2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
36 #include <sys/types.h>
38 #include <sys/param.h>
44 #ifdef HAVE_SYS_FILE_H
61 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
62 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
65 #if defined(__APPLE__) || defined (BSD)
66 # define MDB_USE_POSIX_SEM 1
67 # define MDB_FDATASYNC fsync
68 #elif defined(ANDROID)
69 # define MDB_FDATASYNC fsync
74 #ifdef MDB_USE_POSIX_SEM
75 #include <semaphore.h>
80 #include <valgrind/memcheck.h>
81 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
82 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
83 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
84 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
85 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
87 #define VGMEMP_CREATE(h,r,z)
88 #define VGMEMP_ALLOC(h,a,s)
89 #define VGMEMP_FREE(h,a)
90 #define VGMEMP_DESTROY(h)
91 #define VGMEMP_DEFINED(a,s)
95 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
96 /* Solaris just defines one or the other */
97 # define LITTLE_ENDIAN 1234
98 # define BIG_ENDIAN 4321
99 # ifdef _LITTLE_ENDIAN
100 # define BYTE_ORDER LITTLE_ENDIAN
102 # define BYTE_ORDER BIG_ENDIAN
105 # define BYTE_ORDER __BYTE_ORDER
109 #ifndef LITTLE_ENDIAN
110 #define LITTLE_ENDIAN __LITTLE_ENDIAN
113 #define BIG_ENDIAN __BIG_ENDIAN
116 #if defined(__i386) || defined(__x86_64)
117 #define MISALIGNED_OK 1
123 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
124 # error "Unknown or unsupported endianness (BYTE_ORDER)"
125 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
126 # error "Two's complement, reasonably sized integer types, please"
129 /** @defgroup internal MDB Internals
132 /** @defgroup compat Windows Compatibility Macros
133 * A bunch of macros to minimize the amount of platform-specific ifdefs
134 * needed throughout the rest of the code. When the features this library
135 * needs are similar enough to POSIX to be hidden in a one-or-two line
136 * replacement, this macro approach is used.
140 #define pthread_t DWORD
141 #define pthread_mutex_t HANDLE
142 #define pthread_key_t DWORD
143 #define pthread_self() GetCurrentThreadId()
144 #define pthread_key_create(x,y) \
145 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
146 #define pthread_key_delete(x) TlsFree(x)
147 #define pthread_getspecific(x) TlsGetValue(x)
148 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
149 #define pthread_mutex_unlock(x) ReleaseMutex(x)
150 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
151 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
152 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
153 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
154 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
155 #define getpid() GetCurrentProcessId()
156 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
157 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
158 #define ErrCode() GetLastError()
159 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
160 #define close(fd) CloseHandle(fd)
161 #define munmap(ptr,len) UnmapViewOfFile(ptr)
164 #ifdef MDB_USE_POSIX_SEM
166 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
167 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
168 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
169 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
172 mdb_sem_wait(sem_t *sem)
175 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
180 /** Lock the reader mutex.
182 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
183 /** Unlock the reader mutex.
185 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
187 /** Lock the writer mutex.
188 * Only a single write transaction is allowed at a time. Other writers
189 * will block waiting for this mutex.
191 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
192 /** Unlock the writer mutex.
194 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
195 #endif /* MDB_USE_POSIX_SEM */
197 /** Get the error code for the last failed system function.
199 #define ErrCode() errno
201 /** An abstraction for a file handle.
202 * On POSIX systems file handles are small integers. On Windows
203 * they're opaque pointers.
207 /** A value for an invalid file handle.
208 * Mainly used to initialize file variables and signify that they are
211 #define INVALID_HANDLE_VALUE (-1)
213 /** Get the size of a memory page for the system.
214 * This is the basic size that the platform's memory manager uses, and is
215 * fundamental to the use of memory-mapped files.
217 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
220 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
223 #define MNAME_LEN (sizeof(pthread_mutex_t))
229 /** A flag for opening a file and requesting synchronous data writes.
230 * This is only used when writing a meta page. It's not strictly needed;
231 * we could just do a normal write and then immediately perform a flush.
232 * But if this flag is available it saves us an extra system call.
234 * @note If O_DSYNC is undefined but exists in /usr/include,
235 * preferably set some compiler flag to get the definition.
236 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
239 # define MDB_DSYNC O_DSYNC
243 /** Function for flushing the data of a file. Define this to fsync
244 * if fdatasync() is not supported.
246 #ifndef MDB_FDATASYNC
247 # define MDB_FDATASYNC fdatasync
251 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
262 /** A page number in the database.
263 * Note that 64 bit page numbers are overkill, since pages themselves
264 * already represent 12-13 bits of addressable memory, and the OS will
265 * always limit applications to a maximum of 63 bits of address space.
267 * @note In the #MDB_node structure, we only store 48 bits of this value,
268 * which thus limits us to only 60 bits of addressable data.
270 typedef MDB_ID pgno_t;
272 /** A transaction ID.
273 * See struct MDB_txn.mt_txnid for details.
275 typedef MDB_ID txnid_t;
277 /** @defgroup debug Debug Macros
281 /** Enable debug output.
282 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
283 * read from and written to the database (used for free space management).
288 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
289 # define DPRINTF (void) /* Vararg macros may be unsupported */
291 static int mdb_debug;
292 static txnid_t mdb_debug_start;
294 /** Print a debug message with printf formatting. */
295 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
296 ((void) ((mdb_debug) && \
297 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
299 # define DPRINTF(fmt, ...) ((void) 0)
300 # define MDB_DEBUG_SKIP
302 /** Print a debug string.
303 * The string is printed literally, with no format processing.
305 #define DPUTS(arg) DPRINTF("%s", arg)
308 /** A default memory page size.
309 * The actual size is platform-dependent, but we use this for
310 * boot-strapping. We probably should not be using this any more.
311 * The #GET_PAGESIZE() macro is used to get the actual size.
313 * Note that we don't currently support Huge pages. On Linux,
314 * regular data files cannot use Huge pages, and in general
315 * Huge pages aren't actually pageable. We rely on the OS
316 * demand-pager to read our data and page it out when memory
317 * pressure from other processes is high. So until OSs have
318 * actual paging support for Huge pages, they're not viable.
320 #define MDB_PAGESIZE 4096
322 /** The minimum number of keys required in a database page.
323 * Setting this to a larger value will place a smaller bound on the
324 * maximum size of a data item. Data items larger than this size will
325 * be pushed into overflow pages instead of being stored directly in
326 * the B-tree node. This value used to default to 4. With a page size
327 * of 4096 bytes that meant that any item larger than 1024 bytes would
328 * go into an overflow page. That also meant that on average 2-3KB of
329 * each overflow page was wasted space. The value cannot be lower than
330 * 2 because then there would no longer be a tree structure. With this
331 * value, items larger than 2KB will go into overflow pages, and on
332 * average only 1KB will be wasted.
334 #define MDB_MINKEYS 2
336 /** A stamp that identifies a file as an MDB file.
337 * There's nothing special about this value other than that it is easily
338 * recognizable, and it will reflect any byte order mismatches.
340 #define MDB_MAGIC 0xBEEFC0DE
342 /** The version number for a database's file format. */
343 #define MDB_VERSION 1
345 /** @brief The maximum size of a key in the database.
347 * We require that keys all fit onto a regular page. This limit
348 * could be raised a bit further if needed; to something just
349 * under #MDB_PAGESIZE / #MDB_MINKEYS.
351 * Note that data items in an #MDB_DUPSORT database are actually keys
352 * of a subDB, so they're also limited to this size.
354 #ifndef MDB_MAXKEYSIZE
355 #define MDB_MAXKEYSIZE 511
358 /** @brief The maximum size of a data item.
360 * We only store a 32 bit value for node sizes.
362 #define MAXDATASIZE 0xffffffffUL
367 * This is used for printing a hex dump of a key's contents.
369 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
370 /** Display a key in hex.
372 * Invoke a function to display a key in hex.
374 #define DKEY(x) mdb_dkey(x, kbuf)
376 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
380 /** An invalid page number.
381 * Mainly used to denote an empty tree.
383 #define P_INVALID (~(pgno_t)0)
385 /** Test if the flags \b f are set in a flag word \b w. */
386 #define F_ISSET(w, f) (((w) & (f)) == (f))
388 /** Used for offsets within a single page.
389 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
392 typedef uint16_t indx_t;
394 /** Default size of memory map.
395 * This is certainly too small for any actual applications. Apps should always set
396 * the size explicitly using #mdb_env_set_mapsize().
398 #define DEFAULT_MAPSIZE 1048576
400 /** @defgroup readers Reader Lock Table
401 * Readers don't acquire any locks for their data access. Instead, they
402 * simply record their transaction ID in the reader table. The reader
403 * mutex is needed just to find an empty slot in the reader table. The
404 * slot's address is saved in thread-specific data so that subsequent read
405 * transactions started by the same thread need no further locking to proceed.
407 * No reader table is used if the database is on a read-only filesystem.
409 * Since the database uses multi-version concurrency control, readers don't
410 * actually need any locking. This table is used to keep track of which
411 * readers are using data from which old transactions, so that we'll know
412 * when a particular old transaction is no longer in use. Old transactions
413 * that have discarded any data pages can then have those pages reclaimed
414 * for use by a later write transaction.
416 * The lock table is constructed such that reader slots are aligned with the
417 * processor's cache line size. Any slot is only ever used by one thread.
418 * This alignment guarantees that there will be no contention or cache
419 * thrashing as threads update their own slot info, and also eliminates
420 * any need for locking when accessing a slot.
422 * A writer thread will scan every slot in the table to determine the oldest
423 * outstanding reader transaction. Any freed pages older than this will be
424 * reclaimed by the writer. The writer doesn't use any locks when scanning
425 * this table. This means that there's no guarantee that the writer will
426 * see the most up-to-date reader info, but that's not required for correct
427 * operation - all we need is to know the upper bound on the oldest reader,
428 * we don't care at all about the newest reader. So the only consequence of
429 * reading stale information here is that old pages might hang around a
430 * while longer before being reclaimed. That's actually good anyway, because
431 * the longer we delay reclaiming old pages, the more likely it is that a
432 * string of contiguous pages can be found after coalescing old pages from
433 * many old transactions together.
435 * @todo We don't actually do such coalescing yet, we grab pages from one
436 * old transaction at a time.
439 /** Number of slots in the reader table.
440 * This value was chosen somewhat arbitrarily. 126 readers plus a
441 * couple mutexes fit exactly into 8KB on my development machine.
442 * Applications should set the table size using #mdb_env_set_maxreaders().
444 #define DEFAULT_READERS 126
446 /** The size of a CPU cache line in bytes. We want our lock structures
447 * aligned to this size to avoid false cache line sharing in the
449 * This value works for most CPUs. For Itanium this should be 128.
455 /** The information we store in a single slot of the reader table.
456 * In addition to a transaction ID, we also record the process and
457 * thread ID that owns a slot, so that we can detect stale information,
458 * e.g. threads or processes that went away without cleaning up.
459 * @note We currently don't check for stale records. We simply re-init
460 * the table when we know that we're the only process opening the
463 typedef struct MDB_rxbody {
464 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
465 * Multiple readers that start at the same time will probably have the
466 * same ID here. Again, it's not important to exclude them from
467 * anything; all we need to know is which version of the DB they
468 * started from so we can avoid overwriting any data used in that
469 * particular version.
472 /** The process ID of the process owning this reader txn. */
474 /** The thread ID of the thread owning this txn. */
478 /** The actual reader record, with cacheline padding. */
479 typedef struct MDB_reader {
482 /** shorthand for mrb_txnid */
483 #define mr_txnid mru.mrx.mrb_txnid
484 #define mr_pid mru.mrx.mrb_pid
485 #define mr_tid mru.mrx.mrb_tid
486 /** cache line alignment */
487 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
491 /** The header for the reader table.
492 * The table resides in a memory-mapped file. (This is a different file
493 * than is used for the main database.)
495 * For POSIX the actual mutexes reside in the shared memory of this
496 * mapped file. On Windows, mutexes are named objects allocated by the
497 * kernel; we store the mutex names in this mapped file so that other
498 * processes can grab them. This same approach is also used on
499 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
500 * process-shared POSIX mutexes. For these cases where a named object
501 * is used, the object name is derived from a 64 bit FNV hash of the
502 * environment pathname. As such, naming collisions are extremely
503 * unlikely. If a collision occurs, the results are unpredictable.
505 typedef struct MDB_txbody {
506 /** Stamp identifying this as an MDB file. It must be set
509 /** Version number of this lock file. Must be set to #MDB_VERSION. */
510 uint32_t mtb_version;
511 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
512 char mtb_rmname[MNAME_LEN];
514 /** Mutex protecting access to this table.
515 * This is the reader lock that #LOCK_MUTEX_R acquires.
517 pthread_mutex_t mtb_mutex;
519 /** The ID of the last transaction committed to the database.
520 * This is recorded here only for convenience; the value can always
521 * be determined by reading the main database meta pages.
524 /** The number of slots that have been used in the reader table.
525 * This always records the maximum count, it is not decremented
526 * when readers release their slots.
528 unsigned mtb_numreaders;
531 /** The actual reader table definition. */
532 typedef struct MDB_txninfo {
535 #define mti_magic mt1.mtb.mtb_magic
536 #define mti_version mt1.mtb.mtb_version
537 #define mti_mutex mt1.mtb.mtb_mutex
538 #define mti_rmname mt1.mtb.mtb_rmname
539 #define mti_txnid mt1.mtb.mtb_txnid
540 #define mti_numreaders mt1.mtb.mtb_numreaders
541 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
544 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
545 char mt2_wmname[MNAME_LEN];
546 #define mti_wmname mt2.mt2_wmname
548 pthread_mutex_t mt2_wmutex;
549 #define mti_wmutex mt2.mt2_wmutex
551 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
553 MDB_reader mti_readers[1];
557 /** Common header for all page types.
558 * Overflow records occupy a number of contiguous pages with no
559 * headers on any page after the first.
561 typedef struct MDB_page {
562 #define mp_pgno mp_p.p_pgno
563 #define mp_next mp_p.p_next
565 pgno_t p_pgno; /**< page number */
566 void * p_next; /**< for in-memory list of freed structs */
569 /** @defgroup mdb_page Page Flags
571 * Flags for the page headers.
574 #define P_BRANCH 0x01 /**< branch page */
575 #define P_LEAF 0x02 /**< leaf page */
576 #define P_OVERFLOW 0x04 /**< overflow page */
577 #define P_META 0x08 /**< meta page */
578 #define P_DIRTY 0x10 /**< dirty page */
579 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
580 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
582 uint16_t mp_flags; /**< @ref mdb_page */
583 #define mp_lower mp_pb.pb.pb_lower
584 #define mp_upper mp_pb.pb.pb_upper
585 #define mp_pages mp_pb.pb_pages
588 indx_t pb_lower; /**< lower bound of free space */
589 indx_t pb_upper; /**< upper bound of free space */
591 uint32_t pb_pages; /**< number of overflow pages */
593 indx_t mp_ptrs[1]; /**< dynamic size */
596 /** Size of the page header, excluding dynamic data at the end */
597 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
599 /** Address of first usable data byte in a page, after the header */
600 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
602 /** Number of nodes on a page */
603 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
605 /** The amount of space remaining in the page */
606 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
608 /** The percentage of space used in the page, in tenths of a percent. */
609 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
610 ((env)->me_psize - PAGEHDRSZ))
611 /** The minimum page fill factor, in tenths of a percent.
612 * Pages emptier than this are candidates for merging.
614 #define FILL_THRESHOLD 250
616 /** Test if a page is a leaf page */
617 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
618 /** Test if a page is a LEAF2 page */
619 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
620 /** Test if a page is a branch page */
621 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
622 /** Test if a page is an overflow page */
623 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
624 /** Test if a page is a sub page */
625 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
627 /** The number of overflow pages needed to store the given size. */
628 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
630 /** Header for a single key/data pair within a page.
