len = OPEN_EXISTING;
3372 oflags = GENERIC_READ|GENERIC_WRITE;
3375 mode = FILE_ATTRIBUTE_NORMAL;
3376 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3377 NULL, len, mode, NULL);
3379 if (F_ISSET(flags, MDB_RDONLY))
3382 oflags = O_RDWR | O_CREAT;
3384 env->me_fd = open(dpath, oflags, mode);
3386 if (env->me_fd == INVALID_HANDLE_VALUE) {
3391 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3392 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3393 env->me_mfd = env->me_fd;
3395 /* Synchronous fd for meta writes. Needed even with
3396 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3399 env->me_mfd = CreateFile(dpath, oflags,
3400 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3401 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3403 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3405 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3410 DPRINTF("opened dbenv %p", (void *) env);
3411 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3414 env->me_numdbs = 2; /* this notes that me_txkey was set */
3416 /* Windows TLS callbacks need help finding their TLS info. */
3417 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3418 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3425 rc = mdb_env_share_locks(env, &excl);
3429 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3430 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3431 env->me_path = strdup(path);
3432 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3438 mdb_env_close0(env, excl);
3444 /** Destroy resources from mdb_env_open() and clear our readers */
3446 mdb_env_close0(MDB_env *env, int excl)
3450 if (!(env->me_flags & MDB_ENV_ACTIVE))
3453 free(env->me_dbflags);
3457 if (env->me_numdbs) {
3458 pthread_key_delete(env->me_txkey);
3460 /* Delete our key from the global list */
3461 for (i=0; i<mdb_tls_nkeys; i++)
3462 if (mdb_tls_keys[i] == env->me_txkey) {
3463 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3471 munmap(env->me_map, env->me_mapsize);
3473 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3475 if (env->me_fd != INVALID_HANDLE_VALUE)
3478 pid_t pid = env->me_pid;
3479 /* Clearing readers is done in this function because
3480 * me_txkey with its destructor must be disabled first.
3482 for (i = env->me_numreaders; --i >= 0; )
3483 if (env->me_txns->mti_readers[i].mr_pid == pid)
3484 env->me_txns->mti_readers[i].mr_pid = 0;
3486 if (env->me_rmutex) {
3487 CloseHandle(env->me_rmutex);
3488 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3490 /* Windows automatically destroys the mutexes when
3491 * the last handle closes.
3493 #elif defined(MDB_USE_POSIX_SEM)
3494 if (env->me_rmutex != SEM_FAILED) {
3495 sem_close(env->me_rmutex);
3496 if (env->me_wmutex != SEM_FAILED)
3497 sem_close(env->me_wmutex);
3498 /* If we have the filelock: If we are the
3499 * only remaining user, clean up semaphores.
3502 mdb_env_excl_lock(env, &excl);
3504 sem_unlink(env->me_txns->mti_rmname);
3505 sem_unlink(env->me_txns->mti_wmname);
3509 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3511 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3514 /* Unlock the lockfile. Windows would have unlocked it
3515 * after closing anyway, but not necessarily at once.
3517 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3523 env->me_flags &= ~MDB_ENV_ACTIVE;
3527 mdb_env_copy(MDB_env *env, const char *path)
3529 MDB_txn *txn = NULL;
3533 HANDLE newfd = INVALID_HANDLE_VALUE;
3535 if (env->me_flags & MDB_NOSUBDIR) {
3536 lpath = (char *)path;
3539 len += sizeof(DATANAME);
3540 lpath = malloc(len);
3543 sprintf(lpath, "%s" DATANAME, path);
3546 /* The destination path must exist, but the destination file must not.
3547 * We don't want the OS to cache the writes, since the source data is
3548 * already in the OS cache.
3551 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3552 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3554 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3560 if (!(env->me_flags & MDB_NOSUBDIR))
3562 if (newfd == INVALID_HANDLE_VALUE) {
3567 #ifdef F_NOCACHE /* __APPLE__ */
3568 rc = fcntl(newfd, F_NOCACHE, 1);
3575 /* Do the lock/unlock of the reader mutex before starting the
3576 * write txn. Otherwise other read txns could block writers.
3578 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3582 if (!(env->me_flags & MDB_ROFS)) {
3583 /* We must start the actual read txn after blocking writers */
3584 mdb_txn_reset0(txn);
3586 /* Temporarily block writers until we snapshot the meta pages */
3589 rc = mdb_txn_renew0(txn);
3591 UNLOCK_MUTEX_W(env);
3596 wsize = env->me_psize * 2;
3600 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3601 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3604 rc = write(newfd, env->me_map, wsize);
3605 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3607 if (! (env->me_flags & MDB_ROFS))
3608 UNLOCK_MUTEX_W(env);
3613 ptr = env->me_map + wsize;
3614 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3615 #define MAX_WRITE 2147483648U
3619 if (wsize > MAX_WRITE)
3623 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3624 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3633 if (wsize > MAX_WRITE)
3637 wres = write(newfd, ptr, w2);
3638 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3647 if (newfd != INVALID_HANDLE_VALUE)
3654 mdb_env_close(MDB_env *env)
3661 VGMEMP_DESTROY(env);
3662 while ((dp = env->me_dpages) != NULL) {
3663 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3664 env->me_dpages = dp->mp_next;
3668 mdb_env_close0(env, 0);
3669 mdb_midl_free(env->me_free_pgs);
3673 /** Compare two items pointing at aligned size_t's */
3675 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3677 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3678 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3681 /** Compare two items pointing at aligned int's */
3683 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3685 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3686 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3689 /** Compare two items pointing at ints of unknown alignment.
3690 * Nodes and keys are guaranteed to be 2-byte aligned.
3693 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3695 #if BYTE_ORDER == LITTLE_ENDIAN
3696 unsigned short *u, *c;
3699 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3700 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3703 } while(!x && u > (unsigned short *)a->mv_data);
3706 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3710 /** Compare two items lexically */
3712 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3719 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3725 diff = memcmp(a->mv_data, b->mv_data, len);
3726 return diff ? diff : len_diff<0 ? -1 : len_diff;
3729 /** Compare two items in reverse byte order */
3731 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3733 const unsigned char *p1, *p2, *p1_lim;
3737 p1_lim = (const unsigned char *)a->mv_data;
3738 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3739 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3741 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3747 while (p1 > p1_lim) {
3748 diff = *--p1 - *--p2;
3752 return len_diff<0 ? -1 : len_diff;
3755 /** Search for key within a page, using binary search.
3756 * Returns the smallest entry larger or equal to the key.
3757 * If exactp is non-null, stores whether the found entry was an exact match
3758 * in *exactp (1 or 0).
3759 * Updates the cursor index with the index of the found entry.
3760 * If no entry larger or equal to the key is found, returns NULL.
3763 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3765 unsigned int i = 0, nkeys;
3768 MDB_page *mp = mc->mc_pg[mc->mc_top];
3769 MDB_node *node = NULL;
3774 nkeys = NUMKEYS(mp);
3779 COPY_PGNO(pgno, mp->mp_pgno);
3780 DPRINTF("searching %u keys in %s %spage %zu",
3781 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3788 low = IS_LEAF(mp) ? 0 : 1;
3790 cmp = mc->mc_dbx->md_cmp;
3792 /* Branch pages have no data, so if using integer keys,
3793 * alignment is guaranteed. Use faster mdb_cmp_int.
3795 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3796 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3803 nodekey.mv_size = mc->mc_db->md_pad;
3804 node = NODEPTR(mp, 0); /* fake */
3805 while (low <= high) {
3806 i = (low + high) >> 1;
3807 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3808 rc = cmp(key, &nodekey);
3809 DPRINTF("found leaf index %u [%s], rc = %i",
3810 i, DKEY(&nodekey), rc);
3819 while (low <= high) {
3820 i = (low + high) >> 1;
3822 node = NODEPTR(mp, i);
3823 nodekey.mv_size = NODEKSZ(node);
3824 nodekey.mv_data = NODEKEY(node);
3826 rc = cmp(key, &nodekey);
3829 DPRINTF("found leaf index %u [%s], rc = %i",
3830 i, DKEY(&nodekey), rc);
3832 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3833 i, DKEY(&nodekey), NODEPGNO(node), rc);
3844 if (rc > 0) { /* Found entry is less than the key. */
3845 i++; /* Skip to get the smallest entry larger than key. */
3847 node = NODEPTR(mp, i);
3850 *exactp = (rc == 0);
3851 /* store the key index */
3852 mc->mc_ki[mc->mc_top] = i;
3854 /* There is no entry larger or equal to the key. */
3857 /* nodeptr is fake for LEAF2 */
3863 mdb_cursor_adjust(MDB_cursor *mc, func)
3867 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3868 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3875 /** Pop a page off the top of the cursor's stack. */
3877 mdb_cursor_pop(MDB_cursor *mc)
3880 #ifndef MDB_DEBUG_SKIP
3881 MDB_page *top = mc->mc_pg[mc->mc_top];
3887 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3888 mc->mc_dbi, (void *) mc);
3892 /** Push a page onto the top of the cursor's stack. */
3894 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3896 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3897 mc->mc_dbi, (void *) mc);
3899 if (mc->mc_snum >= CURSOR_STACK) {
3900 assert(mc->mc_snum < CURSOR_STACK);
3901 return MDB_CURSOR_FULL;
3904 mc->mc_top = mc->mc_snum++;
3905 mc->mc_pg[mc->mc_top] = mp;
3906 mc->mc_ki[mc->mc_top] = 0;
3911 /** Find the address of the page corresponding to a given page number.
3912 * @param[in] txn the transaction for this access.
3913 * @param[in] pgno the page number for the page to retrieve.
3914 * @param[out] ret address of a pointer where the page's address will be stored.
3915 * @return 0 on success, non-zero on failure.