631 * We guarantee 2-byte alignment for nodes.
633 typedef struct MDB_node {
634 /** lo and hi are used for data size on leaf nodes and for
635 * child pgno on branch nodes. On 64 bit platforms, flags
636 * is also used for pgno. (Branch nodes have no flags).
637 * They are in host byte order in case that lets some
638 * accesses be optimized into a 32-bit word access.
640 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
641 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
642 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
643 /** @defgroup mdb_node Node Flags
645 * Flags for node headers.
648 #define F_BIGDATA 0x01 /**< data put on overflow page */
649 #define F_SUBDATA 0x02 /**< data is a sub-database */
650 #define F_DUPDATA 0x04 /**< data has duplicates */
652 /** valid flags for #mdb_node_add() */
653 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
656 unsigned short mn_flags; /**< @ref mdb_node */
657 unsigned short mn_ksize; /**< key size */
658 char mn_data[1]; /**< key and data are appended here */
661 /** Size of the node header, excluding dynamic data at the end */
662 #define NODESIZE offsetof(MDB_node, mn_data)
664 /** Bit position of top word in page number, for shifting mn_flags */
665 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
667 /** Size of a node in a branch page with a given key.
668 * This is just the node header plus the key, there is no data.
670 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
672 /** Size of a node in a leaf page with a given key and data.
673 * This is node header plus key plus data size.
675 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
677 /** Address of node \b i in page \b p */
678 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
680 /** Address of the key for the node */
681 #define NODEKEY(node) (void *)((node)->mn_data)
683 /** Address of the data for a node */
684 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
686 /** Get the page number pointed to by a branch node */
687 #define NODEPGNO(node) \
688 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
689 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
690 /** Set the page number in a branch node */
691 #define SETPGNO(node,pgno) do { \
692 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
693 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
695 /** Get the size of the data in a leaf node */
696 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
697 /** Set the size of the data for a leaf node */
698 #define SETDSZ(node,size) do { \
699 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
700 /** The size of a key in a node */
701 #define NODEKSZ(node) ((node)->mn_ksize)
703 /** Copy a page number from src to dst */
705 #define COPY_PGNO(dst,src) dst = src
707 #if SIZE_MAX > 4294967295UL
708 #define COPY_PGNO(dst,src) do { \
709 unsigned short *s, *d; \
710 s = (unsigned short *)&(src); \
711 d = (unsigned short *)&(dst); \
718 #define COPY_PGNO(dst,src) do { \
719 unsigned short *s, *d; \
720 s = (unsigned short *)&(src); \
721 d = (unsigned short *)&(dst); \
727 /** The address of a key in a LEAF2 page.
728 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
729 * There are no node headers, keys are stored contiguously.
731 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
733 /** Set the \b node's key into \b key, if requested. */
734 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
735 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
737 /** Information about a single database in the environment. */
738 typedef struct MDB_db {
739 uint32_t md_pad; /**< also ksize for LEAF2 pages */
740 uint16_t md_flags; /**< @ref mdb_dbi_open */
741 uint16_t md_depth; /**< depth of this tree */
742 pgno_t md_branch_pages; /**< number of internal pages */
743 pgno_t md_leaf_pages; /**< number of leaf pages */
744 pgno_t md_overflow_pages; /**< number of overflow pages */
745 size_t md_entries; /**< number of data items */
746 pgno_t md_root; /**< the root page of this tree */
749 /** Handle for the DB used to track free pages. */
751 /** Handle for the default DB. */
754 /** Meta page content. */
755 typedef struct MDB_meta {
756 /** Stamp identifying this as an MDB file. It must be set
759 /** Version number of this lock file. Must be set to #MDB_VERSION. */
761 void *mm_address; /**< address for fixed mapping */
762 size_t mm_mapsize; /**< size of mmap region */
763 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
764 /** The size of pages used in this DB */
765 #define mm_psize mm_dbs[0].md_pad
766 /** Any persistent environment flags. @ref mdb_env */
767 #define mm_flags mm_dbs[0].md_flags
768 pgno_t mm_last_pg; /**< last used page in file */
769 txnid_t mm_txnid; /**< txnid that committed this page */
772 /** Buffer for a stack-allocated dirty page.
773 * The members define size and alignment, and silence type
774 * aliasing warnings. They are not used directly; that could
775 * mean incorrectly using several union members in parallel.
777 typedef union MDB_pagebuf {
778 char mb_raw[MDB_PAGESIZE];
781 char mm_pad[PAGEHDRSZ];
786 /** Auxiliary DB info.
787 * The information here is mostly static/read-only. There is
788 * only a single copy of this record in the environment.
790 typedef struct MDB_dbx {
791 MDB_val md_name; /**< name of the database */
792 MDB_cmp_func *md_cmp; /**< function for comparing keys */
793 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
794 MDB_rel_func *md_rel; /**< user relocate function */
795 void *md_relctx; /**< user-provided context for md_rel */
798 /** A database transaction.
799 * Every operation requires a transaction handle.
802 MDB_txn *mt_parent; /**< parent of a nested txn */
803 MDB_txn *mt_child; /**< nested txn under this txn */
804 pgno_t mt_next_pgno; /**< next unallocated page */
805 /** The ID of this transaction. IDs are integers incrementing from 1.
806 * Only committed write transactions increment the ID. If a transaction
807 * aborts, the ID may be re-used by the next writer.
810 MDB_env *mt_env; /**< the DB environment */
811 /** The list of pages that became unused during this transaction.
815 MDB_ID2L dirty_list; /**< for write txns: modified pages */
816 MDB_reader *reader; /**< this thread's reader table slot or NULL */
818 /** Array of records for each DB known in the environment. */
820 /** Array of MDB_db records for each known DB */
822 /** @defgroup mt_dbflag Transaction DB Flags
826 #define DB_DIRTY 0x01 /**< DB was written in this txn */
827 #define DB_STALE 0x02 /**< DB record is older than txnID */
829 /** In write txns, array of cursors for each DB */
830 MDB_cursor **mt_cursors;
831 /** Array of flags for each DB */
832 unsigned char *mt_dbflags;
833 /** Number of DB records in use. This number only ever increments;
834 * we don't decrement it when individual DB handles are closed.
838 /** @defgroup mdb_txn Transaction Flags
842 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
843 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
844 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
846 unsigned int mt_flags; /**< @ref mdb_txn */
847 /** Tracks which of the two meta pages was used at the start
848 * of this transaction.
850 unsigned int mt_toggle;
853 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
854 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
855 * raise this on a 64 bit machine.
857 #define CURSOR_STACK 32
861 /** Cursors are used for all DB operations */
863 /** Next cursor on this DB in this txn */
865 /** Original cursor if this is a shadow */
867 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
868 struct MDB_xcursor *mc_xcursor;
869 /** The transaction that owns this cursor */
871 /** The database handle this cursor operates on */
873 /** The database record for this cursor */
875 /** The database auxiliary record for this cursor */
877 /** The @ref mt_dbflag for this database */
878 unsigned char *mc_dbflag;
879 unsigned short mc_snum; /**< number of pushed pages */
880 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
881 /** @defgroup mdb_cursor Cursor Flags
883 * Cursor state flags.
886 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
887 #define C_EOF 0x02 /**< No more data */
888 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
889 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
890 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
891 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
893 unsigned int mc_flags; /**< @ref mdb_cursor */
894 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
895 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
898 /** Context for sorted-dup records.
899 * We could have gone to a fully recursive design, with arbitrarily
900 * deep nesting of sub-databases. But for now we only handle these
901 * levels - main DB, optional sub-DB, sorted-duplicate DB.
903 typedef struct MDB_xcursor {
904 /** A sub-cursor for traversing the Dup DB */
905 MDB_cursor mx_cursor;
906 /** The database record for this Dup DB */
908 /** The auxiliary DB record for this Dup DB */
910 /** The @ref mt_dbflag for this Dup DB */
911 unsigned char mx_dbflag;
914 /** The database environment. */
916 HANDLE me_fd; /**< The main data file */
917 HANDLE me_lfd; /**< The lock file */
918 HANDLE me_mfd; /**< just for writing the meta pages */
919 /** Failed to update the meta page. Probably an I/O error. */
920 #define MDB_FATAL_ERROR 0x80000000U
921 /** Read-only Filesystem. Allow read access, no locking. */
922 #define MDB_ROFS 0x40000000U
923 /** Some fields are initialized. */
924 #define MDB_ENV_ACTIVE 0x20000000U
925 uint32_t me_flags; /**< @ref mdb_env */
926 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
927 unsigned int me_maxreaders; /**< size of the reader table */
928 unsigned int me_numreaders; /**< max numreaders set by this env */
929 MDB_dbi me_numdbs; /**< number of DBs opened */
930 MDB_dbi me_maxdbs; /**< size of the DB table */
931 pid_t me_pid; /**< process ID of this env */
932 char *me_path; /**< path to the DB files */
933 char *me_map; /**< the memory map of the data file */
934 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
935 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
936 MDB_txn *me_txn; /**< current write transaction */
937 size_t me_mapsize; /**< size of the data memory map */
938 off_t me_size; /**< current file size */
939 pgno_t me_maxpg; /**< me_mapsize / me_psize */
940 txnid_t me_pglast; /**< ID of last old page record we used */
941 MDB_dbx *me_dbxs; /**< array of static DB info */
942 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
943 pgno_t *me_pghead; /**< old pages reclaimed from freelist */
944 pgno_t *me_pgfree; /**< memory to free when dropping me_pghead */
945 pthread_key_t me_txkey; /**< thread-key for readers */
946 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
947 /** IDL of pages that became unused in a write txn */
949 /** ID2L of pages that were written during a write txn */
950 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
951 /** Max number of freelist items that can fit in a single overflow page */
952 unsigned int me_maxfree_1pg;
954 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
956 #elif defined(MDB_USE_POSIX_SEM)
957 sem_t *me_rmutex; /* Shared mutexes are not supported */
961 /** max number of pages to commit in one writev() call */
962 #define MDB_COMMIT_PAGES 64
963 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
964 #undef MDB_COMMIT_PAGES
965 #define MDB_COMMIT_PAGES IOV_MAX
968 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
969 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
970 static int mdb_page_touch(MDB_cursor *mc);
972 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
973 static int mdb_page_search_root(MDB_cursor *mc,
974 MDB_val *key, int modify);
975 #define MDB_PS_MODIFY 1
976 #define MDB_PS_ROOTONLY 2
977 static int mdb_page_search(MDB_cursor *mc,
978 MDB_val *key, int flags);
979 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
981 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
982 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
983 pgno_t newpgno, unsigned int nflags);
985 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
986 static int mdb_env_pick_meta(const MDB_env *env);
987 static int mdb_env_write_meta(MDB_txn *txn);
988 static void mdb_env_close0(MDB_env *env, int excl);
990 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
991 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
992 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
993 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
994 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
995 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
996 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
997 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
998 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1000 static int mdb_rebalance(MDB_cursor *mc);
1001 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1003 static void mdb_cursor_pop(MDB_cursor *mc);
1004 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1006 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1007 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1008 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1009 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1010 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1012 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1013 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1015 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1016 static void mdb_xcursor_init0(MDB_cursor *mc);
1017 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1019 static int mdb_drop0(MDB_cursor *mc, int subs);
1020 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1023 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1027 static SECURITY_DESCRIPTOR mdb_null_sd;
1028 static SECURITY_ATTRIBUTES mdb_all_sa;
1029 static int mdb_sec_inited;
1032 /** Return the library version info. */
1034 mdb_version(int *major, int *minor, int *patch)
1036 if (major) *major = MDB_VERSION_MAJOR;
1037 if (minor) *minor = MDB_VERSION_MINOR;
1038 if (patch) *patch = MDB_VERSION_PATCH;
1039 return MDB_VERSION_STRING;
1042 /** Table of descriptions for MDB @ref errors */
1043 static char *const mdb_errstr[] = {
1044 "MDB_KEYEXIST: Key/data pair already exists",
1045 "MDB_NOTFOUND: No matching key/data pair found",
1046 "MDB_PAGE_NOTFOUND: Requested page not found",
1047 "MDB_CORRUPTED: Located page was wrong type",
1048 "MDB_PANIC: Update of meta page failed",
1049 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1050 "MDB_INVALID: File is not an MDB file",
1051 "MDB_MAP_FULL: Environment mapsize limit reached",
1052 "MDB_DBS_FULL: Environment maxdbs limit reached",
1053 "MDB_READERS_FULL: Environment maxreaders limit reached",
1054 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1055 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1056 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1057 "MDB_PAGE_FULL: Internal error - page has no more space",
1058 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1062 mdb_strerror(int err)
1066 return ("Successful return: 0");
1068 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1069 i = err - MDB_KEYEXIST;
1070 return mdb_errstr[i];
1073 return strerror(err);
1077 /** Display a key in hexadecimal and return the address of the result.
1078 * @param[in] key the key to display
1079 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1080 * @return The key in hexadecimal form.
1083 mdb_dkey(MDB_val *key, char *buf)
1086 unsigned char *c = key->mv_data;
1092 if (key->mv_size > MDB_MAXKEYSIZE)
1093 return "MDB_MAXKEYSIZE";
1094 /* may want to make this a dynamic check: if the key is mostly
1095 * printable characters, print it as-is instead of converting to hex.
1099 for (i=0; i<key->mv_size; i++)
1100 ptr += sprintf(ptr, "%02x", *c++);
1102 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1107 /** Display all the keys in the page. */
1109 mdb_page_list(MDB_page *mp)
1112 unsigned int i, nkeys, nsize;
1116 nkeys = NUMKEYS(mp);
1117 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1118 for (i=0; i<nkeys; i++) {
1119 node = NODEPTR(mp, i);
1120 key.mv_size = node->mn_ksize;
1121 key.mv_data = node->mn_data;
1122 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1123 if (IS_BRANCH(mp)) {
1124 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1127 if (F_ISSET(node->mn_flags, F_BIGDATA))
1128 nsize += sizeof(pgno_t);
1130 nsize += NODEDSZ(node);
1131 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1137 mdb_cursor_chk(MDB_cursor *mc)
1143 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1144 for (i=0; i<mc->mc_top; i++) {
1146 node = NODEPTR(mp, mc->mc_ki[i]);
1147 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1150 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1156 /** Count all the pages in each DB and in the freelist
1157 * and make sure it matches the actual number of pages
1160 static void mdb_audit(MDB_txn *txn)
1164 MDB_ID freecount, count;
1169 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1170 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1171 freecount += *(MDB_ID *)data.mv_data;
1174 for (i = 0; i<txn->mt_numdbs; i++) {
1175 MDB_xcursor mx, *mxp;
1176 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1177 mdb_cursor_init(&mc, txn, i, mxp);
1178 if (txn->mt_dbs[i].md_root == P_INVALID)
1180 count += txn->mt_dbs[i].md_branch_pages +
1181 txn->mt_dbs[i].md_leaf_pages +
1182 txn->mt_dbs[i].md_overflow_pages;
1183 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1184 mdb_page_search(&mc, NULL, 0);
1188 mp = mc.mc_pg[mc.mc_top];
1189 for (j=0; j<NUMKEYS(mp); j++) {
1190 MDB_node *leaf = NODEPTR(mp, j);
1191 if (leaf->mn_flags & F_SUBDATA) {
1193 memcpy(&db, NODEDATA(leaf), sizeof(db));
1194 count += db.md_branch_pages + db.md_leaf_pages +
1195 db.md_overflow_pages;
1199 while (mdb_cursor_sibling(&mc, 1) == 0);
1202 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1203 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1204 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1210 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1212 return txn->mt_dbxs[dbi].md_cmp(a, b);
1216 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1218 if (txn->mt_dbxs[dbi].md_dcmp)
1219 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1221 return EINVAL; /* too bad you can't distinguish this from a valid result */
1224 /** Allocate a single page.