3918 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3922 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
3923 if (pgno < txn->mt_next_pgno)
3924 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3927 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3929 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3930 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3931 p = txn->mt_u.dirty_list[x].mptr;
3935 if (pgno < txn->mt_next_pgno)
3936 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3941 DPRINTF("page %zu not found", pgno);
3944 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3947 /** Search for the page a given key should be in.
3948 * Pushes parent pages on the cursor stack. This function continues a
3949 * search on a cursor that has already been initialized. (Usually by
3950 * #mdb_page_search() but also by #mdb_node_move().)
3951 * @param[in,out] mc the cursor for this operation.
3952 * @param[in] key the key to search for. If NULL, search for the lowest
3953 * page. (This is used by #mdb_cursor_first().)
3954 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3955 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3956 * @return 0 on success, non-zero on failure.
3959 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3961 MDB_page *mp = mc->mc_pg[mc->mc_top];
3966 while (IS_BRANCH(mp)) {
3970 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3971 assert(NUMKEYS(mp) > 1);
3972 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3974 if (key == NULL) /* Initialize cursor to first page. */
3976 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
3977 /* cursor to last page */
3981 node = mdb_node_search(mc, key, &exact);
3983 i = NUMKEYS(mp) - 1;
3985 i = mc->mc_ki[mc->mc_top];
3994 DPRINTF("following index %u for key [%s]",
3996 assert(i < NUMKEYS(mp));
3997 node = NODEPTR(mp, i);
3999 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
4002 mc->mc_ki[mc->mc_top] = i;
4003 if ((rc = mdb_cursor_push(mc, mp)))
4007 if ((rc = mdb_page_touch(mc)) != 0)
4009 mp = mc->mc_pg[mc->mc_top];
4014 DPRINTF("internal error, index points to a %02X page!?",
4016 return MDB_CORRUPTED;
4019 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4020 key ? DKEY(key) : NULL);
4025 /** Search for the page a given key should be in.
4026 * Pushes parent pages on the cursor stack. This function just sets up
4027 * the search; it finds the root page for \b mc's database and sets this
4028 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4029 * called to complete the search.
4030 * @param[in,out] mc the cursor for this operation.
4031 * @param[in] key the key to search for. If NULL, search for the lowest
4032 * page. (This is used by #mdb_cursor_first().)
4033 * @param[in] modify If true, visited pages are updated with new page numbers.
4034 * @return 0 on success, non-zero on failure.
4037 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4042 /* Make sure the txn is still viable, then find the root from
4043 * the txn's db table.
4045 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4046 DPUTS("transaction has failed, must abort");
4049 /* Make sure we're using an up-to-date root */
4050 if (mc->mc_dbi > MAIN_DBI) {
4051 if ((*mc->mc_dbflag & DB_STALE) ||
4052 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4054 unsigned char dbflag = 0;
4055 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4056 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4059 if (*mc->mc_dbflag & DB_STALE) {
4062 MDB_node *leaf = mdb_node_search(&mc2,
4063 &mc->mc_dbx->md_name, &exact);
4065 return MDB_NOTFOUND;
4066 mdb_node_read(mc->mc_txn, leaf, &data);
4067 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4069 if (flags & MDB_PS_MODIFY)
4071 *mc->mc_dbflag = dbflag;
4074 root = mc->mc_db->md_root;
4076 if (root == P_INVALID) { /* Tree is empty. */
4077 DPUTS("tree is empty");
4078 return MDB_NOTFOUND;
4083 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4084 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
4090 DPRINTF("db %u root page %zu has flags 0x%X",
4091 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4093 if (flags & MDB_PS_MODIFY) {
4094 if ((rc = mdb_page_touch(mc)))
4098 if (flags & MDB_PS_ROOTONLY)
4101 return mdb_page_search_root(mc, key, flags);
4104 /** Return the data associated with a given node.
4105 * @param[in] txn The transaction for this operation.
4106 * @param[in] leaf The node being read.
4107 * @param[out] data Updated to point to the node's data.
4108 * @return 0 on success, non-zero on failure.
4111 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4113 MDB_page *omp; /* overflow page */
4117 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4118 data->mv_size = NODEDSZ(leaf);
4119 data->mv_data = NODEDATA(leaf);
4123 /* Read overflow data.
4125 data->mv_size = NODEDSZ(leaf);
4126 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4127 if ((rc = mdb_page_get(txn, pgno, &omp))) {
4128 DPRINTF("read overflow page %zu failed", pgno);
4131 data->mv_data = METADATA(omp);
4137 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4138 MDB_val *key, MDB_val *data)
4147 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4149 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4152 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4156 mdb_cursor_init(&mc, txn, dbi, &mx);
4157 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4160 /** Find a sibling for a page.
4161 * Replaces the page at the top of the cursor's stack with the
4162 * specified sibling, if one exists.
4163 * @param[in] mc The cursor for this operation.
4164 * @param[in] move_right Non-zero if the right sibling is requested,
4165 * otherwise the left sibling.
4166 * @return 0 on success, non-zero on failure.
4169 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4175 if (mc->mc_snum < 2) {
4176 return MDB_NOTFOUND; /* root has no siblings */
4180 DPRINTF("parent page is page %zu, index %u",
4181 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4183 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4184 : (mc->mc_ki[mc->mc_top] == 0)) {
4185 DPRINTF("no more keys left, moving to %s sibling",
4186 move_right ? "right" : "left");
4187 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4188 /* undo cursor_pop before returning */
4195 mc->mc_ki[mc->mc_top]++;
4197 mc->mc_ki[mc->mc_top]--;
4198 DPRINTF("just moving to %s index key %u",
4199 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4201 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4203 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4204 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4207 mdb_cursor_push(mc, mp);
4212 /** Move the cursor to the next data item. */
4214 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4220 if (mc->mc_flags & C_EOF) {
4221 return MDB_NOTFOUND;
4224 assert(mc->mc_flags & C_INITIALIZED);
4226 mp = mc->mc_pg[mc->mc_top];
4228 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4229 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4230 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4231 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4232 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4233 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4237 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4238 if (op == MDB_NEXT_DUP)
4239 return MDB_NOTFOUND;
4243 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4245 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4246 DPUTS("=====> move to next sibling page");
4247 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4248 mc->mc_flags |= C_EOF;
4249 mc->mc_flags &= ~C_INITIALIZED;
4250 return MDB_NOTFOUND;
4252 mp = mc->mc_pg[mc->mc_top];
4253 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4255 mc->mc_ki[mc->mc_top]++;
4257 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4258 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4261 key->mv_size = mc->mc_db->md_pad;
4262 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4266 assert(IS_LEAF(mp));
4267 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4269 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4270 mdb_xcursor_init1(mc, leaf);
4273 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4276 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4277 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4278 if (rc != MDB_SUCCESS)
4283 MDB_GET_KEY(leaf, key);
4287 /** Move the cursor to the previous data item. */
4289 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4295 assert(mc->mc_flags & C_INITIALIZED);
4297 mp = mc->mc_pg[mc->mc_top];
4299 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4300 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4301 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4302 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4303 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4304 if (op != MDB_PREV || rc == MDB_SUCCESS)
4307 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4308 if (op == MDB_PREV_DUP)
4309 return MDB_NOTFOUND;
4314 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4316 if (mc->mc_ki[mc->mc_top] == 0) {
4317 DPUTS("=====> move to prev sibling page");
4318 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4319 mc->mc_flags &= ~C_INITIALIZED;
4320 return MDB_NOTFOUND;
4322 mp = mc->mc_pg[mc->mc_top];
4323 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4324 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4326 mc->mc_ki[mc->mc_top]--;
4328 mc->mc_flags &= ~C_EOF;
4330 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4331 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4334 key->mv_size = mc->mc_db->md_pad;
4335 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4339 assert(IS_LEAF(mp));
4340 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4342 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4343 mdb_xcursor_init1(mc, leaf);
4346 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4349 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4350 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4351 if (rc != MDB_SUCCESS)
4356 MDB_GET_KEY(leaf, key);
4360 /** Set the cursor on a specific data item. */
4362 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4363 MDB_cursor_op op, int *exactp)
4367 MDB_node *leaf = NULL;
4372 assert(key->mv_size > 0);
4374 /* See if we're already on the right page */
4375 if (mc->mc_flags & C_INITIALIZED) {
4378 mp = mc->mc_pg[mc->mc_top];
4380 mc->mc_ki[mc->mc_top] = 0;
4381 return MDB_NOTFOUND;
4383 if (mp->mp_flags & P_LEAF2) {
4384 nodekey.mv_size = mc->mc_db->md_pad;
4385 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4387 leaf = NODEPTR(mp, 0);
4388 MDB_GET_KEY(leaf, &nodekey);
4390 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4392 /* Probably happens rarely, but first node on the page
4393 * was the one we wanted.
4395 mc->mc_ki[mc->mc_top] = 0;
4402 unsigned int nkeys = NUMKEYS(mp);
4404 if (mp->mp_flags & P_LEAF2) {
4405 nodekey.mv_data = LEAF2KEY(mp,
4406 nkeys-1, nodekey.mv_size);
4408 leaf = NODEPTR(mp, nkeys-1);
4409 MDB_GET_KEY(leaf, &nodekey);
4411 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4413 /* last node was the one we wanted */
4414 mc->mc_ki[mc->mc_top] = nkeys-1;
4420 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4421 /* This is definitely the right page, skip search_page */
4422 if (mp->mp_flags & P_LEAF2) {
4423 nodekey.mv_data = LEAF2KEY(mp,
4424 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4426 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4427 MDB_GET_KEY(leaf, &nodekey);
4429 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4431 /* current node was the one we wanted */
4441 /* If any parents have right-sibs, search.
4442 * Otherwise, there's nothing further.
4444 for (i=0; i<mc->mc_top; i++)
4446 NUMKEYS(mc->mc_pg[i])-1)
4448 if (i == mc->mc_top) {
4449 /* There are no other pages */
4450 mc->mc_ki[mc->mc_top] = nkeys;
4451 return MDB_NOTFOUND;
4455 /* There are no other pages */
4456 mc->mc_ki[mc->mc_top] = 0;
4457 return MDB_NOTFOUND;
4461 rc = mdb_page_search(mc, key, 0);
4462 if (rc != MDB_SUCCESS)
4465 mp = mc->mc_pg[mc->mc_top];
4466 assert(IS_LEAF(mp));
4469 leaf = mdb_node_search(mc, key, exactp);
4470 if (exactp != NULL && !*exactp) {
4471 /* MDB_SET specified and not an exact match. */
4472 return MDB_NOTFOUND;
4476 DPUTS("===> inexact leaf not found, goto sibling");
4477 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4478 return rc; /* no entries matched */
4479 mp = mc->mc_pg[mc->mc_top];
4480 assert(IS_LEAF(mp));
4481 leaf = NODEPTR(mp, 0);
4485 mc->mc_flags |= C_INITIALIZED;
4486 mc->mc_flags &= ~C_EOF;
4489 key->mv_size = mc->mc_db->md_pad;
4490 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4494 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4495 mdb_xcursor_init1(mc, leaf);
4498 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4499 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4500 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4503 if (op == MDB_GET_BOTH) {
4509 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4510 if (rc != MDB_SUCCESS)
4513 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4515 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4517 rc = mc->mc_dbx->md_dcmp(data, &d2);
4519 if (op == MDB_GET_BOTH || rc > 0)
4520 return MDB_NOTFOUND;
4525 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4526 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4531 /* The key already matches in all other cases */
4532 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4533 MDB_GET_KEY(leaf, key);
4534 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4539 /** Move the cursor to the first item in the database. */
4541 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4546 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4547 rc = mdb_page_search(mc, NULL, 0);
4548 if (rc != MDB_SUCCESS)
4551 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4553 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4554 mc->mc_flags |= C_INITIALIZED;
4555 mc->mc_flags &= ~C_EOF;
4557 mc->mc_ki[mc->mc_top] = 0;
4559 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4560 key->mv_size = mc->mc_db->md_pad;
4561 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4566 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4567 mdb_xcursor_init1(mc, leaf);
4568 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4573 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4574 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4578 MDB_GET_KEY(leaf, key);
4582 /** Move the cursor to the last item in the database. */
4584 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4589 if (!(mc->mc_flags & C_EOF)) {
4591 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4594 lkey.mv_size = MDB_MAXKEYSIZE+1;
4595 lkey.mv_data = NULL;
4596 rc = mdb_page_search(mc, &lkey, 0);
4597 if (rc != MDB_SUCCESS)
4600 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4602 mc->mc_flags |= C_INITIALIZED|C_EOF;
4603 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4605 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4607 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4608 key->mv_size = mc->mc_db->md_pad;
4609 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4614 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4615 mdb_xcursor_init1(mc, leaf);
4616 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4621 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4622 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4627 MDB_GET_KEY(leaf, key);
4632 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4641 case MDB_GET_CURRENT:
4642 if (!(mc->mc_flags & C_INITIALIZED)) {
4645 MDB_page *mp = mc->mc_pg[mc->mc_top];
4647 mc->mc_ki[mc->mc_top] = 0;
4653 key->mv_size = mc->mc_db->md_pad;
4654 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4656 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4657 MDB_GET_KEY(leaf, key);
4659 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4660 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4662 rc = mdb_node_read(mc->mc_txn, leaf, data);
4669 case MDB_GET_BOTH_RANGE:
4670 if (data == NULL || mc->mc_xcursor == NULL) {
4678 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4680 } else if (op == MDB_SET_RANGE)
4681 rc = mdb_cursor_set(mc, key, data, op, NULL);
4683 rc = mdb_cursor_set(mc, key, data, op, &exact);
4685 case MDB_GET_MULTIPLE:
4687 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4688 !(mc->mc_flags & C_INITIALIZED)) {
4693 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4694 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4697 case MDB_NEXT_MULTIPLE:
4699 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4703 if (!(mc->mc_flags & C_INITIALIZED))
4704 rc = mdb_cursor_first(mc, key, data);
4706 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4707 if (rc == MDB_SUCCESS) {
4708 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4711 mx = &mc->mc_xcursor->mx_cursor;
4712 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4714 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4715 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4723 case MDB_NEXT_NODUP:
4724 if (!(mc->mc_flags & C_INITIALIZED))
4725 rc = mdb_cursor_first(mc, key, data);
4727 rc = mdb_cursor_next(mc, key, data, op);
4731 case MDB_PREV_NODUP:
4732 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4733 rc = mdb_cursor_last(mc, key, data);
4734 mc->mc_flags |= C_INITIALIZED;
4735 mc->mc_ki[mc->mc_top]++;
4737 rc = mdb_cursor_prev(mc, key, data, op);
4740 rc = mdb_cursor_first(mc, key, data);
4744 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4745 !(mc->mc_flags & C_INITIALIZED) ||
4746 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4750 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4753 rc = mdb_cursor_last(mc, key, data);
4757 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4758 !(mc->mc_flags & C_INITIALIZED) ||
4759 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4763 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4766 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4774 /** Touch all the pages in the cursor stack.
4775 * Makes sure all the pages are writable, before attempting a write operation.
4776 * @param[in] mc The cursor to operate on.
4779 mdb_cursor_touch(MDB_cursor *mc)
4783 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4786 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4787 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4788 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4791 *mc->mc_dbflag = DB_DIRTY;
4793 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4794 rc = mdb_page_touch(mc);
4798 mc->mc_top = mc->mc_snum-1;
4803 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4806 MDB_node *leaf = NULL;
4807 MDB_val xdata, *rdata, dkey;
4810 int do_sub = 0, insert = 0;
4811 unsigned int mcount = 0;
4815 char dbuf[MDB_MAXKEYSIZE+1];
4816 unsigned int nflags;
4819 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4822 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE)
4825 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4828 #if SIZE_MAX > MAXDATASIZE
4829 if (data->mv_size > MAXDATASIZE)
4833 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4834 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4838 if (flags == MDB_CURRENT) {
4839 if (!(mc->mc_flags & C_INITIALIZED))
4842 } else if (mc->mc_db->md_root == P_INVALID) {
4844 /* new database, write a root leaf page */
4845 DPUTS("allocating new root leaf page");
4846 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4850 mdb_cursor_push(mc, np);
4851 mc->mc_db->md_root = np->mp_pgno;
4852 mc->mc_db->md_depth++;
4853 *mc->mc_dbflag = DB_DIRTY;
4854 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4856 np->mp_flags |= P_LEAF2;
4857 mc->mc_flags |= C_INITIALIZED;
4863 if (flags & MDB_APPEND) {
4865 rc = mdb_cursor_last(mc, &k2, &d2);
4867 rc = mc->mc_dbx->md_cmp(key, &k2);
4870 mc->mc_ki[mc->mc_top]++;
4876 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4878 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4879 DPRINTF("duplicate key [%s]", DKEY(key));
4881 return MDB_KEYEXIST;
4883 if (rc && rc != MDB_NOTFOUND)
4887 /* Cursor is positioned, now make sure all pages are writable */
4888 rc2 = mdb_cursor_touch(mc);
4893 /* The key already exists */
4894 if (rc == MDB_SUCCESS) {
4895 /* there's only a key anyway, so this is a no-op */
4896 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4897 unsigned int ksize = mc->mc_db->md_pad;
4898 if (key->mv_size != ksize)
4900 if (flags == MDB_CURRENT) {
4901 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4902 memcpy(ptr, key->mv_data, ksize);
4907 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4910 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4911 /* Was a single item before, must convert now */
4913 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4914 /* Just overwrite the current item */
4915 if (flags == MDB_CURRENT)
4918 dkey.mv_size = NODEDSZ(leaf);
4919 dkey.mv_data = NODEDATA(leaf);
4920 #if UINT_MAX < SIZE_MAX
4921 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4922 #ifdef MISALIGNED_OK
4923 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4925 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4928 /* if data matches, ignore it */
4929 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4930 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4932 /* create a fake page for the dup items */
4933 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4934 dkey.mv_data = dbuf;
4935 fp = (MDB_page *)&pbuf;
4936 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4937 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4938 fp->mp_lower = PAGEHDRSZ;
4939 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4940 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4941 fp->mp_flags |= P_LEAF2;
4942 fp->mp_pad = data->mv_size;
4943 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4945 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4946 (dkey.mv_size & 1) + (data->mv_size & 1);
4948 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4951 xdata.mv_size = fp->mp_upper;
4956 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4957 /* See if we need to convert from fake page to subDB */
4959 unsigned int offset;
4962 fp = NODEDATA(leaf);
4963 if (flags == MDB_CURRENT) {
4965 fp->mp_flags |= P_DIRTY;
4966 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4967 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4971 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4972 offset = fp->mp_pad;
4973 if (SIZELEFT(fp) >= offset)
4975 offset *= 4; /* space for 4 more */
4977 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4979 offset += offset & 1;
4980 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4981 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4983 /* yes, convert it */
4985 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4986 dummy.md_pad = fp->mp_pad;
4987 dummy.md_flags = MDB_DUPFIXED;
4988 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4989 dummy.md_flags |= MDB_INTEGERKEY;
4992 dummy.md_branch_pages = 0;
4993 dummy.md_leaf_pages = 1;
4994 dummy.md_overflow_pages = 0;
4995 dummy.md_entries = NUMKEYS(fp);
4997 xdata.mv_size = sizeof(MDB_db);
4998 xdata.mv_data = &dummy;
4999 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5001 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5002 flags |= F_DUPDATA|F_SUBDATA;
5003 dummy.md_root = mp->mp_pgno;
5005 /* no, just grow it */
5007 xdata.mv_size = NODEDSZ(leaf) + offset;
5008 xdata.mv_data = &pbuf;
5009 mp = (MDB_page *)&pbuf;
5010 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5013 mp->mp_flags = fp->mp_flags | P_DIRTY;
5014 mp->mp_pad = fp->mp_pad;
5015 mp->mp_lower = fp->mp_lower;
5016 mp->mp_upper = fp->mp_upper + offset;
5018 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5020 nsize = NODEDSZ(leaf) - fp->mp_upper;
5021 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5022 for (i=0; i<NUMKEYS(fp); i++)
5023 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5025 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5029 /* data is on sub-DB, just store it */
5030 flags |= F_DUPDATA|F_SUBDATA;
5034 /* overflow page overwrites need special handling */
5035 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5038 int ovpages, dpages;
5040 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5041 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5042 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5043 mdb_page_get(mc->mc_txn, pg, &omp);
5044 /* Is the ov page writable and large enough? */
5045 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5046 /* yes, overwrite it. Note in this case we don't
5047 * bother to try shrinking the node if the new data
5048 * is smaller than the overflow threshold.