1225 * Re-use old malloc'd pages first, otherwise just malloc.
1228 mdb_page_malloc(MDB_cursor *mc) {
1230 size_t sz = mc->mc_txn->mt_env->me_psize;
1231 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1232 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1233 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1234 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1235 } else if ((ret = malloc(sz)) != NULL) {
1236 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1242 mdb_page_free(MDB_env *env, MDB_page *mp)
1244 mp->mp_next = env->me_dpages;
1245 VGMEMP_FREE(env, mp);
1246 env->me_dpages = mp;
1249 /** Allocate pages for writing.
1250 * If there are free pages available from older transactions, they
1251 * will be re-used first. Otherwise a new page will be allocated.
1252 * @param[in] mc cursor A cursor handle identifying the transaction and
1253 * database for which we are allocating.
1254 * @param[in] num the number of pages to allocate.
1255 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1256 * will always be satisfied by a single contiguous chunk of memory.
1257 * @return 0 on success, non-zero on failure.
1260 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1262 MDB_txn *txn = mc->mc_txn;
1264 pgno_t pgno = P_INVALID;
1266 txnid_t oldest = 0, last;
1271 /* If our dirty list is already full, we can't do anything */
1272 if (txn->mt_u.dirty_list[0].mid >= MDB_IDL_UM_MAX)
1273 return MDB_TXN_FULL;
1275 /* The free list won't have any content at all until txn 2 has
1276 * committed. The pages freed by txn 2 will be unreferenced
1277 * after txn 3 commits, and so will be safe to re-use in txn 4.
1279 if (txn->mt_txnid > 3) {
1280 if (!txn->mt_env->me_pghead &&
1281 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1282 /* See if there's anything in the free DB */
1289 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1290 if (!txn->mt_env->me_pglast) {
1291 mdb_page_search(&m2, NULL, 0);
1292 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1293 kptr = (txnid_t *)NODEKEY(leaf);
1300 last = txn->mt_env->me_pglast + 1;
1302 key.mv_data = &last;
1303 key.mv_size = sizeof(last);
1304 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1307 last = *(txnid_t *)key.mv_data;
1313 oldest = txn->mt_txnid - 1;
1314 nr = txn->mt_env->me_txns->mti_numreaders;
1315 r = txn->mt_env->me_txns->mti_readers;
1316 for (i=0; i<nr; i++) {
1317 if (!r[i].mr_pid) continue;
1324 if (oldest > last) {
1325 /* It's usable, grab it.
1329 if (!txn->mt_env->me_pglast) {
1330 mdb_node_read(txn, leaf, &data);
1332 idl = (MDB_ID *) data.mv_data;
1333 /* We might have a zero-length IDL due to freelist growth
1334 * during a prior commit
1337 txn->mt_env->me_pglast = last;
1340 mop = malloc(MDB_IDL_SIZEOF(idl));
1343 txn->mt_env->me_pglast = last;
1344 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop;
1345 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1350 DPRINTF("IDL read txn %zu root %zu num %zu",
1351 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1352 for (i=0; i<idl[0]; i++) {
1353 DPRINTF("IDL %zu", idl[i+1]);
1360 if (txn->mt_env->me_pghead) {
1361 pgno_t *mop = txn->mt_env->me_pghead;
1364 int retry = 500, readit = 0, n2 = num-1;
1365 unsigned int i, j, k;
1367 /* If current list is too short, must fetch more and coalesce */
1368 if (mop[0] < (unsigned)num)
1371 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1373 /* If on freelist, don't try to read more. If what we have
1374 * right now isn't enough just use new pages.
1375 * TODO: get all of this working. Many circular dependencies...
1377 if (mc->mc_dbi == FREE_DBI) {
1386 last = txn->mt_env->me_pglast + 1;
1388 /* We haven't hit the readers list yet? */
1394 oldest = txn->mt_txnid - 1;
1395 nr = txn->mt_env->me_txns->mti_numreaders;
1396 r = txn->mt_env->me_txns->mti_readers;
1397 for (i=0; i<nr; i++) {
1398 if (!r[i].mr_pid) continue;
1405 /* There's nothing we can use on the freelist */
1406 if (oldest - last < 1)
1410 key.mv_data = &last;
1411 key.mv_size = sizeof(last);
1412 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1415 idl = (MDB_ID *) data.mv_data;
1416 mop2 = malloc(MDB_IDL_SIZEOF(idl) + MDB_IDL_SIZEOF(mop));
1419 /* merge in sorted order */
1420 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1422 while (i>0 || j>0) {
1423 if (i && idl[i] < mop[j])
1424 mop2[k--] = idl[i--];
1426 mop2[k--] = mop[j--];
1428 txn->mt_env->me_pglast = last;
1429 free(txn->mt_env->me_pgfree);
1430 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop2;
1432 /* Keep trying to read until we have enough */
1433 if (mop[0] < (unsigned)num) {
1438 /* current list has enough pages, but are they contiguous? */
1439 for (i=mop[0]; i>=(unsigned)num; i--) {
1440 if (mop[i-n2] == mop[i] + n2) {
1443 /* move any stragglers down */
1444 for (j=i+num; j<=mop[0]; j++)
1451 /* Stop if we succeeded, or no more retries */
1452 if (!retry || pgno != P_INVALID)
1459 /* peel pages off tail, so we only have to truncate the list */
1460 pgno = MDB_IDL_LAST(mop);
1463 if (MDB_IDL_IS_ZERO(mop)) {
1464 free(txn->mt_env->me_pgfree);
1465 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1470 if (pgno == P_INVALID) {
1471 /* DB size is maxed out */
1472 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1473 DPUTS("DB size maxed out");
1474 return MDB_MAP_FULL;
1477 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1478 if (pgno == P_INVALID) {
1479 pgno = txn->mt_next_pgno;
1480 txn->mt_next_pgno += num;
1482 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1485 if (txn->mt_env->me_dpages && num == 1) {
1486 np = txn->mt_env->me_dpages;
1487 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1488 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1489 txn->mt_env->me_dpages = np->mp_next;
1491 size_t sz = txn->mt_env->me_psize * num;
1492 if ((np = malloc(sz)) == NULL)
1494 VGMEMP_ALLOC(txn->mt_env, np, sz);
1496 if (pgno == P_INVALID) {
1497 np->mp_pgno = txn->mt_next_pgno;
1498 txn->mt_next_pgno += num;
1503 mid.mid = np->mp_pgno;
1505 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1506 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1508 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1515 /** Copy a page: avoid copying unused portions of the page.
1516 * @param[in] dst page to copy into
1517 * @param[in] src page to copy from
1520 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1522 dst->mp_flags = src->mp_flags | P_DIRTY;
1523 dst->mp_pages = src->mp_pages;
1525 if (IS_LEAF2(src)) {
1526 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1528 unsigned int i, nkeys = NUMKEYS(src);
1529 for (i=0; i<nkeys; i++)
1530 dst->mp_ptrs[i] = src->mp_ptrs[i];
1531 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1532 psize - src->mp_upper);
1536 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1537 * @param[in] mc cursor pointing to the page to be touched
1538 * @return 0 on success, non-zero on failure.
1541 mdb_page_touch(MDB_cursor *mc)
1543 MDB_page *mp = mc->mc_pg[mc->mc_top];
1547 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1549 if ((rc = mdb_page_alloc(mc, 1, &np)))
1551 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1552 assert(mp->mp_pgno != np->mp_pgno);
1553 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1555 /* If page isn't full, just copy the used portion */
1556 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1559 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1561 np->mp_flags |= P_DIRTY;
1566 /* Adjust other cursors pointing to mp */
1567 if (mc->mc_flags & C_SUB) {
1568 MDB_cursor *m2, *m3;
1569 MDB_dbi dbi = mc->mc_dbi-1;
1571 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1572 if (m2 == mc) continue;
1573 m3 = &m2->mc_xcursor->mx_cursor;
1574 if (m3->mc_snum < mc->mc_snum) continue;
1575 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1576 m3->mc_pg[mc->mc_top] = mp;
1582 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1583 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1584 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1585 m2->mc_pg[mc->mc_top] = mp;
1589 mc->mc_pg[mc->mc_top] = mp;
1590 /** If this page has a parent, update the parent to point to
1594 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1596 mc->mc_db->md_root = mp->mp_pgno;
1597 } else if (mc->mc_txn->mt_parent) {
1600 /* If txn has a parent, make sure the page is in our
1603 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1604 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1605 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1606 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1607 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1608 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1609 mc->mc_pg[mc->mc_top] = mp;
1614 if (mc->mc_txn->mt_u.dirty_list[0].mid >= MDB_IDL_UM_MAX)
1615 return MDB_TXN_FULL;
1617 np = mdb_page_malloc(mc);
1620 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1621 mid.mid = np->mp_pgno;
1623 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1631 mdb_env_sync(MDB_env *env, int force)
1634 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1635 if (env->me_flags & MDB_WRITEMAP) {
1636 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1637 ? MS_ASYNC : MS_SYNC;
1638 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1641 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1645 if (MDB_FDATASYNC(env->me_fd))
1652 /** Make shadow copies of all of parent txn's cursors */
1654 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1656 MDB_cursor *mc, *m2;
1657 unsigned int i, j, size;
1659 for (i=0;i<src->mt_numdbs; i++) {
1660 if (src->mt_cursors[i]) {
1661 size = sizeof(MDB_cursor);
1662 if (src->mt_cursors[i]->mc_xcursor)
1663 size += sizeof(MDB_xcursor);
1664 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1671 mc->mc_db = &dst->mt_dbs[i];
1672 mc->mc_dbx = m2->mc_dbx;
1673 mc->mc_dbflag = &dst->mt_dbflags[i];
1674 mc->mc_snum = m2->mc_snum;
1675 mc->mc_top = m2->mc_top;
1676 mc->mc_flags = m2->mc_flags | C_SHADOW;
1677 for (j=0; j<mc->mc_snum; j++) {
1678 mc->mc_pg[j] = m2->mc_pg[j];
1679 mc->mc_ki[j] = m2->mc_ki[j];
1681 if (m2->mc_xcursor) {
1682 MDB_xcursor *mx, *mx2;
1683 mx = (MDB_xcursor *)(mc+1);
1684 mc->mc_xcursor = mx;
1685 mx2 = m2->mc_xcursor;
1686 mx->mx_db = mx2->mx_db;
1687 mx->mx_dbx = mx2->mx_dbx;
1688 mx->mx_dbflag = mx2->mx_dbflag;
1689 mx->mx_cursor.mc_txn = dst;
1690 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1691 mx->mx_cursor.mc_db = &mx->mx_db;
1692 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1693 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1694 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1695 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1696 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1697 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1698 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1699 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1702 mc->mc_xcursor = NULL;
1704 mc->mc_next = dst->mt_cursors[i];
1705 dst->mt_cursors[i] = mc;
1712 /** Merge shadow cursors back into parent's */
1714 mdb_cursor_merge(MDB_txn *txn)
1717 for (i=0; i<txn->mt_numdbs; i++) {
1718 if (txn->mt_cursors[i]) {
1720 while ((mc = txn->mt_cursors[i])) {
1721 txn->mt_cursors[i] = mc->mc_next;
1722 if (mc->mc_flags & C_SHADOW) {
1723 MDB_cursor *m2 = mc->mc_orig;
1725 m2->mc_snum = mc->mc_snum;
1726 m2->mc_top = mc->mc_top;
1727 for (j=0; j<mc->mc_snum; j++) {
1728 m2->mc_pg[j] = mc->mc_pg[j];
1729 m2->mc_ki[j] = mc->mc_ki[j];
1732 if (mc->mc_flags & C_ALLOCD)
1740 mdb_txn_reset0(MDB_txn *txn);
1742 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1743 * @param[in] txn the transaction handle to initialize
1744 * @return 0 on success, non-zero on failure. This can only
1745 * fail for read-only transactions, and then only if the
1746 * reader table is full.