5050 if (F_ISSET(flags, MDB_RESERVE))
5051 data->mv_data = METADATA(omp);
5053 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5056 /* no, free ovpages */
5058 mc->mc_db->md_overflow_pages -= ovpages;
5059 for (i=0; i<ovpages; i++) {
5060 DPRINTF("freed ov page %zu", pg);
5061 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5065 } else if (NODEDSZ(leaf) == data->mv_size) {
5066 /* same size, just replace it. Note that we could
5067 * also reuse this node if the new data is smaller,
5068 * but instead we opt to shrink the node in that case.
5070 if (F_ISSET(flags, MDB_RESERVE))
5071 data->mv_data = NODEDATA(leaf);
5072 else if (data->mv_size)
5073 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5075 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5078 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5079 mc->mc_db->md_entries--;
5081 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5088 nflags = flags & NODE_ADD_FLAGS;
5089 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5090 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5091 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5092 nflags &= ~MDB_APPEND;
5094 nflags |= MDB_SPLIT_REPLACE;
5095 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5097 /* There is room already in this leaf page. */
5098 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5099 if (rc == 0 && !do_sub && insert) {
5100 /* Adjust other cursors pointing to mp */
5101 MDB_cursor *m2, *m3;
5102 MDB_dbi dbi = mc->mc_dbi;
5103 unsigned i = mc->mc_top;
5104 MDB_page *mp = mc->mc_pg[i];
5106 if (mc->mc_flags & C_SUB)
5109 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5110 if (mc->mc_flags & C_SUB)
5111 m3 = &m2->mc_xcursor->mx_cursor;
5114 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5115 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5122 if (rc != MDB_SUCCESS)
5123 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5125 /* Now store the actual data in the child DB. Note that we're
5126 * storing the user data in the keys field, so there are strict
5127 * size limits on dupdata. The actual data fields of the child
5128 * DB are all zero size.
5135 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5136 if (flags & MDB_CURRENT) {
5137 xflags = MDB_CURRENT;
5139 mdb_xcursor_init1(mc, leaf);
5140 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5142 /* converted, write the original data first */
5144 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5148 /* Adjust other cursors pointing to mp */
5150 unsigned i = mc->mc_top;
5151 MDB_page *mp = mc->mc_pg[i];
5153 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5154 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5155 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5156 mdb_xcursor_init1(m2, leaf);
5161 if (flags & MDB_APPENDDUP)
5162 xflags |= MDB_APPEND;
5163 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5164 if (flags & F_SUBDATA) {
5165 void *db = NODEDATA(leaf);
5166 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5169 /* sub-writes might have failed so check rc again.
5170 * Don't increment count if we just replaced an existing item.
5172 if (!rc && !(flags & MDB_CURRENT))
5173 mc->mc_db->md_entries++;
5174 if (flags & MDB_MULTIPLE) {
5176 if (mcount < data[1].mv_size) {
5177 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5178 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5184 /* If we succeeded and the key didn't exist before, make sure
5185 * the cursor is marked valid.
5188 mc->mc_flags |= C_INITIALIZED;
5193 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5198 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5201 if (!(mc->mc_flags & C_INITIALIZED))
5204 rc = mdb_cursor_touch(mc);
5208 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5210 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5211 if (flags != MDB_NODUPDATA) {
5212 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5213 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5215 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5216 /* If sub-DB still has entries, we're done */
5217 if (mc->mc_xcursor->mx_db.md_entries) {
5218 if (leaf->mn_flags & F_SUBDATA) {
5219 /* update subDB info */
5220 void *db = NODEDATA(leaf);
5221 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5223 /* shrink fake page */
5224 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5226 mc->mc_db->md_entries--;
5229 /* otherwise fall thru and delete the sub-DB */
5232 if (leaf->mn_flags & F_SUBDATA) {
5233 /* add all the child DB's pages to the free list */
5234 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5235 if (rc == MDB_SUCCESS) {
5236 mc->mc_db->md_entries -=
5237 mc->mc_xcursor->mx_db.md_entries;
5242 return mdb_cursor_del0(mc, leaf);
5245 /** Allocate and initialize new pages for a database.
5246 * @param[in] mc a cursor on the database being added to.
5247 * @param[in] flags flags defining what type of page is being allocated.
5248 * @param[in] num the number of pages to allocate. This is usually 1,
5249 * unless allocating overflow pages for a large record.
5250 * @param[out] mp Address of a page, or NULL on failure.
5251 * @return 0 on success, non-zero on failure.
5254 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5259 if ((rc = mdb_page_alloc(mc, num, &np)))
5261 DPRINTF("allocated new mpage %zu, page size %u",
5262 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5263 np->mp_flags = flags | P_DIRTY;
5264 np->mp_lower = PAGEHDRSZ;
5265 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5268 mc->mc_db->md_branch_pages++;
5269 else if (IS_LEAF(np))
5270 mc->mc_db->md_leaf_pages++;
5271 else if (IS_OVERFLOW(np)) {
5272 mc->mc_db->md_overflow_pages += num;
5280 /** Calculate the size of a leaf node.
5281 * The size depends on the environment's page size; if a data item
5282 * is too large it will be put onto an overflow page and the node
5283 * size will only include the key and not the data. Sizes are always
5284 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5285 * of the #MDB_node headers.
5286 * @param[in] env The environment handle.
5287 * @param[in] key The key for the node.
5288 * @param[in] data The data for the node.
5289 * @return The number of bytes needed to store the node.
5292 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5296 sz = LEAFSIZE(key, data);
5297 if (sz >= env->me_psize / MDB_MINKEYS) {
5298 /* put on overflow page */
5299 sz -= data->mv_size - sizeof(pgno_t);
5303 return sz + sizeof(indx_t);
5306 /** Calculate the size of a branch node.
5307 * The size should depend on the environment's page size but since
5308 * we currently don't support spilling large keys onto overflow
5309 * pages, it's simply the size of the #MDB_node header plus the
5310 * size of the key. Sizes are always rounded up to an even number
5311 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5312 * @param[in] env The environment handle.
5313 * @param[in] key The key for the node.
5314 * @return The number of bytes needed to store the node.
5317 mdb_branch_size(MDB_env *env, MDB_val *key)
5322 if (sz >= env->me_psize / MDB_MINKEYS) {
5323 /* put on overflow page */
5324 /* not implemented */
5325 /* sz -= key->size - sizeof(pgno_t); */
5328 return sz + sizeof(indx_t);
5331 /** Add a node to the page pointed to by the cursor.
5332 * @param[in] mc The cursor for this operation.
5333 * @param[in] indx The index on the page where the new node should be added.
5334 * @param[in] key The key for the new node.
5335 * @param[in] data The data for the new node, if any.
5336 * @param[in] pgno The page number, if adding a branch node.
5337 * @param[in] flags Flags for the node.
5338 * @return 0 on success, non-zero on failure. Possible errors are:
5340 * <li>ENOMEM - failed to allocate overflow pages for the node.
5341 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5342 * should never happen since all callers already calculate the
5343 * page's free space before calling this function.
5347 mdb_node_add(MDB_cursor *mc, indx_t indx,
5348 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5351 size_t node_size = NODESIZE;
5354 MDB_page *mp = mc->mc_pg[mc->mc_top];
5355 MDB_page *ofp = NULL; /* overflow page */
5358 assert(mp->mp_upper >= mp->mp_lower);
5360 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5361 IS_LEAF(mp) ? "leaf" : "branch",
5362 IS_SUBP(mp) ? "sub-" : "",
5363 mp->mp_pgno, indx, data ? data->mv_size : 0,
5364 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5367 /* Move higher keys up one slot. */
5368 int ksize = mc->mc_db->md_pad, dif;
5369 char *ptr = LEAF2KEY(mp, indx, ksize);
5370 dif = NUMKEYS(mp) - indx;
5372 memmove(ptr+ksize, ptr, dif*ksize);
5373 /* insert new key */
5374 memcpy(ptr, key->mv_data, ksize);
5376 /* Just using these for counting */
5377 mp->mp_lower += sizeof(indx_t);
5378 mp->mp_upper -= ksize - sizeof(indx_t);
5383 node_size += key->mv_size;
5387 if (F_ISSET(flags, F_BIGDATA)) {
5388 /* Data already on overflow page. */
5389 node_size += sizeof(pgno_t);
5390 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5391 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5393 /* Put data on overflow page. */
5394 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5395 data->mv_size, node_size+data->mv_size);
5396 node_size += sizeof(pgno_t);
5397 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5399 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5402 node_size += data->mv_size;
5405 node_size += node_size & 1;
5407 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5408 DPRINTF("not enough room in page %zu, got %u ptrs",
5409 mp->mp_pgno, NUMKEYS(mp));
5410 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5411 mp->mp_upper - mp->mp_lower);
5412 DPRINTF("node size = %zu", node_size);
5413 return MDB_PAGE_FULL;
5416 /* Move higher pointers up one slot. */
5417 for (i = NUMKEYS(mp); i > indx; i--)
5418 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5420 /* Adjust free space offsets. */
5421 ofs = mp->mp_upper - node_size;
5422 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5423 mp->mp_ptrs[indx] = ofs;
5425 mp->mp_lower += sizeof(indx_t);
5427 /* Write the node data. */
5428 node = NODEPTR(mp, indx);
5429 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5430 node->mn_flags = flags;
5432 SETDSZ(node,data->mv_size);
5437 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5442 if (F_ISSET(flags, F_BIGDATA))
5443 memcpy(node->mn_data + key->mv_size, data->mv_data,
5445 else if (F_ISSET(flags, MDB_RESERVE))
5446 data->mv_data = node->mn_data + key->mv_size;
5448 memcpy(node->mn_data + key->mv_size, data->mv_data,
5451 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5453 if (F_ISSET(flags, MDB_RESERVE))
5454 data->mv_data = METADATA(ofp);
5456 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5463 /** Delete the specified node from a page.