1749 mdb_txn_renew0(MDB_txn *txn)
1751 MDB_env *env = txn->mt_env;
1756 txn->mt_numdbs = env->me_numdbs;
1757 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1759 if (txn->mt_flags & MDB_TXN_RDONLY) {
1760 if (env->me_flags & MDB_ROFS) {
1761 i = mdb_env_pick_meta(env);
1762 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1763 txn->mt_u.reader = NULL;
1765 MDB_reader *r = pthread_getspecific(env->me_txkey);
1767 pid_t pid = env->me_pid;
1768 pthread_t tid = pthread_self();
1771 for (i=0; i<env->me_txns->mti_numreaders; i++)
1772 if (env->me_txns->mti_readers[i].mr_pid == 0)
1774 if (i == env->me_maxreaders) {
1775 UNLOCK_MUTEX_R(env);
1776 return MDB_READERS_FULL;
1778 env->me_txns->mti_readers[i].mr_pid = pid;
1779 env->me_txns->mti_readers[i].mr_tid = tid;
1780 if (i >= env->me_txns->mti_numreaders)
1781 env->me_txns->mti_numreaders = i+1;
1782 /* Save numreaders for un-mutexed mdb_env_close() */
1783 env->me_numreaders = env->me_txns->mti_numreaders;
1784 UNLOCK_MUTEX_R(env);
1785 r = &env->me_txns->mti_readers[i];
1786 if ((rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1787 env->me_txns->mti_readers[i].mr_pid = 0;
1791 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1792 txn->mt_u.reader = r;
1794 txn->mt_toggle = txn->mt_txnid & 1;
1795 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1799 txn->mt_txnid = env->me_txns->mti_txnid;
1800 txn->mt_toggle = txn->mt_txnid & 1;
1801 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1804 if (txn->mt_txnid == mdb_debug_start)
1807 txn->mt_u.dirty_list = env->me_dirty_list;
1808 txn->mt_u.dirty_list[0].mid = 0;
1809 txn->mt_free_pgs = env->me_free_pgs;
1810 txn->mt_free_pgs[0] = 0;
1814 /* Copy the DB info and flags */
1815 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1816 for (i=2; i<txn->mt_numdbs; i++)
1817 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1818 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1819 if (txn->mt_numdbs > 2)
1820 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1822 if (env->me_maxpg < txn->mt_next_pgno) {
1823 mdb_txn_reset0(txn);
1824 return MDB_MAP_RESIZED;
1831 mdb_txn_renew(MDB_txn *txn)
1835 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1838 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1839 DPUTS("environment had fatal error, must shutdown!");
1843 rc = mdb_txn_renew0(txn);
1844 if (rc == MDB_SUCCESS) {
1845 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1846 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1847 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1853 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1858 if (env->me_flags & MDB_FATAL_ERROR) {
1859 DPUTS("environment had fatal error, must shutdown!");
1862 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1865 /* Nested transactions: Max 1 child, write txns only, no writemap */
1866 if (parent->mt_child ||
1867 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1868 (env->me_flags & MDB_WRITEMAP))
1873 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1874 if (!(flags & MDB_RDONLY))
1875 size += env->me_maxdbs * sizeof(MDB_cursor *);
1877 if ((txn = calloc(1, size)) == NULL) {
1878 DPRINTF("calloc: %s", strerror(ErrCode()));
1881 txn->mt_dbs = (MDB_db *)(txn+1);
1882 if (flags & MDB_RDONLY) {
1883 txn->mt_flags |= MDB_TXN_RDONLY;
1884 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1886 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1887 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1892 txn->mt_free_pgs = mdb_midl_alloc();
1893 if (!txn->mt_free_pgs) {
1897 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1898 if (!txn->mt_u.dirty_list) {
1899 free(txn->mt_free_pgs);
1903 txn->mt_txnid = parent->mt_txnid;
1904 txn->mt_toggle = parent->mt_toggle;
1905 txn->mt_u.dirty_list[0].mid = 0;
1906 txn->mt_free_pgs[0] = 0;
1907 txn->mt_next_pgno = parent->mt_next_pgno;
1908 parent->mt_child = txn;
1909 txn->mt_parent = parent;
1910 txn->mt_numdbs = parent->mt_numdbs;
1911 txn->mt_dbxs = parent->mt_dbxs;
1912 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1913 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1914 mdb_cursor_shadow(parent, txn);
1917 rc = mdb_txn_renew0(txn);
1923 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1924 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1925 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1931 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1932 * @param[in] txn the transaction handle to reset
1935 mdb_txn_reset0(MDB_txn *txn)
1937 MDB_env *env = txn->mt_env;
1939 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1940 if (!(env->me_flags & MDB_ROFS))
1941 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1946 /* close(free) all cursors */
1947 for (i=0; i<txn->mt_numdbs; i++) {
1948 if (txn->mt_cursors[i]) {
1950 while ((mc = txn->mt_cursors[i])) {
1951 txn->mt_cursors[i] = mc->mc_next;
1952 if (mc->mc_flags & C_ALLOCD)
1958 if (!(env->me_flags & MDB_WRITEMAP)) {
1959 /* return all dirty pages to dpage list */
1960 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1961 dp = txn->mt_u.dirty_list[i].mptr;
1962 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1963 mdb_page_free(txn->mt_env, dp);
1965 /* large pages just get freed directly */
1966 VGMEMP_FREE(txn->mt_env, dp);
1972 if (txn->mt_parent) {
1973 txn->mt_parent->mt_child = NULL;
1974 mdb_midl_free(txn->mt_free_pgs);
1975 free(txn->mt_u.dirty_list);
1978 if (mdb_midl_shrink(&txn->mt_free_pgs))
1979 env->me_free_pgs = txn->mt_free_pgs;
1982 free(txn->mt_env->me_pgfree);
1983 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1984 txn->mt_env->me_pglast = 0;
1987 /* The writer mutex was locked in mdb_txn_begin. */
1988 UNLOCK_MUTEX_W(env);
1993 mdb_txn_reset(MDB_txn *txn)
1998 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1999 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2000 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2002 mdb_txn_reset0(txn);
2006 mdb_txn_abort(MDB_txn *txn)
2011 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2012 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2013 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2016 mdb_txn_abort(txn->mt_child);
2018 mdb_txn_reset0(txn);
2023 mdb_txn_commit(MDB_txn *txn)
2031 pgno_t next, freecnt;
2032 txnid_t oldpg_txnid, id;
2035 assert(txn != NULL);
2036 assert(txn->mt_env != NULL);
2038 if (txn->mt_child) {
2039 mdb_txn_commit(txn->mt_child);
2040 txn->mt_child = NULL;
2045 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2046 if (txn->mt_numdbs > env->me_numdbs) {
2047 /* update the DB flags */
2049 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2050 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2057 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2058 DPUTS("error flag is set, can't commit");
2060 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2065 if (txn->mt_parent) {
2071 /* Merge (and close) our cursors with parent's */
2072 mdb_cursor_merge(txn);
2074 /* Update parent's DB table */
2075 ip = &txn->mt_parent->mt_dbs[2];
2076 jp = &txn->mt_dbs[2];
2077 for (i = 2; i < txn->mt_numdbs; i++) {
2078 if (ip->md_root != jp->md_root)
2082 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2084 /* Append our free list to parent's */
2085 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
2087 mdb_midl_free(txn->mt_free_pgs);
2089 /* Merge our dirty list with parent's */
2090 dst = txn->mt_parent->mt_u.dirty_list;
2091 src = txn->mt_u.dirty_list;
2092 x = mdb_mid2l_search(dst, src[1].mid);
2093 for (y=1; y<=src[0].mid; y++) {
2094 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
2098 dst[x].mptr = src[y].mptr;
2101 for (; y<=src[0].mid; y++) {
2102 if (++x >= MDB_IDL_UM_MAX) {
2104 return MDB_TXN_FULL;
2109 free(txn->mt_u.dirty_list);
2110 txn->mt_parent->mt_child = NULL;
2115 if (txn != env->me_txn) {
2116 DPUTS("attempt to commit unknown transaction");
2121 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2124 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2125 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2127 /* Update DB root pointers. Their pages have already been
2128 * touched so this is all in-place and cannot fail.
2130 if (txn->mt_numdbs > 2) {
2133 data.mv_size = sizeof(MDB_db);
2135 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2136 for (i = 2; i < txn->mt_numdbs; i++) {
2137 if (txn->mt_dbflags[i] & DB_DIRTY) {
2138 data.mv_data = &txn->mt_dbs[i];
2139 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2144 /* Save the freelist as of this transaction to the freeDB. This
2145 * can change the freelist, so keep trying until it stabilizes.
2147 * env->me_pglast and the length of txn->mt_free_pgs cannot decrease,
2148 * except the code below can decrease env->me_pglast to split pghead.
2149 * Page numbers cannot disappear from txn->mt_free_pgs. New pages
2150 * can only appear in env->me_pghead when env->me_pglast increases.
2151 * Until then, the me_pghead pointer won't move but can become NULL.
2154 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2155 oldpg_txnid = id = 0;
2158 /* should only be one record now */
2159 if (env->me_pghead || env->me_pglast) {
2160 /* make sure first page of freeDB is touched and on freelist */
2161 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2162 if (rc && rc != MDB_NOTFOUND) {
2169 /* Delete IDLs we used from the free list */
2170 if (env->me_pglast) {
2175 rc = mdb_cursor_first(&mc, &key, NULL);
2178 oldpg_txnid = *(txnid_t *)key.mv_data;
2180 assert(oldpg_txnid <= env->me_pglast);
2182 rc = mdb_cursor_del(&mc, 0);
2185 } while (oldpg_txnid < env->me_pglast);
2188 /* Save IDL of pages freed by this txn, to freeDB */
2190 if (freecnt != txn->mt_free_pgs[0]) {
2193 /* make sure last page of freeDB is touched and on freelist */
2194 key.mv_size = MDB_MAXKEYSIZE+1;
2196 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2197 if (rc && rc != MDB_NOTFOUND)
2203 MDB_IDL idl = txn->mt_free_pgs;
2204 mdb_midl_sort(txn->mt_free_pgs);
2205 DPRINTF("IDL write txn %zu root %zu num %zu",
2206 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2207 for (i=1; i<=idl[0]; i++) {
2208 DPRINTF("IDL %zu", idl[i]);
2212 /* write to last page of freeDB */
2213 key.mv_size = sizeof(pgno_t);
2214 key.mv_data = &txn->mt_txnid;
2215 /* The free list can still grow during this call,
2216 * despite the pre-emptive touches above. So retry
2217 * until the reserved space remains big enough.
2220 assert(freecnt < txn->mt_free_pgs[0]);
2221 freecnt = txn->mt_free_pgs[0];
2222 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2223 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2226 } while (freecnt != txn->mt_free_pgs[0]);
2227 mdb_midl_sort(txn->mt_free_pgs);
2228 memcpy(data.mv_data, txn->mt_free_pgs, data.mv_size);
2229 if (oldpg_txnid < env->me_pglast || (!env->me_pghead && id))
2230 goto free_pgfirst; /* used up freeDB[oldpg_txnid] */
2233 /* Put back page numbers we took from freeDB but did not use */
2234 if (env->me_pghead) {
2239 mop = env->me_pghead;
2240 id = env->me_pglast;
2241 key.mv_size = sizeof(id);
2243 /* These steps may grow the freelist again
2244 * due to freed overflow pages...
2249 if (orig > env->me_maxfree_1pg && id > 4)
2250 orig = env->me_maxfree_1pg; /* Do not use more than 1 page */
2251 data.mv_size = (orig + 1) * sizeof(pgno_t);
2252 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2255 assert(!env->me_pghead || env->me_pglast);
2256 /* mop could have been used again here */
2257 if (id != env->me_pglast || env->me_pghead == NULL)
2258 goto again; /* was completely used up */
2259 assert(mop == env->me_pghead);
2260 } while (mop[0] < orig && --i);
2261 memcpy(data.mv_data, mop, data.mv_size);
2264 *(pgno_t *)data.mv_data = orig;
2265 mop[orig] = mop[0] - orig;
2266 env->me_pghead = mop += orig;
2267 /* Save more oldpages at the previous txnid. */
2268 assert(env->me_pglast == id && id == oldpg_txnid);
2269 env->me_pglast = --oldpg_txnid;
2273 /* Check for growth of freelist again */
2274 if (freecnt != txn->mt_free_pgs[0])
2277 free(env->me_pgfree);
2278 env->me_pghead = env->me_pgfree = NULL;
2280 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2281 if (mdb_midl_shrink(&txn->mt_free_pgs))
2282 env->me_free_pgs = txn->mt_free_pgs;
2289 if (env->me_flags & MDB_WRITEMAP) {
2290 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2291 dp = txn->mt_u.dirty_list[i].mptr;
2292 /* clear dirty flag */
2293 dp->mp_flags &= ~P_DIRTY;
2294 txn->mt_u.dirty_list[i].mid = 0;
2296 txn->mt_u.dirty_list[0].mid = 0;
2300 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2306 /* Windows actually supports scatter/gather I/O, but only on
2307 * unbuffered file handles. Since we're relying on the OS page
2308 * cache for all our data, that's self-defeating. So we just
2309 * write pages one at a time. We use the ov structure to set
2310 * the write offset, to at least save the overhead of a Seek
2314 memset(&ov, 0, sizeof(ov));
2315 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2317 dp = txn->mt_u.dirty_list[i].mptr;
2318 DPRINTF("committing page %zu", dp->mp_pgno);
2319 size = dp->mp_pgno * env->me_psize;
2320 ov.Offset = size & 0xffffffff;
2321 ov.OffsetHigh = size >> 16;
2322 ov.OffsetHigh >>= 16;
2323 /* clear dirty flag */
2324 dp->mp_flags &= ~P_DIRTY;
2325 wsize = env->me_psize;
2326 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2327 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2330 DPRINTF("WriteFile: %d", n);
2337 struct iovec iov[MDB_COMMIT_PAGES];
2341 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2342 dp = txn->mt_u.dirty_list[i].mptr;
2343 if (dp->mp_pgno != next) {
2345 rc = writev(env->me_fd, iov, n);
2349 DPUTS("short write, filesystem full?");
2351 DPRINTF("writev: %s", strerror(n));
2358 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2361 DPRINTF("committing page %zu", dp->mp_pgno);
2362 iov[n].iov_len = env->me_psize;
2363 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2364 iov[n].iov_base = (char *)dp;
2365 size += iov[n].iov_len;
2366 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2367 /* clear dirty flag */
2368 dp->mp_flags &= ~P_DIRTY;
2369 if (++n >= MDB_COMMIT_PAGES) {
2379 rc = writev(env->me_fd, iov, n);
2383 DPUTS("short write, filesystem full?");
2385 DPRINTF("writev: %s", strerror(n));
2392 /* Drop the dirty pages.
2394 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2395 dp = txn->mt_u.dirty_list[i].mptr;
2396 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2397 mdb_page_free(txn->mt_env, dp);
2399 VGMEMP_FREE(txn->mt_env, dp);
2402 txn->mt_u.dirty_list[i].mid = 0;
2404 txn->mt_u.dirty_list[0].mid = 0;
2407 if ((n = mdb_env_sync(env, 0)) != 0 ||
2408 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2416 if (txn->mt_numdbs > env->me_numdbs) {
2417 /* update the DB flags */
2419 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2420 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2424 UNLOCK_MUTEX_W(env);
2430 /** Read the environment parameters of a DB environment before
2431 * mapping it into memory.
2432 * @param[in] env the environment handle
2433 * @param[out] meta address of where to store the meta information
2434 * @return 0 on success, non-zero on failure.
2437 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2444 /* We don't know the page size yet, so use a minimum value.
2445 * Read both meta pages so we can use the latest one.
2448 for (i=0; i<2; i++) {
2450 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2452 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2457 else if (rc != MDB_PAGESIZE) {
2461 DPRINTF("read: %s", strerror(err));
2465 p = (MDB_page *)&pbuf;
2467 if (!F_ISSET(p->mp_flags, P_META)) {
2468 DPRINTF("page %zu not a meta page", p->mp_pgno);
2473 if (m->mm_magic != MDB_MAGIC) {
2474 DPUTS("meta has invalid magic");
2478 if (m->mm_version != MDB_VERSION) {
2479 DPRINTF("database is version %u, expected version %u",
2480 m->mm_version, MDB_VERSION);
2481 return MDB_VERSION_MISMATCH;
2485 if (m->mm_txnid > meta->mm_txnid)
2486 memcpy(meta, m, sizeof(*m));
2488 memcpy(meta, m, sizeof(*m));
2490 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2492 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2500 /** Write the environment parameters of a freshly created DB environment.
2501 * @param[in] env the environment handle
2502 * @param[out] meta address of where to store the meta information
2503 * @return 0 on success, non-zero on failure.
2506 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2513 DPUTS("writing new meta page");
2515 GET_PAGESIZE(psize);
2517 meta->mm_magic = MDB_MAGIC;
2518 meta->mm_version = MDB_VERSION;
2519 meta->mm_mapsize = env->me_mapsize;
2520 meta->mm_psize = psize;
2521 meta->mm_last_pg = 1;
2522 meta->mm_flags = env->me_flags & 0xffff;
2523 meta->mm_flags |= MDB_INTEGERKEY;
2524 meta->mm_dbs[0].md_root = P_INVALID;
2525 meta->mm_dbs[1].md_root = P_INVALID;
2527 p = calloc(2, psize);
2529 p->mp_flags = P_META;
2532 memcpy(m, meta, sizeof(*meta));
2534 q = (MDB_page *)((char *)p + psize);
2537 q->mp_flags = P_META;
2540 memcpy(m, meta, sizeof(*meta));
2545 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2546 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2547 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2550 lseek(env->me_fd, 0, SEEK_SET);
2551 rc = write(env->me_fd, p, psize * 2);
2552 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2558 /** Update the environment info to commit a transaction.
2559 * @param[in] txn the transaction that's being committed
2560 * @return 0 on success, non-zero on failure.
2563 mdb_env_write_meta(MDB_txn *txn)
2566 MDB_meta meta, metab, *mp;
2568 int rc, len, toggle;
2575 assert(txn != NULL);
2576 assert(txn->mt_env != NULL);
2578 toggle = !txn->mt_toggle;
2579 DPRINTF("writing meta page %d for root page %zu",
2580 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2583 mp = env->me_metas[toggle];
2585 if (env->me_flags & MDB_WRITEMAP) {
2586 /* Persist any increases of mapsize config */
2587 if (env->me_mapsize > mp->mm_mapsize)
2588 mp->mm_mapsize = env->me_mapsize;
2589 mp->mm_dbs[0] = txn->mt_dbs[0];
2590 mp->mm_dbs[1] = txn->mt_dbs[1];
2591 mp->mm_last_pg = txn->mt_next_pgno - 1;
2592 mp->mm_txnid = txn->mt_txnid;
2593 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2594 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2597 ptr += env->me_psize;
2598 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2605 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2606 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2608 ptr = (char *)&meta;
2609 if (env->me_mapsize > mp->mm_mapsize) {
2610 /* Persist any increases of mapsize config */
2611 meta.mm_mapsize = env->me_mapsize;
2612 off = offsetof(MDB_meta, mm_mapsize);
2614 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2616 len = sizeof(MDB_meta) - off;
2619 meta.mm_dbs[0] = txn->mt_dbs[0];
2620 meta.mm_dbs[1] = txn->mt_dbs[1];
2621 meta.mm_last_pg = txn->mt_next_pgno - 1;
2622 meta.mm_txnid = txn->mt_txnid;
2625 off += env->me_psize;
2628 /* Write to the SYNC fd */
2629 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2630 env->me_fd : env->me_mfd;
2633 memset(&ov, 0, sizeof(ov));
2635 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2638 rc = pwrite(mfd, ptr, len, off);
2643 DPUTS("write failed, disk error?");
2644 /* On a failure, the pagecache still contains the new data.