5464 * @param[in] mp The page to operate on.
5465 * @param[in] indx The index of the node to delete.
5466 * @param[in] ksize The size of a node. Only used if the page is
5467 * part of a #MDB_DUPFIXED database.
5470 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5473 indx_t i, j, numkeys, ptr;
5480 COPY_PGNO(pgno, mp->mp_pgno);
5481 DPRINTF("delete node %u on %s page %zu", indx,
5482 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5485 assert(indx < NUMKEYS(mp));
5488 int x = NUMKEYS(mp) - 1 - indx;
5489 base = LEAF2KEY(mp, indx, ksize);
5491 memmove(base, base + ksize, x * ksize);
5492 mp->mp_lower -= sizeof(indx_t);
5493 mp->mp_upper += ksize - sizeof(indx_t);
5497 node = NODEPTR(mp, indx);
5498 sz = NODESIZE + node->mn_ksize;
5500 if (F_ISSET(node->mn_flags, F_BIGDATA))
5501 sz += sizeof(pgno_t);
5503 sz += NODEDSZ(node);
5507 ptr = mp->mp_ptrs[indx];
5508 numkeys = NUMKEYS(mp);
5509 for (i = j = 0; i < numkeys; i++) {
5511 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5512 if (mp->mp_ptrs[i] < ptr)
5513 mp->mp_ptrs[j] += sz;
5518 base = (char *)mp + mp->mp_upper;
5519 memmove(base + sz, base, ptr - mp->mp_upper);
5521 mp->mp_lower -= sizeof(indx_t);
5525 /** Compact the main page after deleting a node on a subpage.
5526 * @param[in] mp The main page to operate on.
5527 * @param[in] indx The index of the subpage on the main page.
5530 mdb_node_shrink(MDB_page *mp, indx_t indx)
5537 indx_t i, numkeys, ptr;
5539 node = NODEPTR(mp, indx);
5540 sp = (MDB_page *)NODEDATA(node);
5541 osize = NODEDSZ(node);
5543 delta = sp->mp_upper - sp->mp_lower;
5544 SETDSZ(node, osize - delta);
5545 xp = (MDB_page *)((char *)sp + delta);
5547 /* shift subpage upward */
5549 nsize = NUMKEYS(sp) * sp->mp_pad;
5550 memmove(METADATA(xp), METADATA(sp), nsize);
5553 nsize = osize - sp->mp_upper;
5554 numkeys = NUMKEYS(sp);
5555 for (i=numkeys-1; i>=0; i--)
5556 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5558 xp->mp_upper = sp->mp_lower;
5559 xp->mp_lower = sp->mp_lower;
5560 xp->mp_flags = sp->mp_flags;
5561 xp->mp_pad = sp->mp_pad;
5562 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5564 /* shift lower nodes upward */
5565 ptr = mp->mp_ptrs[indx];
5566 numkeys = NUMKEYS(mp);
5567 for (i = 0; i < numkeys; i++) {
5568 if (mp->mp_ptrs[i] <= ptr)
5569 mp->mp_ptrs[i] += delta;
5572 base = (char *)mp + mp->mp_upper;
5573 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5574 mp->mp_upper += delta;
5577 /** Initial setup of a sorted-dups cursor.
5578 * Sorted duplicates are implemented as a sub-database for the given key.
5579 * The duplicate data items are actually keys of the sub-database.
5580 * Operations on the duplicate data items are performed using a sub-cursor
5581 * initialized when the sub-database is first accessed. This function does
5582 * the preliminary setup of the sub-cursor, filling in the fields that
5583 * depend only on the parent DB.
5584 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5587 mdb_xcursor_init0(MDB_cursor *mc)
5589 MDB_xcursor *mx = mc->mc_xcursor;
5591 mx->mx_cursor.mc_xcursor = NULL;
5592 mx->mx_cursor.mc_txn = mc->mc_txn;
5593 mx->mx_cursor.mc_db = &mx->mx_db;
5594 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5595 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5596 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5597 mx->mx_cursor.mc_snum = 0;
5598 mx->mx_cursor.mc_top = 0;
5599 mx->mx_cursor.mc_flags = C_SUB;
5600 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5601 mx->mx_dbx.md_dcmp = NULL;
5602 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5605 /** Final setup of a sorted-dups cursor.
5606 * Sets up the fields that depend on the data from the main cursor.
5607 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5608 * @param[in] node The data containing the #MDB_db record for the
5609 * sorted-dup database.
5612 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5614 MDB_xcursor *mx = mc->mc_xcursor;
5616 if (node->mn_flags & F_SUBDATA) {
5617 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5618 mx->mx_cursor.mc_pg[0] = 0;
5619 mx->mx_cursor.mc_snum = 0;
5620 mx->mx_cursor.mc_flags = C_SUB;
5622 MDB_page *fp = NODEDATA(node);
5623 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5624 mx->mx_db.md_flags = 0;
5625 mx->mx_db.md_depth = 1;
5626 mx->mx_db.md_branch_pages = 0;
5627 mx->mx_db.md_leaf_pages = 1;
5628 mx->mx_db.md_overflow_pages = 0;
5629 mx->mx_db.md_entries = NUMKEYS(fp);
5630 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5631 mx->mx_cursor.mc_snum = 1;
5632 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5633 mx->mx_cursor.mc_top = 0;
5634 mx->mx_cursor.mc_pg[0] = fp;
5635 mx->mx_cursor.mc_ki[0] = 0;
5636 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5637 mx->mx_db.md_flags = MDB_DUPFIXED;
5638 mx->mx_db.md_pad = fp->mp_pad;
5639 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5640 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5643 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5645 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5647 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5648 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5649 #if UINT_MAX < SIZE_MAX
5650 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5651 #ifdef MISALIGNED_OK
5652 mx->mx_dbx.md_cmp = mdb_cmp_long;
5654 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5659 /** Initialize a cursor for a given transaction and database. */
5661 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5666 mc->mc_db = &txn->mt_dbs[dbi];
5667 mc->mc_dbx = &txn->mt_dbxs[dbi];
5668 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5673 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5675 mc->mc_xcursor = mx;
5676 mdb_xcursor_init0(mc);
5678 mc->mc_xcursor = NULL;
5680 if (*mc->mc_dbflag & DB_STALE) {
5681 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5686 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5689 MDB_xcursor *mx = NULL;
5690 size_t size = sizeof(MDB_cursor);
5692 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5695 /* Allow read access to the freelist */
5696 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5699 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5700 size += sizeof(MDB_xcursor);
5702 if ((mc = malloc(size)) != NULL) {
5703 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5704 mx = (MDB_xcursor *)(mc + 1);
5706 mdb_cursor_init(mc, txn, dbi, mx);
5707 if (txn->mt_cursors) {
5708 mc->mc_next = txn->mt_cursors[dbi];
5709 txn->mt_cursors[dbi] = mc;
5711 mc->mc_flags |= C_ALLOCD;
5722 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5724 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5727 if (txn->mt_cursors)
5730 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5734 /* Return the count of duplicate data items for the current key */
5736 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5740 if (mc == NULL || countp == NULL)
5743 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5746 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5747 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5750 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5753 *countp = mc->mc_xcursor->mx_db.md_entries;
5759 mdb_cursor_close(MDB_cursor *mc)
5762 /* remove from txn, if tracked */
5763 if (mc->mc_txn->mt_cursors) {
5764 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5765 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5767 *prev = mc->mc_next;
5769 if (mc->mc_flags & C_ALLOCD)
5775 mdb_cursor_txn(MDB_cursor *mc)
5777 if (!mc) return NULL;
5782 mdb_cursor_dbi(MDB_cursor *mc)
5788 /** Replace the key for a node with a new key.
5789 * @param[in] mp The page containing the node to operate on.
5790 * @param[in] indx The index of the node to operate on.
5791 * @param[in] key The new key to use.
5792 * @return 0 on success, non-zero on failure.
5795 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5802 indx_t ptr, i, numkeys, indx;
5805 indx = mc->mc_ki[mc->mc_top];
5806 mp = mc->mc_pg[mc->mc_top];
5807 node = NODEPTR(mp, indx);
5808 ptr = mp->mp_ptrs[indx];
5812 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5813 k2.mv_data = NODEKEY(node);
5814 k2.mv_size = node->mn_ksize;
5815 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5817 mdb_dkey(&k2, kbuf2),
5823 delta0 = delta = key->mv_size - node->mn_ksize;
5825 /* Must be 2-byte aligned. If new key is
5826 * shorter by 1, the shift will be skipped.
5828 delta += (delta & 1);
5830 if (delta > 0 && SIZELEFT(mp) < delta) {
5832 /* not enough space left, do a delete and split */
5833 DPRINTF("Not enough room, delta = %d, splitting...", delta);
5834 pgno = NODEPGNO(node);
5835 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5836 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
5839 numkeys = NUMKEYS(mp);
5840 for (i = 0; i < numkeys; i++) {
5841 if (mp->mp_ptrs[i] <= ptr)
5842 mp->mp_ptrs[i] -= delta;
5845 base = (char *)mp + mp->mp_upper;
5846 len = ptr - mp->mp_upper + NODESIZE;
5847 memmove(base - delta, base, len);
5848 mp->mp_upper -= delta;
5850 node = NODEPTR(mp, indx);
5853 /* But even if no shift was needed, update ksize */
5855 node->mn_ksize = key->mv_size;
5858 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5864 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
5866 /** Move a node from csrc to cdst.
5869 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5876 unsigned short flags;
5880 /* Mark src and dst as dirty. */
5881 if ((rc = mdb_page_touch(csrc)) ||
5882 (rc = mdb_page_touch(cdst)))
5885 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5886 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5887 key.mv_size = csrc->mc_db->md_pad;
5888 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5890 data.mv_data = NULL;
5894 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5895 assert(!((long)srcnode&1));
5896 srcpg = NODEPGNO(srcnode);
5897 flags = srcnode->mn_flags;
5898 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5899 unsigned int snum = csrc->mc_snum;
5901 /* must find the lowest key below src */
5902 mdb_page_search_root(csrc, NULL, 0);
5903 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5904 key.mv_size = csrc->mc_db->md_pad;
5905 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5907 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5908 key.mv_size = NODEKSZ(s2);
5909 key.mv_data = NODEKEY(s2);
5911 csrc->mc_snum = snum--;
5912 csrc->mc_top = snum;
5914 key.mv_size = NODEKSZ(srcnode);
5915 key.mv_data = NODEKEY(srcnode);
5917 data.mv_size = NODEDSZ(srcnode);
5918 data.mv_data = NODEDATA(srcnode);
5920 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5921 unsigned int snum = cdst->mc_snum;
5924 /* must find the lowest key below dst */
5925 mdb_page_search_root(cdst, NULL, 0);
5926 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5927 bkey.mv_size = cdst->mc_db->md_pad;
5928 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5930 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5931 bkey.mv_size = NODEKSZ(s2);
5932 bkey.mv_data = NODEKEY(s2);
5934 cdst->mc_snum = snum--;
5935 cdst->mc_top = snum;
5936 mdb_cursor_copy(cdst, &mn);
5938 rc = mdb_update_key(&mn, &bkey);
5943 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5944 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5945 csrc->mc_ki[csrc->mc_top],
5947 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5948 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5950 /* Add the node to the destination page.
5952 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5953 if (rc != MDB_SUCCESS)
5956 /* Delete the node from the source page.
5958 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5961 /* Adjust other cursors pointing to mp */
5962 MDB_cursor *m2, *m3;
5963 MDB_dbi dbi = csrc->mc_dbi;
5964 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5966 if (csrc->mc_flags & C_SUB)
5969 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5970 if (m2 == csrc) continue;
5971 if (csrc->mc_flags & C_SUB)
5972 m3 = &m2->mc_xcursor->mx_cursor;
5975 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5976 csrc->mc_ki[csrc->mc_top]) {
5977 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5978 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5983 /* Update the parent separators.
5985 if (csrc->mc_ki[csrc->mc_top] == 0) {
5986 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5987 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5988 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5990 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5991 key.mv_size = NODEKSZ(srcnode);
5992 key.mv_data = NODEKEY(srcnode);
5994 DPRINTF("update separator for source page %zu to [%s]",
5995 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5996 mdb_cursor_copy(csrc, &mn);
5999 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6002 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6004 indx_t ix = csrc->mc_ki[csrc->mc_top];
6005 nullkey.mv_size = 0;
6006 csrc->mc_ki[csrc->mc_top] = 0;
6007 rc = mdb_update_key(csrc, &nullkey);
6008 csrc->mc_ki[csrc->mc_top] = ix;
6009 assert(rc == MDB_SUCCESS);
6013 if (cdst->mc_ki[cdst->mc_top] == 0) {
6014 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6015 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6016 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6018 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6019 key.mv_size = NODEKSZ(srcnode);
6020 key.mv_data = NODEKEY(srcnode);
6022 DPRINTF("update separator for destination page %zu to [%s]",
6023 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6024 mdb_cursor_copy(cdst, &mn);
6027 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6030 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6032 indx_t ix = cdst->mc_ki[cdst->mc_top];
6033 nullkey.mv_size = 0;
6034 cdst->mc_ki[cdst->mc_top] = 0;
6035 rc = mdb_update_key(cdst, &nullkey);
6036 cdst->mc_ki[cdst->mc_top] = ix;
6037 assert(rc == MDB_SUCCESS);
6044 /** Merge one page into another.
6045 * The nodes from the page pointed to by \b csrc will
6046 * be copied to the page pointed to by \b cdst and then
6047 * the \b csrc page will be freed.
6048 * @param[in] csrc Cursor pointing to the source page.
6049 * @param[in] cdst Cursor pointing to the destination page.
6052 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6060 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6061 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6063 assert(csrc->mc_snum > 1); /* can't merge root page */
6064 assert(cdst->mc_snum > 1);
6066 /* Mark dst as dirty. */
6067 if ((rc = mdb_page_touch(cdst)))
6070 /* Move all nodes from src to dst.
6072 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6073 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6074 key.mv_size = csrc->mc_db->md_pad;
6075 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6076 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6077 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6078 if (rc != MDB_SUCCESS)
6080 key.mv_data = (char *)key.mv_data + key.mv_size;
6083 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6084 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6085 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6086 unsigned int snum = csrc->mc_snum;
6088 /* must find the lowest key below src */
6089 mdb_page_search_root(csrc, NULL, 0);
6090 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6091 key.mv_size = csrc->mc_db->md_pad;
6092 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6094 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6095 key.mv_size = NODEKSZ(s2);
6096 key.mv_data = NODEKEY(s2);
6098 csrc->mc_snum = snum--;
6099 csrc->mc_top = snum;
6101 key.mv_size = srcnode->mn_ksize;
6102 key.mv_data = NODEKEY(srcnode);
6105 data.mv_size = NODEDSZ(srcnode);
6106 data.mv_data = NODEDATA(srcnode);
6107 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6108 if (rc != MDB_SUCCESS)
6113 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6114 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);
6116 /* Unlink the src page from parent and add to free list.
6118 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6119 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6122 rc = mdb_update_key(csrc, &key);
6128 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6129 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6130 csrc->mc_db->md_leaf_pages--;
6132 csrc->mc_db->md_branch_pages--;
6134 /* Adjust other cursors pointing to mp */
6135 MDB_cursor *m2, *m3;
6136 MDB_dbi dbi = csrc->mc_dbi;
6137 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6139 if (csrc->mc_flags & C_SUB)
6142 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6143 if (csrc->mc_flags & C_SUB)
6144 m3 = &m2->mc_xcursor->mx_cursor;
6147 if (m3 == csrc) continue;
6148 if (m3->mc_snum < csrc->mc_snum) continue;
6149 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6150 m3->mc_pg[csrc->mc_top] = mp;
6151 m3->mc_ki[csrc->mc_top] += nkeys;
6155 mdb_cursor_pop(csrc);
6157 return mdb_rebalance(csrc);
6160 /** Copy the contents of a cursor.
6161 * @param[in] csrc The cursor to copy from.
6162 * @param[out] cdst The cursor to copy to.
6165 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6169 cdst->mc_txn = csrc->mc_txn;
6170 cdst->mc_dbi = csrc->mc_dbi;
6171 cdst->mc_db = csrc->mc_db;
6172 cdst->mc_dbx = csrc->mc_dbx;
6173 cdst->mc_snum = csrc->mc_snum;
6174 cdst->mc_top = csrc->mc_top;
6175 cdst->mc_flags = csrc->mc_flags;
6177 for (i=0; i<csrc->mc_snum; i++) {
6178 cdst->mc_pg[i] = csrc->mc_pg[i];
6179 cdst->mc_ki[i] = csrc->mc_ki[i];
6183 /** Rebalance the tree after a delete operation.
6184 * @param[in] mc Cursor pointing to the page where rebalancing
6186 * @return 0 on success, non-zero on failure.
6189 mdb_rebalance(MDB_cursor *mc)
6199 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6200 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6201 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6202 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6206 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
6209 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6210 DPRINTF("no need to rebalance page %zu, above fill threshold",
6216 if (mc->mc_snum < 2) {
6217 MDB_page *mp = mc->mc_pg[0];
6218 if (NUMKEYS(mp) == 0) {
6219 DPUTS("tree is completely empty");
6220 mc->mc_db->md_root = P_INVALID;
6221 mc->mc_db->md_depth = 0;
6222 mc->mc_db->md_leaf_pages = 0;
6223 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6227 /* Adjust other cursors pointing to mp */
6228 MDB_cursor *m2, *m3;
6229 MDB_dbi dbi = mc->mc_dbi;
6231 if (mc->mc_flags & C_SUB)
6234 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6235 if (m2 == mc) continue;
6236 if (mc->mc_flags & C_SUB)
6237 m3 = &m2->mc_xcursor->mx_cursor;
6240 if (m3->mc_snum < mc->mc_snum) continue;
6241 if (m3->mc_pg[0] == mp) {
6247 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6248 DPUTS("collapsing root page!");
6249 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6250 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6251 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6254 mc->mc_db->md_depth--;
6255 mc->mc_db->md_branch_pages--;
6257 /* Adjust other cursors pointing to mp */
6258 MDB_cursor *m2, *m3;
6259 MDB_dbi dbi = mc->mc_dbi;
6261 if (mc->mc_flags & C_SUB)
6264 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6265 if (m2 == mc) continue;
6266 if (mc->mc_flags & C_SUB)
6267 m3 = &m2->mc_xcursor->mx_cursor;
6270 if (m3->mc_snum < mc->mc_snum) continue;
6271 if (m3->mc_pg[0] == mp) {
6272 m3->mc_pg[0] = mc->mc_pg[0];
6277 DPUTS("root page doesn't need rebalancing");
6281 /* The parent (branch page) must have at least 2 pointers,
6282 * otherwise the tree is invalid.