2645 * Write some old data back, to prevent it from being used.
2646 * Use the non-SYNC fd; we know it will fail anyway.
2648 meta.mm_last_pg = metab.mm_last_pg;
2649 meta.mm_txnid = metab.mm_txnid;
2651 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2653 r2 = pwrite(env->me_fd, ptr, len, off);
2656 env->me_flags |= MDB_FATAL_ERROR;
2660 /* Memory ordering issues are irrelevant; since the entire writer
2661 * is wrapped by wmutex, all of these changes will become visible
2662 * after the wmutex is unlocked. Since the DB is multi-version,
2663 * readers will get consistent data regardless of how fresh or
2664 * how stale their view of these values is.
2666 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2671 /** Check both meta pages to see which one is newer.
2672 * @param[in] env the environment handle
2673 * @return meta toggle (0 or 1).
2676 mdb_env_pick_meta(const MDB_env *env)
2678 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2682 mdb_env_create(MDB_env **env)
2686 e = calloc(1, sizeof(MDB_env));
2690 e->me_free_pgs = mdb_midl_alloc();
2691 if (!e->me_free_pgs) {
2695 e->me_maxreaders = DEFAULT_READERS;
2697 e->me_fd = INVALID_HANDLE_VALUE;
2698 e->me_lfd = INVALID_HANDLE_VALUE;
2699 e->me_mfd = INVALID_HANDLE_VALUE;
2700 #ifdef MDB_USE_POSIX_SEM
2701 e->me_rmutex = SEM_FAILED;
2702 e->me_wmutex = SEM_FAILED;
2704 e->me_pid = getpid();
2705 VGMEMP_CREATE(e,0,0);
2711 mdb_env_set_mapsize(MDB_env *env, size_t size)
2715 env->me_mapsize = size;
2717 env->me_maxpg = env->me_mapsize / env->me_psize;
2722 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2726 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2731 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2733 if (env->me_map || readers < 1)
2735 env->me_maxreaders = readers;
2740 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2742 if (!env || !readers)
2744 *readers = env->me_maxreaders;
2748 /** Further setup required for opening an MDB environment
2751 mdb_env_open2(MDB_env *env)
2753 unsigned int flags = env->me_flags;
2754 int i, newenv = 0, prot;
2758 memset(&meta, 0, sizeof(meta));
2760 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2763 DPUTS("new mdbenv");
2767 /* Was a mapsize configured? */
2768 if (!env->me_mapsize) {
2769 /* If this is a new environment, take the default,
2770 * else use the size recorded in the existing env.
2772 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2773 } else if (env->me_mapsize < meta.mm_mapsize) {
2774 /* If the configured size is smaller, make sure it's
2775 * still big enough. Silently round up to minimum if not.
2777 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2778 if (env->me_mapsize < minsize)
2779 env->me_mapsize = minsize;
2785 LONG sizelo, sizehi;
2786 sizelo = env->me_mapsize & 0xffffffff;
2787 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2789 /* Windows won't create mappings for zero length files.
2790 * Just allocate the maxsize right now.
2793 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2794 if (!SetEndOfFile(env->me_fd))
2796 SetFilePointer(env->me_fd, 0, NULL, 0);
2798 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2799 PAGE_READWRITE : PAGE_READONLY,
2800 sizehi, sizelo, NULL);
2803 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2804 FILE_MAP_WRITE : FILE_MAP_READ,
2805 0, 0, env->me_mapsize, meta.mm_address);
2813 if (flags & MDB_WRITEMAP) {
2815 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2818 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2820 if (env->me_map == MAP_FAILED) {
2827 if (flags & MDB_FIXEDMAP)
2828 meta.mm_address = env->me_map;
2829 i = mdb_env_init_meta(env, &meta);
2830 if (i != MDB_SUCCESS) {
2833 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2834 /* Can happen because the address argument to mmap() is just a
2835 * hint. mmap() can pick another, e.g. if the range is in use.
2836 * The MAP_FIXED flag would prevent that, but then mmap could
2837 * instead unmap existing pages to make room for the new map.
2839 return EBUSY; /* TODO: Make a new MDB_* error code? */
2841 env->me_psize = meta.mm_psize;
2842 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2844 env->me_maxpg = env->me_mapsize / env->me_psize;
2846 p = (MDB_page *)env->me_map;
2847 env->me_metas[0] = METADATA(p);
2848 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2852 int toggle = mdb_env_pick_meta(env);
2853 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2855 DPRINTF("opened database version %u, pagesize %u",
2856 env->me_metas[0]->mm_version, env->me_psize);
2857 DPRINTF("using meta page %d", toggle);
2858 DPRINTF("depth: %u", db->md_depth);
2859 DPRINTF("entries: %zu", db->md_entries);
2860 DPRINTF("branch pages: %zu", db->md_branch_pages);
2861 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2862 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2863 DPRINTF("root: %zu", db->md_root);
2871 /** Release a reader thread's slot in the reader lock table.
2872 * This function is called automatically when a thread exits.
2873 * @param[in] ptr This points to the slot in the reader lock table.
2876 mdb_env_reader_dest(void *ptr)
2878 MDB_reader *reader = ptr;
2884 /** Junk for arranging thread-specific callbacks on Windows. This is
2885 * necessarily platform and compiler-specific. Windows supports up
2886 * to 1088 keys. Let's assume nobody opens more than 64 environments
2887 * in a single process, for now. They can override this if needed.
2889 #ifndef MAX_TLS_KEYS
2890 #define MAX_TLS_KEYS 64
2892 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2893 static int mdb_tls_nkeys;
2895 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2899 case DLL_PROCESS_ATTACH: break;
2900 case DLL_THREAD_ATTACH: break;
2901 case DLL_THREAD_DETACH:
2902 for (i=0; i<mdb_tls_nkeys; i++) {
2903 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2904 mdb_env_reader_dest(r);
2907 case DLL_PROCESS_DETACH: break;
2912 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2914 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2918 /* Force some symbol references.
2919 * _tls_used forces the linker to create the TLS directory if not already done
2920 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2922 #pragma comment(linker, "/INCLUDE:_tls_used")
2923 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2924 #pragma const_seg(".CRT$XLB")
2925 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2926 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2929 #pragma comment(linker, "/INCLUDE:__tls_used")
2930 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2931 #pragma data_seg(".CRT$XLB")
2932 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2934 #endif /* WIN 32/64 */
2935 #endif /* !__GNUC__ */
2938 /** Downgrade the exclusive lock on the region back to shared */
2940 mdb_env_share_locks(MDB_env *env, int *excl)
2942 int rc = 0, toggle = mdb_env_pick_meta(env);
2944 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2949 /* First acquire a shared lock. The Unlock will
2950 * then release the existing exclusive lock.
2952 memset(&ov, 0, sizeof(ov));
2953 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2956 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2962 struct flock lock_info;
2963 /* The shared lock replaces the existing lock */
2964 memset((void *)&lock_info, 0, sizeof(lock_info));
2965 lock_info.l_type = F_RDLCK;
2966 lock_info.l_whence = SEEK_SET;
2967 lock_info.l_start = 0;
2968 lock_info.l_len = 1;
2969 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2970 (rc = ErrCode()) == EINTR) ;
2971 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
2978 /** Try to get exlusive lock, otherwise shared.
2979 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
2982 mdb_env_excl_lock(MDB_env *env, int *excl)
2986 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2990 memset(&ov, 0, sizeof(ov));
2991 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2998 struct flock lock_info;
2999 memset((void *)&lock_info, 0, sizeof(lock_info));
3000 lock_info.l_type = F_WRLCK;
3001 lock_info.l_whence = SEEK_SET;
3002 lock_info.l_start = 0;
3003 lock_info.l_len = 1;
3004 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3005 (rc = ErrCode()) == EINTR) ;
3009 # ifdef MDB_USE_POSIX_SEM
3010 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3013 lock_info.l_type = F_RDLCK;
3014 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3015 (rc = ErrCode()) == EINTR) ;
3023 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3025 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3027 * @(#) $Revision: 5.1 $
3028 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3029 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3031 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3035 * Please do not copyright this code. This code is in the public domain.
3037 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3038 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3039 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3040 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3041 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3042 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3043 * PERFORMANCE OF THIS SOFTWARE.
3046 * chongo <Landon Curt Noll> /\oo/\
3047 * http://www.isthe.com/chongo/
3049 * Share and Enjoy! :-)
3052 typedef unsigned long long mdb_hash_t;
3053 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3055 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3056 * @param[in] str string to hash
3057 * @param[in] hval initial value for hash
3058 * @return 64 bit hash
3060 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3061 * hval arg on the first call.
3064 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3066 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3067 unsigned char *end = s + val->mv_size;
3069 * FNV-1a hash each octet of the string
3072 /* xor the bottom with the current octet */
3073 hval ^= (mdb_hash_t)*s++;
3075 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3076 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3077 (hval << 7) + (hval << 8) + (hval << 40);
3079 /* return our new hash value */
3083 /** Hash the string and output the hash in hex.
3084 * @param[in] str string to hash
3085 * @param[out] hexbuf an array of 17 chars to hold the hash
3088 mdb_hash_hex(MDB_val *val, char *hexbuf)
3091 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3092 for (i=0; i<8; i++) {
3093 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3099 /** Open and/or initialize the lock region for the environment.
3100 * @param[in] env The MDB environment.
3101 * @param[in] lpath The pathname of the file used for the lock region.
3102 * @param[in] mode The Unix permissions for the file, if we create it.
3103 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3104 * @return 0 on success, non-zero on failure.
3107 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3115 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3116 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3117 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
3119 if (rc == ERROR_WRITE_PROTECT && (env->me_flags & MDB_RDONLY)) {
3120 env->me_flags |= MDB_ROFS;
3125 /* Try to get exclusive lock. If we succeed, then
3126 * nobody is using the lock region and we should initialize it.
3128 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3129 size = GetFileSize(env->me_lfd, NULL);
3135 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
3137 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3138 env->me_flags |= MDB_ROFS;
3143 /* Lose record locks when exec*() */
3144 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3145 fcntl(env->me_lfd, F_SETFD, fdflags);
3147 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
3148 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1) {
3150 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3151 env->me_flags |= MDB_ROFS;
3158 /* Try to get exclusive lock. If we succeed, then
3159 * nobody is using the lock region and we should initialize it.
3161 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3163 size = lseek(env->me_lfd, 0, SEEK_END);
3165 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3166 if (size < rsize && *excl > 0) {
3168 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3169 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3171 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3175 size = rsize - sizeof(MDB_txninfo);
3176 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3181 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3183 if (!mh) goto fail_errno;
3184 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3186 if (!env->me_txns) goto fail_errno;
3188 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3190 if (m == MAP_FAILED) goto fail_errno;
3196 BY_HANDLE_FILE_INFORMATION stbuf;
3205 if (!mdb_sec_inited) {
3206 InitializeSecurityDescriptor(&mdb_null_sd,
3207 SECURITY_DESCRIPTOR_REVISION);
3208 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3209 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3210 mdb_all_sa.bInheritHandle = FALSE;
3211 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3214 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3215 idbuf.volume = stbuf.dwVolumeSerialNumber;
3216 idbuf.nhigh = stbuf.nFileIndexHigh;
3217 idbuf.nlow = stbuf.nFileIndexLow;
3218 val.mv_data = &idbuf;
3219 val.mv_size = sizeof(idbuf);
3220 mdb_hash_hex(&val, hexbuf);
3221 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3222 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3223 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3224 if (!env->me_rmutex) goto fail_errno;
3225 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3226 if (!env->me_wmutex) goto fail_errno;
3227 #elif defined(MDB_USE_POSIX_SEM)
3236 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3237 idbuf.dev = stbuf.st_dev;
3238 idbuf.ino = stbuf.st_ino;
3239 val.mv_data = &idbuf;
3240 val.mv_size = sizeof(idbuf);
3241 mdb_hash_hex(&val, hexbuf);
3242 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3243 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3244 /* Clean up after a previous run, if needed: Try to
3245 * remove both semaphores before doing anything else.
3247 sem_unlink(env->me_txns->mti_rmname);
3248 sem_unlink(env->me_txns->mti_wmname);
3249 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3250 O_CREAT|O_EXCL, mode, 1);
3251 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3252 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3253 O_CREAT|O_EXCL, mode, 1);
3254 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3255 #else /* MDB_USE_POSIX_SEM */
3256 pthread_mutexattr_t mattr;
3258 if ((rc = pthread_mutexattr_init(&mattr))
3259 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3260 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3261 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3263 pthread_mutexattr_destroy(&mattr);
3264 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3266 env->me_txns->mti_version = MDB_VERSION;
3267 env->me_txns->mti_magic = MDB_MAGIC;
3268 env->me_txns->mti_txnid = 0;
3269 env->me_txns->mti_numreaders = 0;
3272 if (env->me_txns->mti_magic != MDB_MAGIC) {
3273 DPUTS("lock region has invalid magic");
3277 if (env->me_txns->mti_version != MDB_VERSION) {
3278 DPRINTF("lock region is version %u, expected version %u",
3279 env->me_txns->mti_version, MDB_VERSION);
3280 rc = MDB_VERSION_MISMATCH;
3284 if (rc != EACCES && rc != EAGAIN) {
3288 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3289 if (!env->me_rmutex) goto fail_errno;
3290 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3291 if (!env->me_wmutex) goto fail_errno;
3292 #elif defined(MDB_USE_POSIX_SEM)
3293 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3294 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3295 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3296 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3307 /** The name of the lock file in the DB environment */
3308 #define LOCKNAME "/lock.mdb"
3309 /** The name of the data file in the DB environment */
3310 #define DATANAME "/data.mdb"
3311 /** The suffix of the lock file when no subdir is used */
3312 #define LOCKSUFF "-lock"
3313 /** Only a subset of the @ref mdb_env flags can be changed
3314 * at runtime. Changing other flags requires closing the
3315 * environment and re-opening it with the new flags.
3317 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3318 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP)
3321 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3323 int oflags, rc, len, excl;
3324 char *lpath, *dpath;
3326 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3330 if (flags & MDB_NOSUBDIR) {
3331 rc = len + sizeof(LOCKSUFF) + len + 1;
3333 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3338 if (flags & MDB_NOSUBDIR) {
3339 dpath = lpath + len + sizeof(LOCKSUFF);
3340 sprintf(lpath, "%s" LOCKSUFF, path);
3341 strcpy(dpath, path);
3343 dpath = lpath + len + sizeof(LOCKNAME);
3344 sprintf(lpath, "%s" LOCKNAME, path);
3345 sprintf(dpath, "%s" DATANAME, path);
3348 flags |= env->me_flags;
3349 /* silently ignore WRITEMAP if we're only getting read access */
3350 if (F_ISSET(flags, MDB_RDONLY|MDB_WRITEMAP))
3351 flags ^= MDB_WRITEMAP;
3352 env->me_flags = flags |= MDB_ENV_ACTIVE;
3354 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3359 if (F_ISSET(flags, MDB_RDONLY)) {
3360 oflags = GENERIC_READ;
3361 len = OPEN_EXISTING;
3363 oflags = GENERIC_READ|GENERIC_WRITE;
3366 mode = FILE_ATTRIBUTE_NORMAL;
3367 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3368 NULL, len, mode, NULL);
3370 if (F_ISSET(flags, MDB_RDONLY))
3373 oflags = O_RDWR | O_CREAT;
3375 env->me_fd = open(dpath, oflags, mode);
3377 if (env->me_fd == INVALID_HANDLE_VALUE) {
3382 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3383 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3384 env->me_mfd = env->me_fd;
3386 /* Synchronous fd for meta writes. Needed even with
3387 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3390 env->me_mfd = CreateFile(dpath, oflags,
3391 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3392 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3394 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3396 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3401 DPRINTF("opened dbenv %p", (void *) env);
3402 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3405 env->me_numdbs = 2; /* this notes that me_txkey was set */
3407 /* Windows TLS callbacks need help finding their TLS info. */
3408 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3409 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3416 rc = mdb_env_share_locks(env, &excl);
3420 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3421 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3422 env->me_path = strdup(path);
3423 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3429 mdb_env_close0(env, excl);
3435 /** Destroy resources from mdb_env_open() and clear our readers */
3437 mdb_env_close0(MDB_env *env, int excl)
3441 if (!(env->me_flags & MDB_ENV_ACTIVE))
3444 free(env->me_dbflags);
3448 if (env->me_numdbs) {
3449 pthread_key_delete(env->me_txkey);
3451 /* Delete our key from the global list */
3452 for (i=0; i<mdb_tls_nkeys; i++)
3453 if (mdb_tls_keys[i] == env->me_txkey) {
3454 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3462 munmap(env->me_map, env->me_mapsize);
3464 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3466 if (env->me_fd != INVALID_HANDLE_VALUE)
3469 pid_t pid = env->me_pid;
3470 /* Clearing readers is done in this function because
3471 * me_txkey with its destructor must be disabled first.
3473 for (i = env->me_numreaders; --i >= 0; )
3474 if (env->me_txns->mti_readers[i].mr_pid == pid)
3475 env->me_txns->mti_readers[i].mr_pid = 0;
3477 if (env->me_rmutex) {
3478 CloseHandle(env->me_rmutex);
3479 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3481 /* Windows automatically destroys the mutexes when
3482 * the last handle closes.
3484 #elif defined(MDB_USE_POSIX_SEM)
3485 if (env->me_rmutex != SEM_FAILED) {
3486 sem_close(env->me_rmutex);
3487 if (env->me_wmutex != SEM_FAILED)
3488 sem_close(env->me_wmutex);
3489 /* If we have the filelock: If we are the
3490 * only remaining user, clean up semaphores.
3493 mdb_env_excl_lock(env, &excl);
3495 sem_unlink(env->me_txns->mti_rmname);
3496 sem_unlink(env->me_txns->mti_wmname);
3500 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3502 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3505 /* Unlock the lockfile. Windows would have unlocked it
3506 * after closing anyway, but not necessarily at once.
3508 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3514 env->me_flags &= ~MDB_ENV_ACTIVE;
3518 mdb_env_copy(MDB_env *env, const char *path)
3520 MDB_txn *txn = NULL;
3524 HANDLE newfd = INVALID_HANDLE_VALUE;
3526 if (env->me_flags & MDB_NOSUBDIR) {
3527 lpath = (char *)path;
3530 len += sizeof(DATANAME);
3531 lpath = malloc(len);
3534 sprintf(lpath, "%s" DATANAME, path);
3537 /* The destination path must exist, but the destination file must not.
3538 * We don't want the OS to cache the writes, since the source data is
3539 * already in the OS cache.
3542 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3543 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3545 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3551 if (!(env->me_flags & MDB_NOSUBDIR))
3553 if (newfd == INVALID_HANDLE_VALUE) {
3558 #ifdef F_NOCACHE /* __APPLE__ */
3559 rc = fcntl(newfd, F_NOCACHE, 1);
3566 /* Do the lock/unlock of the reader mutex before starting the
3567 * write txn. Otherwise other read txns could block writers.
3569 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3573 if (!(env->me_flags & MDB_ROFS)) {
3574 /* We must start the actual read txn after blocking writers */
3575 mdb_txn_reset0(txn);
3577 /* Temporarily block writers until we snapshot the meta pages */
3580 rc = mdb_txn_renew0(txn);
3582 UNLOCK_MUTEX_W(env);
3587 wsize = env->me_psize * 2;
3591 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3592 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3595 rc = write(newfd, env->me_map, wsize);
3596 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3598 if (! (env->me_flags & MDB_ROFS))
3599 UNLOCK_MUTEX_W(env);
3604 ptr = env->me_map + wsize;
3605 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3606 #define MAX_WRITE 2147483648U
3610 if (wsize > MAX_WRITE)
3614 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3615 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3624 if (wsize > MAX_WRITE)
3628 wres = write(newfd, ptr, w2);
3629 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3638 if (newfd != INVALID_HANDLE_VALUE)
3645 mdb_env_close(MDB_env *env)
3652 VGMEMP_DESTROY(env);
3653 while ((dp = env->me_dpages) != NULL) {
3654 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3655 env->me_dpages = dp->mp_next;
3659 mdb_env_close0(env, 0);
3660 mdb_midl_free(env->me_free_pgs);
3664 /** Compare two items pointing at aligned size_t's */
3666 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3668 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3669 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3672 /** Compare two items pointing at aligned int's */
3674 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3676 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3677 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3680 /** Compare two items pointing at ints of unknown alignment.
3681 * Nodes and keys are guaranteed to be 2-byte aligned.
3684 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3686 #if BYTE_ORDER == LITTLE_ENDIAN
3687 unsigned short *u, *c;
3690 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3691 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3694 } while(!x && u > (unsigned short *)a->mv_data);
3697 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3701 /** Compare two items lexically */
3703 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3710 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3716 diff = memcmp(a->mv_data, b->mv_data, len);
3717 return diff ? diff : len_diff<0 ? -1 : len_diff;
3720 /** Compare two items in reverse byte order */
3722 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3724 const unsigned char *p1, *p2, *p1_lim;
3728 p1_lim = (const unsigned char *)a->mv_data;
3729 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3730 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3732 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3738 while (p1 > p1_lim) {
3739 diff = *--p1 - *--p2;
3743 return len_diff<0 ? -1 : len_diff;
3746 /** Search for key within a page, using binary search.
3747 * Returns the smallest entry larger or equal to the key.
3748 * If exactp is non-null, stores whether the found entry was an exact match
3749 * in *exactp (1 or 0).
3750 * Updates the cursor index with the index of the found entry.
3751 * If no entry larger or equal to the key is found, returns NULL.
3754 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3756 unsigned int i = 0, nkeys;
3759 MDB_page *mp = mc->mc_pg[mc->mc_top];
3760 MDB_node *node = NULL;
3765 nkeys = NUMKEYS(mp);
3770 COPY_PGNO(pgno, mp->mp_pgno);
3771 DPRINTF("searching %u keys in %s %spage %zu",
3772 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3779 low = IS_LEAF(mp) ? 0 : 1;
3781 cmp = mc->mc_dbx->md_cmp;
3783 /* Branch pages have no data, so if using integer keys,
3784 * alignment is guaranteed. Use faster mdb_cmp_int.
3786 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3787 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3794 nodekey.mv_size = mc->mc_db->md_pad;
3795 node = NODEPTR(mp, 0); /* fake */
3796 while (low <= high) {
3797 i = (low + high) >> 1;
3798 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3799 rc = cmp(key, &nodekey);
3800 DPRINTF("found leaf index %u [%s], rc = %i",
3801 i, DKEY(&nodekey), rc);
3810 while (low <= high) {
3811 i = (low + high) >> 1;
3813 node = NODEPTR(mp, i);
3814 nodekey.mv_size = NODEKSZ(node);
3815 nodekey.mv_data = NODEKEY(node);
3817 rc = cmp(key, &nodekey);
3820 DPRINTF("found leaf index %u [%s], rc = %i",
3821 i, DKEY(&nodekey), rc);
3823 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3824 i, DKEY(&nodekey), NODEPGNO(node), rc);
3835 if (rc > 0) { /* Found entry is less than the key. */
3836 i++; /* Skip to get the smallest entry larger than key. */
3838 node = NODEPTR(mp, i);
3841 *exactp = (rc == 0);
3842 /* store the key index */
3843 mc->mc_ki[mc->mc_top] = i;
3845 /* There is no entry larger or equal to the key. */
3848 /* nodeptr is fake for LEAF2 */
3854 mdb_cursor_adjust(MDB_cursor *mc, func)
3858 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3859 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3866 /** Pop a page off the top of the cursor's stack. */
3868 mdb_cursor_pop(MDB_cursor *mc)
3871 #ifndef MDB_DEBUG_SKIP
3872 MDB_page *top = mc->mc_pg[mc->mc_top];
3878 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3879 mc->mc_dbi, (void *) mc);
3883 /** Push a page onto the top of the cursor's stack. */
3885 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3887 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3888 mc->mc_dbi, (void *) mc);
3890 if (mc->mc_snum >= CURSOR_STACK) {
3891 assert(mc->mc_snum < CURSOR_STACK);
3892 return MDB_CURSOR_FULL;
3895 mc->mc_top = mc->mc_snum++;
3896 mc->mc_pg[mc->mc_top] = mp;
3897 mc->mc_ki[mc->mc_top] = 0;
3902 /** Find the address of the page corresponding to a given page number.
3903 * @param[in] txn the transaction for this access.
3904 * @param[in] pgno the page number for the page to retrieve.
3905 * @param[out] ret address of a pointer where the page's address will be stored.
3906 * @return 0 on success, non-zero on failure.
3909 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3913 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
3914 if (pgno < txn->mt_next_pgno)
3915 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3918 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3920 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3921 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3922 p = txn->mt_u.dirty_list[x].mptr;
3926 if (pgno < txn->mt_next_pgno)
3927 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3932 DPRINTF("page %zu not found", pgno);
3935 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3938 /** Search for the page a given key should be in.
3939 * Pushes parent pages on the cursor stack. This function continues a
3940 * search on a cursor that has already been initialized. (Usually by
3941 * #mdb_page_search() but also by #mdb_node_move().)
3942 * @param[in,out] mc the cursor for this operation.
3943 * @param[in] key the key to search for. If NULL, search for the lowest
3944 * page. (This is used by #mdb_cursor_first().)
3945 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3946 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3947 * @return 0 on success, non-zero on failure.
3950 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3952 MDB_page *mp = mc->mc_pg[mc->mc_top];
3957 while (IS_BRANCH(mp)) {
3961 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3962 assert(NUMKEYS(mp) > 1);
3963 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3965 if (key == NULL) /* Initialize cursor to first page. */
3967 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
3968 /* cursor to last page */
3972 node = mdb_node_search(mc, key, &exact);
3974 i = NUMKEYS(mp) - 1;
3976 i = mc->mc_ki[mc->mc_top];
3985 DPRINTF("following index %u for key [%s]",
3987 assert(i < NUMKEYS(mp));
3988 node = NODEPTR(mp, i);
3990 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3993 mc->mc_ki[mc->mc_top] = i;
3994 if ((rc = mdb_cursor_push(mc, mp)))
3998 if ((rc = mdb_page_touch(mc)) != 0)
4000 mp = mc->mc_pg[mc->mc_top];
4005 DPRINTF("internal error, index points to a %02X page!?",
4007 return MDB_CORRUPTED;
4010 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4011 key ? DKEY(key) : NULL);
4016 /** Search for the page a given key should be in.
4017 * Pushes parent pages on the cursor stack. This function just sets up
4018 * the search; it finds the root page for \b mc's database and sets this
4019 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4020 * called to complete the search.
4021 * @param[in,out] mc the cursor for this operation.
4022 * @param[in] key the key to search for. If NULL, search for the lowest
4023 * page. (This is used by #mdb_cursor_first().)
4024 * @param[in] modify If true, visited pages are updated with new page numbers.
4025 * @return 0 on success, non-zero on failure.
4028 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4033 /* Make sure the txn is still viable, then find the root from
4034 * the txn's db table.
4036 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4037 DPUTS("transaction has failed, must abort");
4040 /* Make sure we're using an up-to-date root */
4041 if (mc->mc_dbi > MAIN_DBI) {
4042 if ((*mc->mc_dbflag & DB_STALE) ||
4043 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4045 unsigned char dbflag = 0;
4046 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4047 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4050 if (*mc->mc_dbflag & DB_STALE) {
4053 MDB_node *leaf = mdb_node_search(&mc2,
4054 &mc->mc_dbx->md_name, &exact);
4056 return MDB_NOTFOUND;
4057 mdb_node_read(mc->mc_txn, leaf, &data);
4058 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4060 if (flags & MDB_PS_MODIFY)
4062 *mc->mc_dbflag = dbflag;
4065 root = mc->mc_db->md_root;
4067 if (root == P_INVALID) { /* Tree is empty. */
4068 DPUTS("tree is empty");
4069 return MDB_NOTFOUND;
4074 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4075 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
4081 DPRINTF("db %u root page %zu has flags 0x%X",
4082 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4084 if (flags & MDB_PS_MODIFY) {
4085 if ((rc = mdb_page_touch(mc)))
4089 if (flags & MDB_PS_ROOTONLY)
4092 return mdb_page_search_root(mc, key, flags);
4095 /** Return the data associated with a given node.
4096 * @param[in] txn The transaction for this operation.
4097 * @param[in] leaf The node being read.
4098 * @param[out] data Updated to point to the node's data.
4099 * @return 0 on success, non-zero on failure.
4102 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4104 MDB_page *omp; /* overflow page */
4108 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4109 data->mv_size = NODEDSZ(leaf);
4110 data->mv_data = NODEDATA(leaf);
4114 /* Read overflow data.
4116 data->mv_size = NODEDSZ(leaf);
4117 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4118 if ((rc = mdb_page_get(txn, pgno, &omp))) {
4119 DPRINTF("read overflow page %zu failed", pgno);
4122 data->mv_data = METADATA(omp);
4128 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4129 MDB_val *key, MDB_val *data)
4138 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4140 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4143 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4147 mdb_cursor_init(&mc, txn, dbi, &mx);
4148 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4151 /** Find a sibling for a page.
4152 * Replaces the page at the top of the cursor's stack with the
4153 * specified sibling, if one exists.
4154 * @param[in] mc The cursor for this operation.
4155 * @param[in] move_right Non-zero if the right sibling is requested,
4156 * otherwise the left sibling.
4157 * @return 0 on success, non-zero on failure.