6284 ptop = mc->mc_top-1;
6285 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6287 /* Leaf page fill factor is below the threshold.
6288 * Try to move keys from left or right neighbor, or
6289 * merge with a neighbor page.
6294 mdb_cursor_copy(mc, &mn);
6295 mn.mc_xcursor = NULL;
6297 if (mc->mc_ki[ptop] == 0) {
6298 /* We're the leftmost leaf in our parent.
6300 DPUTS("reading right 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] = 0;
6306 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6308 /* There is at least one neighbor to the left.
6310 DPUTS("reading left neighbor");
6312 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6313 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6315 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6316 mc->mc_ki[mc->mc_top] = 0;
6319 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6320 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);
6322 /* If the neighbor page is above threshold and has at least two
6323 * keys, move one key from it.
6325 * Otherwise we should try to merge them.
6327 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6328 return mdb_node_move(&mn, mc);
6330 if (mc->mc_ki[ptop] == 0)
6331 rc = mdb_page_merge(&mn, mc);
6333 rc = mdb_page_merge(mc, &mn);
6334 mc->mc_flags &= ~C_INITIALIZED;
6339 /** Complete a delete operation started by #mdb_cursor_del(). */
6341 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6345 /* add overflow pages to free list */
6346 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6350 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6351 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6352 mc->mc_db->md_overflow_pages -= ovpages;
6353 for (i=0; i<ovpages; i++) {
6354 DPRINTF("freed ov page %zu", pg);
6355 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6359 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6360 mc->mc_db->md_entries--;
6361 rc = mdb_rebalance(mc);
6362 if (rc != MDB_SUCCESS)
6363 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6369 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6370 MDB_val *key, MDB_val *data)
6375 MDB_val rdata, *xdata;
6379 assert(key != NULL);
6381 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6383 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6386 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6390 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6394 mdb_cursor_init(&mc, txn, dbi, &mx);
6405 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6407 /* let mdb_page_split know about this cursor if needed:
6408 * delete will trigger a rebalance; if it needs to move
6409 * a node from one page to another, it will have to
6410 * update the parent's separator key(s). If the new sepkey
6411 * is larger than the current one, the parent page may
6412 * run out of space, triggering a split. We need this
6413 * cursor to be consistent until the end of the rebalance.
6415 mc.mc_next = txn->mt_cursors[dbi];
6416 txn->mt_cursors[dbi] = &mc;
6417 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6418 txn->mt_cursors[dbi] = mc.mc_next;
6423 /** Split a page and insert a new node.
6424 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6425 * The cursor will be updated to point to the actual page and index where
6426 * the node got inserted after the split.
6427 * @param[in] newkey The key for the newly inserted node.
6428 * @param[in] newdata The data for the newly inserted node.
6429 * @param[in] newpgno The page number, if the new node is a branch node.
6430 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6431 * @return 0 on success, non-zero on failure.
6434 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6435 unsigned int nflags)
6438 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6441 unsigned int i, j, split_indx, nkeys, pmax;
6443 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6445 MDB_page *mp, *rp, *pp;
6450 mp = mc->mc_pg[mc->mc_top];
6451 newindx = mc->mc_ki[mc->mc_top];
6453 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6454 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6455 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6457 /* Create a right sibling. */
6458 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6460 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6462 if (mc->mc_snum < 2) {
6463 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6465 /* shift current top to make room for new parent */
6466 mc->mc_pg[1] = mc->mc_pg[0];
6467 mc->mc_ki[1] = mc->mc_ki[0];
6470 mc->mc_db->md_root = pp->mp_pgno;
6471 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6472 mc->mc_db->md_depth++;
6475 /* Add left (implicit) pointer. */
6476 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6477 /* undo the pre-push */
6478 mc->mc_pg[0] = mc->mc_pg[1];
6479 mc->mc_ki[0] = mc->mc_ki[1];
6480 mc->mc_db->md_root = mp->mp_pgno;
6481 mc->mc_db->md_depth--;
6488 ptop = mc->mc_top-1;
6489 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6492 mc->mc_flags |= C_SPLITTING;
6493 mdb_cursor_copy(mc, &mn);
6494 mn.mc_pg[mn.mc_top] = rp;
6495 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6497 if (nflags & MDB_APPEND) {
6498 mn.mc_ki[mn.mc_top] = 0;
6500 split_indx = newindx;
6505 nkeys = NUMKEYS(mp);
6506 split_indx = nkeys / 2;
6507 if (newindx < split_indx)
6513 unsigned int lsize, rsize, ksize;
6514 /* Move half of the keys to the right sibling */
6516 x = mc->mc_ki[mc->mc_top] - split_indx;
6517 ksize = mc->mc_db->md_pad;
6518 split = LEAF2KEY(mp, split_indx, ksize);
6519 rsize = (nkeys - split_indx) * ksize;
6520 lsize = (nkeys - split_indx) * sizeof(indx_t);
6521 mp->mp_lower -= lsize;
6522 rp->mp_lower += lsize;
6523 mp->mp_upper += rsize - lsize;
6524 rp->mp_upper -= rsize - lsize;
6525 sepkey.mv_size = ksize;
6526 if (newindx == split_indx) {
6527 sepkey.mv_data = newkey->mv_data;
6529 sepkey.mv_data = split;
6532 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6533 memcpy(rp->mp_ptrs, split, rsize);
6534 sepkey.mv_data = rp->mp_ptrs;
6535 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6536 memcpy(ins, newkey->mv_data, ksize);
6537 mp->mp_lower += sizeof(indx_t);
6538 mp->mp_upper -= ksize - sizeof(indx_t);
6541 memcpy(rp->mp_ptrs, split, x * ksize);
6542 ins = LEAF2KEY(rp, x, ksize);
6543 memcpy(ins, newkey->mv_data, ksize);
6544 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6545 rp->mp_lower += sizeof(indx_t);
6546 rp->mp_upper -= ksize - sizeof(indx_t);
6547 mc->mc_ki[mc->mc_top] = x;
6548 mc->mc_pg[mc->mc_top] = rp;
6553 /* For leaf pages, check the split point based on what
6554 * fits where, since otherwise mdb_node_add can fail.
6556 * This check is only needed when the data items are
6557 * relatively large, such that being off by one will
6558 * make the difference between success or failure.
6560 * It's also relevant if a page happens to be laid out
6561 * such that one half of its nodes are all "small" and
6562 * the other half of its nodes are "large." If the new
6563 * item is also "large" and falls on the half with
6564 * "large" nodes, it also may not fit.
6567 unsigned int psize, nsize;
6568 /* Maximum free space in an empty page */
6569 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6570 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6571 if ((nkeys < 20) || (nsize > pmax/16)) {
6572 if (newindx <= split_indx) {
6575 for (i=0; i<split_indx; i++) {
6576 node = NODEPTR(mp, i);
6577 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6578 if (F_ISSET(node->mn_flags, F_BIGDATA))
6579 psize += sizeof(pgno_t);
6581 psize += NODEDSZ(node);
6585 split_indx = newindx;
6596 for (i=nkeys-1; i>=split_indx; i--) {
6597 node = NODEPTR(mp, i);
6598 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6599 if (F_ISSET(node->mn_flags, F_BIGDATA))
6600 psize += sizeof(pgno_t);
6602 psize += NODEDSZ(node);
6606 split_indx = newindx;
6617 /* First find the separating key between the split pages.
6618 * The case where newindx == split_indx is ambiguous; the
6619 * new item could go to the new page or stay on the original
6620 * page. If newpos == 1 it goes to the new page.
6622 if (newindx == split_indx && newpos) {
6623 sepkey.mv_size = newkey->mv_size;
6624 sepkey.mv_data = newkey->mv_data;
6626 node = NODEPTR(mp, split_indx);
6627 sepkey.mv_size = node->mn_ksize;
6628 sepkey.mv_data = NODEKEY(node);
6632 DPRINTF("separator is [%s]", DKEY(&sepkey));
6634 /* Copy separator key to the parent.
6636 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6640 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6643 if (mn.mc_snum == mc->mc_snum) {
6644 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6645 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6646 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6647 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6652 /* Right page might now have changed parent.
6653 * Check if left page also changed parent.