4160 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4166 if (mc->mc_snum < 2) {
4167 return MDB_NOTFOUND; /* root has no siblings */
4171 DPRINTF("parent page is page %zu, index %u",
4172 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4174 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4175 : (mc->mc_ki[mc->mc_top] == 0)) {
4176 DPRINTF("no more keys left, moving to %s sibling",
4177 move_right ? "right" : "left");
4178 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4179 /* undo cursor_pop before returning */
4186 mc->mc_ki[mc->mc_top]++;
4188 mc->mc_ki[mc->mc_top]--;
4189 DPRINTF("just moving to %s index key %u",
4190 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4192 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4194 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4195 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4198 mdb_cursor_push(mc, mp);
4203 /** Move the cursor to the next data item. */
4205 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4211 if (mc->mc_flags & C_EOF) {
4212 return MDB_NOTFOUND;
4215 assert(mc->mc_flags & C_INITIALIZED);
4217 mp = mc->mc_pg[mc->mc_top];
4219 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4220 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4221 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4222 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4223 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4224 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4228 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4229 if (op == MDB_NEXT_DUP)
4230 return MDB_NOTFOUND;
4234 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4236 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4237 DPUTS("=====> move to next sibling page");
4238 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4239 mc->mc_flags |= C_EOF;
4240 mc->mc_flags &= ~C_INITIALIZED;
4241 return MDB_NOTFOUND;
4243 mp = mc->mc_pg[mc->mc_top];
4244 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4246 mc->mc_ki[mc->mc_top]++;
4248 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4249 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4252 key->mv_size = mc->mc_db->md_pad;
4253 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4257 assert(IS_LEAF(mp));
4258 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4260 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4261 mdb_xcursor_init1(mc, leaf);
4264 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4267 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4268 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4269 if (rc != MDB_SUCCESS)
4274 MDB_GET_KEY(leaf, key);
4278 /** Move the cursor to the previous data item. */
4280 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4286 assert(mc->mc_flags & C_INITIALIZED);
4288 mp = mc->mc_pg[mc->mc_top];
4290 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4291 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4292 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4293 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4294 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4295 if (op != MDB_PREV || rc == MDB_SUCCESS)
4298 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4299 if (op == MDB_PREV_DUP)
4300 return MDB_NOTFOUND;
4305 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4307 if (mc->mc_ki[mc->mc_top] == 0) {
4308 DPUTS("=====> move to prev sibling page");
4309 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4310 mc->mc_flags &= ~C_INITIALIZED;
4311 return MDB_NOTFOUND;
4313 mp = mc->mc_pg[mc->mc_top];
4314 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4315 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4317 mc->mc_ki[mc->mc_top]--;
4319 mc->mc_flags &= ~C_EOF;
4321 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4322 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4325 key->mv_size = mc->mc_db->md_pad;
4326 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4330 assert(IS_LEAF(mp));
4331 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4333 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4334 mdb_xcursor_init1(mc, leaf);
4337 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4340 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4341 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4342 if (rc != MDB_SUCCESS)
4347 MDB_GET_KEY(leaf, key);
4351 /** Set the cursor on a specific data item. */
4353 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4354 MDB_cursor_op op, int *exactp)
4358 MDB_node *leaf = NULL;
4363 assert(key->mv_size > 0);
4365 /* See if we're already on the right page */
4366 if (mc->mc_flags & C_INITIALIZED) {
4369 mp = mc->mc_pg[mc->mc_top];
4371 mc->mc_ki[mc->mc_top] = 0;
4372 return MDB_NOTFOUND;
4374 if (mp->mp_flags & P_LEAF2) {
4375 nodekey.mv_size = mc->mc_db->md_pad;
4376 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4378 leaf = NODEPTR(mp, 0);
4379 MDB_GET_KEY(leaf, &nodekey);
4381 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4383 /* Probably happens rarely, but first node on the page
4384 * was the one we wanted.
4386 mc->mc_ki[mc->mc_top] = 0;
4393 unsigned int nkeys = NUMKEYS(mp);
4395 if (mp->mp_flags & P_LEAF2) {
4396 nodekey.mv_data = LEAF2KEY(mp,
4397 nkeys-1, nodekey.mv_size);
4399 leaf = NODEPTR(mp, nkeys-1);
4400 MDB_GET_KEY(leaf, &nodekey);
4402 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4404 /* last node was the one we wanted */
4405 mc->mc_ki[mc->mc_top] = nkeys-1;
4411 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4412 /* This is definitely the right page, skip search_page */
4413 if (mp->mp_flags & P_LEAF2) {
4414 nodekey.mv_data = LEAF2KEY(mp,
4415 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4417 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4418 MDB_GET_KEY(leaf, &nodekey);
4420 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4422 /* current node was the one we wanted */
4432 /* If any parents have right-sibs, search.
4433 * Otherwise, there's nothing further.
4435 for (i=0; i<mc->mc_top; i++)
4437 NUMKEYS(mc->mc_pg[i])-1)
4439 if (i == mc->mc_top) {
4440 /* There are no other pages */
4441 mc->mc_ki[mc->mc_top] = nkeys;
4442 return MDB_NOTFOUND;
4446 /* There are no other pages */
4447 mc->mc_ki[mc->mc_top] = 0;
4448 return MDB_NOTFOUND;
4452 rc = mdb_page_search(mc, key, 0);
4453 if (rc != MDB_SUCCESS)
4456 mp = mc->mc_pg[mc->mc_top];
4457 assert(IS_LEAF(mp));
4460 leaf = mdb_node_search(mc, key, exactp);
4461 if (exactp != NULL && !*exactp) {
4462 /* MDB_SET specified and not an exact match. */
4463 return MDB_NOTFOUND;
4467 DPUTS("===> inexact leaf not found, goto sibling");
4468 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4469 return rc; /* no entries matched */
4470 mp = mc->mc_pg[mc->mc_top];
4471 assert(IS_LEAF(mp));
4472 leaf = NODEPTR(mp, 0);
4476 mc->mc_flags |= C_INITIALIZED;
4477 mc->mc_flags &= ~C_EOF;
4480 key->mv_size = mc->mc_db->md_pad;
4481 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4485 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4486 mdb_xcursor_init1(mc, leaf);
4489 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4490 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4491 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4494 if (op == MDB_GET_BOTH) {
4500 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4501 if (rc != MDB_SUCCESS)
4504 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4506 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4508 rc = mc->mc_dbx->md_dcmp(data, &d2);
4510 if (op == MDB_GET_BOTH || rc > 0)
4511 return MDB_NOTFOUND;
4516 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4517 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4522 /* The key already matches in all other cases */
4523 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4524 MDB_GET_KEY(leaf, key);
4525 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4530 /** Move the cursor to the first item in the database. */
4532 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4537 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4538 rc = mdb_page_search(mc, NULL, 0);
4539 if (rc != MDB_SUCCESS)
4542 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4544 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4545 mc->mc_flags |= C_INITIALIZED;
4546 mc->mc_flags &= ~C_EOF;
4548 mc->mc_ki[mc->mc_top] = 0;
4550 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4551 key->mv_size = mc->mc_db->md_pad;
4552 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4557 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4558 mdb_xcursor_init1(mc, leaf);
4559 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4564 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4565 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4569 MDB_GET_KEY(leaf, key);
4573 /** Move the cursor to the last item in the database. */
4575 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4580 if (!(mc->mc_flags & C_EOF)) {
4582 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4585 lkey.mv_size = MDB_MAXKEYSIZE+1;
4586 lkey.mv_data = NULL;
4587 rc = mdb_page_search(mc, &lkey, 0);
4588 if (rc != MDB_SUCCESS)
4591 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4593 mc->mc_flags |= C_INITIALIZED|C_EOF;
4594 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4596 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4598 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4599 key->mv_size = mc->mc_db->md_pad;
4600 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4605 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4606 mdb_xcursor_init1(mc, leaf);
4607 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4612 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4613 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4618 MDB_GET_KEY(leaf, key);
4623 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4632 case MDB_GET_CURRENT:
4633 if (!(mc->mc_flags & C_INITIALIZED)) {
4636 MDB_page *mp = mc->mc_pg[mc->mc_top];
4638 mc->mc_ki[mc->mc_top] = 0;
4644 key->mv_size = mc->mc_db->md_pad;
4645 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4647 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4648 MDB_GET_KEY(leaf, key);
4650 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4651 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4653 rc = mdb_node_read(mc->mc_txn, leaf, data);
4660 case MDB_GET_BOTH_RANGE:
4661 if (data == NULL || mc->mc_xcursor == NULL) {
4669 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4671 } else if (op == MDB_SET_RANGE)
4672 rc = mdb_cursor_set(mc, key, data, op, NULL);
4674 rc = mdb_cursor_set(mc, key, data, op, &exact);
4676 case MDB_GET_MULTIPLE:
4678 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4679 !(mc->mc_flags & C_INITIALIZED)) {
4684 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4685 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4688 case MDB_NEXT_MULTIPLE:
4690 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4694 if (!(mc->mc_flags & C_INITIALIZED))
4695 rc = mdb_cursor_first(mc, key, data);
4697 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4698 if (rc == MDB_SUCCESS) {
4699 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4702 mx = &mc->mc_xcursor->mx_cursor;
4703 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4705 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4706 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4714 case MDB_NEXT_NODUP:
4715 if (!(mc->mc_flags & C_INITIALIZED))
4716 rc = mdb_cursor_first(mc, key, data);
4718 rc = mdb_cursor_next(mc, key, data, op);
4722 case MDB_PREV_NODUP:
4723 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4724 rc = mdb_cursor_last(mc, key, data);
4725 mc->mc_flags |= C_INITIALIZED;
4726 mc->mc_ki[mc->mc_top]++;
4728 rc = mdb_cursor_prev(mc, key, data, op);
4731 rc = mdb_cursor_first(mc, key, data);
4735 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4736 !(mc->mc_flags & C_INITIALIZED) ||
4737 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4741 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4744 rc = mdb_cursor_last(mc, key, data);
4748 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4749 !(mc->mc_flags & C_INITIALIZED) ||
4750 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4754 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4757 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4765 /** Touch all the pages in the cursor stack.
4766 * Makes sure all the pages are writable, before attempting a write operation.
4767 * @param[in] mc The cursor to operate on.
4770 mdb_cursor_touch(MDB_cursor *mc)
4774 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4777 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4778 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4779 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4782 *mc->mc_dbflag = DB_DIRTY;
4784 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4785 rc = mdb_page_touch(mc);
4789 mc->mc_top = mc->mc_snum-1;
4794 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4797 MDB_node *leaf = NULL;
4798 MDB_val xdata, *rdata, dkey;
4801 int do_sub = 0, insert = 0;
4802 unsigned int mcount = 0;
4806 char dbuf[MDB_MAXKEYSIZE+1];
4807 unsigned int nflags;
4810 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4813 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4816 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4819 #if SIZE_MAX > MAXDATASIZE
4820 if (data->mv_size > MAXDATASIZE)
4824 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4825 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4829 if (flags == MDB_CURRENT) {
4830 if (!(mc->mc_flags & C_INITIALIZED))
4833 } else if (mc->mc_db->md_root == P_INVALID) {
4835 /* new database, write a root leaf page */
4836 DPUTS("allocating new root leaf page");
4837 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4841 mdb_cursor_push(mc, np);
4842 mc->mc_db->md_root = np->mp_pgno;
4843 mc->mc_db->md_depth++;
4844 *mc->mc_dbflag = DB_DIRTY;
4845 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4847 np->mp_flags |= P_LEAF2;
4848 mc->mc_flags |= C_INITIALIZED;
4854 if (flags & MDB_APPEND) {
4856 rc = mdb_cursor_last(mc, &k2, &d2);
4858 rc = mc->mc_dbx->md_cmp(key, &k2);
4861 mc->mc_ki[mc->mc_top]++;
4867 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4869 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4870 DPRINTF("duplicate key [%s]", DKEY(key));
4872 return MDB_KEYEXIST;
4874 if (rc && rc != MDB_NOTFOUND)
4878 /* Cursor is positioned, now make sure all pages are writable */
4879 rc2 = mdb_cursor_touch(mc);
4884 /* The key already exists */
4885 if (rc == MDB_SUCCESS) {
4886 /* there's only a key anyway, so this is a no-op */
4887 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4888 unsigned int ksize = mc->mc_db->md_pad;
4889 if (key->mv_size != ksize)
4891 if (flags == MDB_CURRENT) {
4892 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4893 memcpy(ptr, key->mv_data, ksize);
4898 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4901 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4902 /* Was a single item before, must convert now */
4904 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4905 /* Just overwrite the current item */
4906 if (flags == MDB_CURRENT)
4909 dkey.mv_size = NODEDSZ(leaf);
4910 dkey.mv_data = NODEDATA(leaf);
4911 #if UINT_MAX < SIZE_MAX
4912 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4913 #ifdef MISALIGNED_OK
4914 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4916 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4919 /* if data matches, ignore it */
4920 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4921 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4923 /* create a fake page for the dup items */
4924 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4925 dkey.mv_data = dbuf;
4926 fp = (MDB_page *)&pbuf;
4927 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4928 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4929 fp->mp_lower = PAGEHDRSZ;
4930 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4931 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4932 fp->mp_flags |= P_LEAF2;
4933 fp->mp_pad = data->mv_size;
4934 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4936 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4937 (dkey.mv_size & 1) + (data->mv_size & 1);
4939 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4942 xdata.mv_size = fp->mp_upper;
4947 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4948 /* See if we need to convert from fake page to subDB */
4950 unsigned int offset;
4953 fp = NODEDATA(leaf);
4954 if (flags == MDB_CURRENT) {
4956 fp->mp_flags |= P_DIRTY;
4957 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4958 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4962 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4963 offset = fp->mp_pad;
4964 if (SIZELEFT(fp) >= offset)
4966 offset *= 4; /* space for 4 more */
4968 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4970 offset += offset & 1;
4971 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4972 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4974 /* yes, convert it */
4976 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4977 dummy.md_pad = fp->mp_pad;
4978 dummy.md_flags = MDB_DUPFIXED;
4979 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4980 dummy.md_flags |= MDB_INTEGERKEY;
4983 dummy.md_branch_pages = 0;
4984 dummy.md_leaf_pages = 1;
4985 dummy.md_overflow_pages = 0;
4986 dummy.md_entries = NUMKEYS(fp);
4988 xdata.mv_size = sizeof(MDB_db);
4989 xdata.mv_data = &dummy;
4990 if ((rc = mdb_page_alloc(mc, 1, &mp)))
4992 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4993 flags |= F_DUPDATA|F_SUBDATA;
4994 dummy.md_root = mp->mp_pgno;
4996 /* no, just grow it */
4998 xdata.mv_size = NODEDSZ(leaf) + offset;
4999 xdata.mv_data = &pbuf;
5000 mp = (MDB_page *)&pbuf;
5001 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5004 mp->mp_flags = fp->mp_flags | P_DIRTY;
5005 mp->mp_pad = fp->mp_pad;
5006 mp->mp_lower = fp->mp_lower;
5007 mp->mp_upper = fp->mp_upper + offset;
5009 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5011 nsize = NODEDSZ(leaf) - fp->mp_upper;
5012 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5013 for (i=0; i<NUMKEYS(fp); i++)
5014 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5016 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5020 /* data is on sub-DB, just store it */
5021 flags |= F_DUPDATA|F_SUBDATA;
5025 /* overflow page overwrites need special handling */
5026 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5029 int ovpages, dpages;
5031 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5032 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5033 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5034 mdb_page_get(mc->mc_txn, pg, &omp);
5035 /* Is the ov page writable and large enough? */
5036 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5037 /* yes, overwrite it. Note in this case we don't
5038 * bother to try shrinking the node if the new data
5039 * is smaller than the overflow threshold.
5041 if (F_ISSET(flags, MDB_RESERVE))
5042 data->mv_data = METADATA(omp);
5044 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5047 /* no, free ovpages */
5049 mc->mc_db->md_overflow_pages -= ovpages;
5050 for (i=0; i<ovpages; i++) {
5051 DPRINTF("freed ov page %zu", pg);
5052 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5056 } else if (NODEDSZ(leaf) == data->mv_size) {
5057 /* same size, just replace it. Note that we could
5058 * also reuse this node if the new data is smaller,
5059 * but instead we opt to shrink the node in that case.
5061 if (F_ISSET(flags, MDB_RESERVE))
5062 data->mv_data = NODEDATA(leaf);
5063 else if (data->mv_size)
5064 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5066 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5069 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5070 mc->mc_db->md_entries--;
5072 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5079 nflags = flags & NODE_ADD_FLAGS;
5080 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5081 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5082 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5083 nflags &= ~MDB_APPEND;
5085 nflags |= MDB_SPLIT_REPLACE;
5086 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5088 /* There is room already in this leaf page. */
5089 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5090 if (rc == 0 && !do_sub && insert) {
5091 /* Adjust other cursors pointing to mp */
5092 MDB_cursor *m2, *m3;
5093 MDB_dbi dbi = mc->mc_dbi;
5094 unsigned i = mc->mc_top;
5095 MDB_page *mp = mc->mc_pg[i];
5097 if (mc->mc_flags & C_SUB)
5100 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5101 if (mc->mc_flags & C_SUB)
5102 m3 = &m2->mc_xcursor->mx_cursor;
5105 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5106 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5113 if (rc != MDB_SUCCESS)
5114 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5116 /* Now store the actual data in the child DB. Note that we're
5117 * storing the user data in the keys field, so there are strict
5118 * size limits on dupdata. The actual data fields of the child
5119 * DB are all zero size.