6655 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6656 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6657 for (i=0; i<ptop; i++) {
6658 mc->mc_pg[i] = mn.mc_pg[i];
6659 mc->mc_ki[i] = mn.mc_ki[i];
6661 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6662 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6666 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6669 mc->mc_flags ^= C_SPLITTING;
6670 if (rc != MDB_SUCCESS) {
6673 if (nflags & MDB_APPEND) {
6674 mc->mc_pg[mc->mc_top] = rp;
6675 mc->mc_ki[mc->mc_top] = 0;
6676 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6679 for (i=0; i<mc->mc_top; i++)
6680 mc->mc_ki[i] = mn.mc_ki[i];
6687 /* Move half of the keys to the right sibling. */
6689 /* grab a page to hold a temporary copy */
6690 copy = mdb_page_malloc(mc);
6694 copy->mp_pgno = mp->mp_pgno;
6695 copy->mp_flags = mp->mp_flags;
6696 copy->mp_lower = PAGEHDRSZ;
6697 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6698 mc->mc_pg[mc->mc_top] = copy;
6699 for (i = j = 0; i <= nkeys; j++) {
6700 if (i == split_indx) {
6701 /* Insert in right sibling. */
6702 /* Reset insert index for right sibling. */
6703 if (i != newindx || (newpos ^ ins_new)) {
6705 mc->mc_pg[mc->mc_top] = rp;
6709 if (i == newindx && !ins_new) {
6710 /* Insert the original entry that caused the split. */
6711 rkey.mv_data = newkey->mv_data;
6712 rkey.mv_size = newkey->mv_size;
6721 /* Update index for the new key. */
6722 mc->mc_ki[mc->mc_top] = j;
6723 } else if (i == nkeys) {
6726 node = NODEPTR(mp, i);
6727 rkey.mv_data = NODEKEY(node);
6728 rkey.mv_size = node->mn_ksize;
6730 xdata.mv_data = NODEDATA(node);
6731 xdata.mv_size = NODEDSZ(node);
6734 pgno = NODEPGNO(node);
6735 flags = node->mn_flags;
6740 if (!IS_LEAF(mp) && j == 0) {
6741 /* First branch index doesn't need key data. */
6745 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6749 nkeys = NUMKEYS(copy);
6750 for (i=0; i<nkeys; i++)
6751 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6752 mp->mp_lower = copy->mp_lower;
6753 mp->mp_upper = copy->mp_upper;
6754 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6755 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6757 /* reset back to original page */
6758 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6759 mc->mc_pg[mc->mc_top] = mp;
6760 if (nflags & MDB_RESERVE) {
6761 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6762 if (!(node->mn_flags & F_BIGDATA))
6763 newdata->mv_data = NODEDATA(node);
6769 /* return tmp page to freelist */
6770 mdb_page_free(mc->mc_txn->mt_env, copy);
6773 /* Adjust other cursors pointing to mp */
6774 MDB_cursor *m2, *m3;
6775 MDB_dbi dbi = mc->mc_dbi;
6776 int fixup = NUMKEYS(mp);
6778 if (mc->mc_flags & C_SUB)
6781 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6782 if (m2 == mc) continue;
6783 if (mc->mc_flags & C_SUB)
6784 m3 = &m2->mc_xcursor->mx_cursor;
6787 if (!(m3->mc_flags & C_INITIALIZED))
6789 if (m3->mc_flags & C_SPLITTING)
6794 for (k=m3->mc_top; k>=0; k--) {
6795 m3->mc_ki[k+1] = m3->mc_ki[k];
6796 m3->mc_pg[k+1] = m3->mc_pg[k];
6798 if (m3->mc_ki[0] >= split_indx) {
6803 m3->mc_pg[0] = mc->mc_pg[0];
6807 if (m3->mc_pg[mc->mc_top] == mp) {
6808 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6809 m3->mc_ki[mc->mc_top]++;
6810 if (m3->mc_ki[mc->mc_top] >= fixup) {
6811 m3->mc_pg[mc->mc_top] = rp;
6812 m3->mc_ki[mc->mc_top] -= fixup;
6813 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6815 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6816 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6825 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6826 MDB_val *key, MDB_val *data, unsigned int flags)
6831 assert(key != NULL);
6832 assert(data != NULL);
6834 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6837 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6841 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6845 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6848 mdb_cursor_init(&mc, txn, dbi, &mx);
6849 return mdb_cursor_put(&mc, key, data, flags);
6853 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6855 if ((flag & CHANGEABLE) != flag)
6858 env->me_flags |= flag;
6860 env->me_flags &= ~flag;
6865 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6870 *arg = env->me_flags;
6875 mdb_env_get_path(MDB_env *env, const char **arg)
6880 *arg = env->me_path;
6884 /** Common code for #mdb_stat() and #mdb_env_stat().
6885 * @param[in] env the environment to operate in.
6886 * @param[in] db the #MDB_db record containing the stats to return.
6887 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6888 * @return 0, this function always succeeds.
6891 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6893 arg->ms_psize = env->me_psize;
6894 arg->ms_depth = db->md_depth;
6895 arg->ms_branch_pages = db->md_branch_pages;
6896 arg->ms_leaf_pages = db->md_leaf_pages;
6897 arg->ms_overflow_pages = db->md_overflow_pages;
6898 arg->ms_entries = db->md_entries;
6903 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6907 if (env == NULL || arg == NULL)
6910 toggle = mdb_env_pick_meta(env);
6912 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6916 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
6920 if (env == NULL || arg == NULL)
6923 toggle = mdb_env_pick_meta(env);
6924 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
6925 arg->me_mapsize = env->me_mapsize;
6926 arg->me_maxreaders = env->me_maxreaders;
6927 arg->me_numreaders = env->me_numreaders;
6928 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
6929 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
6933 /** Set the default comparison functions for a database.
6934 * Called immediately after a database is opened to set the defaults.
6935 * The user can then override them with #mdb_set_compare() or
6936 * #mdb_set_dupsort().
6937 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6938 * @param[in] dbi A database handle returned by #mdb_dbi_open()
6941 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6943 uint16_t f = txn->mt_dbs[dbi].md_flags;
6945 txn->mt_dbxs[dbi].md_cmp =
6946 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6947 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6949 txn->mt_dbxs[dbi].md_dcmp =
6950 !(f & MDB_DUPSORT) ? 0 :
6951 ((f & MDB_INTEGERDUP)
6952 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6953 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6956 #define PERSISTENT_FLAGS 0xffff
6957 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
6958 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
6959 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6964 int rc, dbflag, exact;
6965 unsigned int unused = 0;
6968 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6969 mdb_default_cmp(txn, FREE_DBI);
6972 if ((flags & VALID_FLAGS) != flags)
6978 if (flags & PERSISTENT_FLAGS) {
6979 uint16_t f2 = flags & PERSISTENT_FLAGS;
6980 /* make sure flag changes get committed */
6981 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
6982 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
6983 txn->mt_flags |= MDB_TXN_DIRTY;
6986 mdb_default_cmp(txn, MAIN_DBI);
6990 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6991 mdb_default_cmp(txn, MAIN_DBI);
6994 /* Is the DB already open? */
6996 for (i=2; i<txn->mt_numdbs; i++) {
6997 if (!txn->mt_dbxs[i].md_name.mv_size) {
6998 /* Remember this free slot */
6999 if (!unused) unused = i;
7002 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7003 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7009 /* If no free slot and max hit, fail */
7010 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7011 return MDB_DBS_FULL;
7013 /* Find the DB info */
7017 key.mv_data = (void *)name;
7018 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7019 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7020 if (rc == MDB_SUCCESS) {
7021 /* make sure this is actually a DB */
7022 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7023 if (!(node->mn_flags & F_SUBDATA))
7025 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7026 /* Create if requested */
7028 data.mv_size = sizeof(MDB_db);
7029 data.mv_data = &dummy;
7030 memset(&dummy, 0, sizeof(dummy));
7031 dummy.md_root = P_INVALID;
7032 dummy.md_flags = flags & PERSISTENT_FLAGS;
7033 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7037 /* OK, got info, add to table */
7038 if (rc == MDB_SUCCESS) {
7039 unsigned int slot = unused ? unused : txn->mt_numdbs;
7040 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7041 txn->mt_dbxs[slot].md_name.mv_size = len;
7042 txn->mt_dbxs[slot].md_rel = NULL;
7043 txn->mt_dbflags[slot] = dbflag;
7044 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7046 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7047 mdb_default_cmp(txn, slot);
7050 txn->mt_env->me_numdbs++;
7057 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7059 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7062 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7065 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7068 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
7070 ptr = env->me_dbxs[dbi].md_name.mv_data;
7071 env->me_dbxs[dbi].md_name.mv_data = NULL;
7072 env->me_dbxs[dbi].md_name.mv_size = 0;
7076 /** Add all the DB's pages to the free list.
7077 * @param[in] mc Cursor on the DB to free.
7078 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7079 * @return 0 on success, non-zero on failure.
7082 mdb_drop0(MDB_cursor *mc, int subs)
7086 rc = mdb_page_search(mc, NULL, 0);
7087 if (rc == MDB_SUCCESS) {
7092 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7093 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
7096 mdb_cursor_copy(mc, &mx);
7097 while (mc->mc_snum > 0) {
7098 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7099 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7100 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7101 if (ni->mn_flags & F_SUBDATA) {
7102 mdb_xcursor_init1(mc, ni);
7103 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7109 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7111 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7114 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7119 rc = mdb_cursor_sibling(mc, 1);
7121 /* no more siblings, go back to beginning
7122 * of previous level.
7125 for (i=1; i<mc->mc_top; i++)
7126 mc->mc_pg[i] = mx.mc_pg[i];
7130 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7131 mc->mc_db->md_root);
7136 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7141 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1)
7144 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7147 rc = mdb_cursor_open(txn, dbi, &mc);
7151 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7155 /* Can't delete the main DB */
7156 if (del && dbi > MAIN_DBI) {
7157 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7159 mdb_dbi_close(txn->mt_env, dbi);
7161 /* reset the DB record, mark it dirty */
7162 txn->mt_dbflags[dbi] |= DB_DIRTY;
7163 txn->mt_dbs[dbi].md_depth = 0;
7164 txn->mt_dbs[dbi].md_branch_pages = 0;
7165 txn->mt_dbs[dbi].md_leaf_pages = 0;
7166 txn->mt_dbs[dbi].md_overflow_pages = 0;
7167 txn->mt_dbs[dbi].md_entries = 0;
7168 txn->mt_dbs[dbi].md_root = P_INVALID;
7170 if (!txn->mt_u.dirty_list[0].mid) {
7173 /* make sure we have at least one dirty page in this txn
7174 * otherwise these changes will be ignored.
7176 key.mv_size = sizeof(txnid_t);
7177 key.mv_data = &txn->mt_txnid;
7178 data.mv_size = sizeof(MDB_ID);
7179 data.mv_data = txn->mt_free_pgs;
7180 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
7181 rc = mdb_cursor_put(&m2, &key, &data, 0);
7185 mdb_cursor_close(mc);
7189 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7191 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7194 txn->mt_dbxs[dbi].md_cmp = cmp;
7198 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7200 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7203 txn->mt_dbxs[dbi].md_dcmp = cmp;
7207 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7209 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7212 txn->mt_dbxs[dbi].md_rel = rel;
7216 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7218 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7221 txn->mt_dbxs[dbi].md_relctx = ctx;