5126 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5127 if (flags & MDB_CURRENT) {
5128 xflags = MDB_CURRENT;
5130 mdb_xcursor_init1(mc, leaf);
5131 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5133 /* converted, write the original data first */
5135 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5139 /* Adjust other cursors pointing to mp */
5141 unsigned i = mc->mc_top;
5142 MDB_page *mp = mc->mc_pg[i];
5144 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5145 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5146 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5147 mdb_xcursor_init1(m2, leaf);
5152 if (flags & MDB_APPENDDUP)
5153 xflags |= MDB_APPEND;
5154 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5155 if (flags & F_SUBDATA) {
5156 void *db = NODEDATA(leaf);
5157 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5160 /* sub-writes might have failed so check rc again.
5161 * Don't increment count if we just replaced an existing item.
5163 if (!rc && !(flags & MDB_CURRENT))
5164 mc->mc_db->md_entries++;
5165 if (flags & MDB_MULTIPLE) {
5167 if (mcount < data[1].mv_size) {
5168 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5169 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5175 /* If we succeeded and the key didn't exist before, make sure
5176 * the cursor is marked valid.
5179 mc->mc_flags |= C_INITIALIZED;
5184 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5189 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5192 if (!(mc->mc_flags & C_INITIALIZED))
5195 rc = mdb_cursor_touch(mc);
5199 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5201 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5202 if (flags != MDB_NODUPDATA) {
5203 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5204 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5206 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5207 /* If sub-DB still has entries, we're done */
5208 if (mc->mc_xcursor->mx_db.md_entries) {
5209 if (leaf->mn_flags & F_SUBDATA) {
5210 /* update subDB info */
5211 void *db = NODEDATA(leaf);
5212 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5214 /* shrink fake page */
5215 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5217 mc->mc_db->md_entries--;
5220 /* otherwise fall thru and delete the sub-DB */
5223 if (leaf->mn_flags & F_SUBDATA) {
5224 /* add all the child DB's pages to the free list */
5225 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5226 if (rc == MDB_SUCCESS) {
5227 mc->mc_db->md_entries -=
5228 mc->mc_xcursor->mx_db.md_entries;
5233 return mdb_cursor_del0(mc, leaf);
5236 /** Allocate and initialize new pages for a database.
5237 * @param[in] mc a cursor on the database being added to.
5238 * @param[in] flags flags defining what type of page is being allocated.
5239 * @param[in] num the number of pages to allocate. This is usually 1,
5240 * unless allocating overflow pages for a large record.
5241 * @param[out] mp Address of a page, or NULL on failure.
5242 * @return 0 on success, non-zero on failure.
5245 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5250 if ((rc = mdb_page_alloc(mc, num, &np)))
5252 DPRINTF("allocated new mpage %zu, page size %u",
5253 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5254 np->mp_flags = flags | P_DIRTY;
5255 np->mp_lower = PAGEHDRSZ;
5256 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5259 mc->mc_db->md_branch_pages++;
5260 else if (IS_LEAF(np))
5261 mc->mc_db->md_leaf_pages++;
5262 else if (IS_OVERFLOW(np)) {
5263 mc->mc_db->md_overflow_pages += num;
5271 /** Calculate the size of a leaf node.
5272 * The size depends on the environment's page size; if a data item
5273 * is too large it will be put onto an overflow page and the node
5274 * size will only include the key and not the data. Sizes are always
5275 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5276 * of the #MDB_node headers.
5277 * @param[in] env The environment handle.
5278 * @param[in] key The key for the node.
5279 * @param[in] data The data for the node.
5280 * @return The number of bytes needed to store the node.
5283 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5287 sz = LEAFSIZE(key, data);
5288 if (sz >= env->me_psize / MDB_MINKEYS) {
5289 /* put on overflow page */
5290 sz -= data->mv_size - sizeof(pgno_t);
5294 return sz + sizeof(indx_t);
5297 /** Calculate the size of a branch node.
5298 * The size should depend on the environment's page size but since
5299 * we currently don't support spilling large keys onto overflow
5300 * pages, it's simply the size of the #MDB_node header plus the
5301 * size of the key. Sizes are always rounded up to an even number
5302 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5303 * @param[in] env The environment handle.
5304 * @param[in] key The key for the node.
5305 * @return The number of bytes needed to store the node.
5308 mdb_branch_size(MDB_env *env, MDB_val *key)
5313 if (sz >= env->me_psize / MDB_MINKEYS) {
5314 /* put on overflow page */
5315 /* not implemented */
5316 /* sz -= key->size - sizeof(pgno_t); */
5319 return sz + sizeof(indx_t);
5322 /** Add a node to the page pointed to by the cursor.
5323 * @param[in] mc The cursor for this operation.
5324 * @param[in] indx The index on the page where the new node should be added.
5325 * @param[in] key The key for the new node.
5326 * @param[in] data The data for the new node, if any.
5327 * @param[in] pgno The page number, if adding a branch node.
5328 * @param[in] flags Flags for the node.
5329 * @return 0 on success, non-zero on failure. Possible errors are:
5331 * <li>ENOMEM - failed to allocate overflow pages for the node.
5332 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5333 * should never happen since all callers already calculate the
5334 * page's free space before calling this function.
5338 mdb_node_add(MDB_cursor *mc, indx_t indx,
5339 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5342 size_t node_size = NODESIZE;
5345 MDB_page *mp = mc->mc_pg[mc->mc_top];
5346 MDB_page *ofp = NULL; /* overflow page */
5349 assert(mp->mp_upper >= mp->mp_lower);
5351 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5352 IS_LEAF(mp) ? "leaf" : "branch",
5353 IS_SUBP(mp) ? "sub-" : "",
5354 mp->mp_pgno, indx, data ? data->mv_size : 0,
5355 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5358 /* Move higher keys up one slot. */
5359 int ksize = mc->mc_db->md_pad, dif;
5360 char *ptr = LEAF2KEY(mp, indx, ksize);
5361 dif = NUMKEYS(mp) - indx;
5363 memmove(ptr+ksize, ptr, dif*ksize);
5364 /* insert new key */
5365 memcpy(ptr, key->mv_data, ksize);
5367 /* Just using these for counting */
5368 mp->mp_lower += sizeof(indx_t);
5369 mp->mp_upper -= ksize - sizeof(indx_t);
5374 node_size += key->mv_size;
5378 if (F_ISSET(flags, F_BIGDATA)) {
5379 /* Data already on overflow page. */
5380 node_size += sizeof(pgno_t);
5381 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5382 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5384 /* Put data on overflow page. */
5385 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5386 data->mv_size, node_size+data->mv_size);
5387 node_size += sizeof(pgno_t);
5388 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5390 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5393 node_size += data->mv_size;
5396 node_size += node_size & 1;
5398 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5399 DPRINTF("not enough room in page %zu, got %u ptrs",
5400 mp->mp_pgno, NUMKEYS(mp));
5401 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5402 mp->mp_upper - mp->mp_lower);
5403 DPRINTF("node size = %zu", node_size);
5404 return MDB_PAGE_FULL;
5407 /* Move higher pointers up one slot. */
5408 for (i = NUMKEYS(mp); i > indx; i--)
5409 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5411 /* Adjust free space offsets. */
5412 ofs = mp->mp_upper - node_size;
5413 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5414 mp->mp_ptrs[indx] = ofs;
5416 mp->mp_lower += sizeof(indx_t);
5418 /* Write the node data. */
5419 node = NODEPTR(mp, indx);
5420 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5421 node->mn_flags = flags;
5423 SETDSZ(node,data->mv_size);
5428 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5433 if (F_ISSET(flags, F_BIGDATA))
5434 memcpy(node->mn_data + key->mv_size, data->mv_data,
5436 else if (F_ISSET(flags, MDB_RESERVE))
5437 data->mv_data = node->mn_data + key->mv_size;
5439 memcpy(node->mn_data + key->mv_size, data->mv_data,
5442 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5444 if (F_ISSET(flags, MDB_RESERVE))
5445 data->mv_data = METADATA(ofp);
5447 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5454 /** Delete the specified node from a page.
5455 * @param[in] mp The page to operate on.
5456 * @param[in] indx The index of the node to delete.
5457 * @param[in] ksize The size of a node. Only used if the page is
5458 * part of a #MDB_DUPFIXED database.
5461 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5464 indx_t i, j, numkeys, ptr;
5471 COPY_PGNO(pgno, mp->mp_pgno);
5472 DPRINTF("delete node %u on %s page %zu", indx,
5473 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5476 assert(indx < NUMKEYS(mp));
5479 int x = NUMKEYS(mp) - 1 - indx;
5480 base = LEAF2KEY(mp, indx, ksize);
5482 memmove(base, base + ksize, x * ksize);
5483 mp->mp_lower -= sizeof(indx_t);
5484 mp->mp_upper += ksize - sizeof(indx_t);
5488 node = NODEPTR(mp, indx);
5489 sz = NODESIZE + node->mn_ksize;
5491 if (F_ISSET(node->mn_flags, F_BIGDATA))
5492 sz += sizeof(pgno_t);
5494 sz += NODEDSZ(node);
5498 ptr = mp->mp_ptrs[indx];
5499 numkeys = NUMKEYS(mp);
5500 for (i = j = 0; i < numkeys; i++) {
5502 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5503 if (mp->mp_ptrs[i] < ptr)
5504 mp->mp_ptrs[j] += sz;
5509 base = (char *)mp + mp->mp_upper;
5510 memmove(base + sz, base, ptr - mp->mp_upper);
5512 mp->mp_lower -= sizeof(indx_t);
5516 /** Compact the main page after deleting a node on a subpage.
5517 * @param[in] mp The main page to operate on.
5518 * @param[in] indx The index of the subpage on the main page.
5521 mdb_node_shrink(MDB_page *mp, indx_t indx)
5528 indx_t i, numkeys, ptr;
5530 node = NODEPTR(mp, indx);
5531 sp = (MDB_page *)NODEDATA(node);
5532 osize = NODEDSZ(node);
5534 delta = sp->mp_upper - sp->mp_lower;
5535 SETDSZ(node, osize - delta);
5536 xp = (MDB_page *)((char *)sp + delta);
5538 /* shift subpage upward */
5540 nsize = NUMKEYS(sp) * sp->mp_pad;
5541 memmove(METADATA(xp), METADATA(sp), nsize);
5544 nsize = osize - sp->mp_upper;
5545 numkeys = NUMKEYS(sp);
5546 for (i=numkeys-1; i>=0; i--)
5547 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5549 xp->mp_upper = sp->mp_lower;
5550 xp->mp_lower = sp->mp_lower;
5551 xp->mp_flags = sp->mp_flags;
5552 xp->mp_pad = sp->mp_pad;
5553 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5555 /* shift lower nodes upward */
5556 ptr = mp->mp_ptrs[indx];
5557 numkeys = NUMKEYS(mp);
5558 for (i = 0; i < numkeys; i++) {
5559 if (mp->mp_ptrs[i] <= ptr)
5560 mp->mp_ptrs[i] += delta;
5563 base = (char *)mp + mp->mp_upper;
5564 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5565 mp->mp_upper += delta;
5568 /** Initial setup of a sorted-dups cursor.
5569 * Sorted duplicates are implemented as a sub-database for the given key.
5570 * The duplicate data items are actually keys of the sub-database.
5571 * Operations on the duplicate data items are performed using a sub-cursor
5572 * initialized when the sub-database is first accessed. This function does
5573 * the preliminary setup of the sub-cursor, filling in the fields that
5574 * depend only on the parent DB.
5575 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5578 mdb_xcursor_init0(MDB_cursor *mc)
5580 MDB_xcursor *mx = mc->mc_xcursor;
5582 mx->mx_cursor.mc_xcursor = NULL;
5583 mx->mx_cursor.mc_txn = mc->mc_txn;
5584 mx->mx_cursor.mc_db = &mx->mx_db;
5585 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5586 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5587 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5588 mx->mx_cursor.mc_snum = 0;
5589 mx->mx_cursor.mc_top = 0;
5590 mx->mx_cursor.mc_flags = C_SUB;
5591 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5592 mx->mx_dbx.md_dcmp = NULL;
5593 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5596 /** Final setup of a sorted-dups cursor.
5597 * Sets up the fields that depend on the data from the main cursor.
5598 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5599 * @param[in] node The data containing the #MDB_db record for the
5600 * sorted-dup database.
5603 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5605 MDB_xcursor *mx = mc->mc_xcursor;
5607 if (node->mn_flags & F_SUBDATA) {
5608 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5609 mx->mx_cursor.mc_pg[0] = 0;
5610 mx->mx_cursor.mc_snum = 0;
5611 mx->mx_cursor.mc_flags = C_SUB;
5613 MDB_page *fp = NODEDATA(node);
5614 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5615 mx->mx_db.md_flags = 0;
5616 mx->mx_db.md_depth = 1;
5617 mx->mx_db.md_branch_pages = 0;
5618 mx->mx_db.md_leaf_pages = 1;
5619 mx->mx_db.md_overflow_pages = 0;
5620 mx->mx_db.md_entries = NUMKEYS(fp);
5621 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5622 mx->mx_cursor.mc_snum = 1;
5623 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5624 mx->mx_cursor.mc_top = 0;
5625 mx->mx_cursor.mc_pg[0] = fp;
5626 mx->mx_cursor.mc_ki[0] = 0;
5627 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5628 mx->mx_db.md_flags = MDB_DUPFIXED;
5629 mx->mx_db.md_pad = fp->mp_pad;
5630 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5631 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5634 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5636 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5638 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5639 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5640 #if UINT_MAX < SIZE_MAX
5641 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5642 #ifdef MISALIGNED_OK
5643 mx->mx_dbx.md_cmp = mdb_cmp_long;
5645 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5650 /** Initialize a cursor for a given transaction and database. */
5652 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5657 mc->mc_db = &txn->mt_dbs[dbi];
5658 mc->mc_dbx = &txn->mt_dbxs[dbi];
5659 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5664 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5666 mc->mc_xcursor = mx;
5667 mdb_xcursor_init0(mc);
5669 mc->mc_xcursor = NULL;
5671 if (*mc->mc_dbflag & DB_STALE) {
5672 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5677 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5680 MDB_xcursor *mx = NULL;
5681 size_t size = sizeof(MDB_cursor);
5683 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5686 /* Allow read access to the freelist */
5687 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5690 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5691 size += sizeof(MDB_xcursor);
5693 if ((mc = malloc(size)) != NULL) {
5694 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5695 mx = (MDB_xcursor *)(mc + 1);
5697 mdb_cursor_init(mc, txn, dbi, mx);
5698 if (txn->mt_cursors) {
5699 mc->mc_next = txn->mt_cursors[dbi];
5700 txn->mt_cursors[dbi] = mc;
5702 mc->mc_flags |= C_ALLOCD;
5713 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5715 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5718 if (txn->mt_cursors)
5721 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5725 /* Return the count of duplicate data items for the current key */
5727 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5731 if (mc == NULL || countp == NULL)
5734 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5737 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5738 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5741 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5744 *countp = mc->mc_xcursor->mx_db.md_entries;
5750 mdb_cursor_close(MDB_cursor *mc)
5753 /* remove from txn, if tracked */
5754 if (mc->mc_txn->mt_cursors) {
5755 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5756 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5758 *prev = mc->mc_next;
5760 if (mc->mc_flags & C_ALLOCD)
5766 mdb_cursor_txn(MDB_cursor *mc)
5768 if (!mc) return NULL;
5773 mdb_cursor_dbi(MDB_cursor *mc)
5779 /** Replace the key for a node with a new key.
5780 * @param[in] mp The page containing the node to operate on.
5781 * @param[in] indx The index of the node to operate on.
5782 * @param[in] key The new key to use.
5783 * @return 0 on success, non-zero on failure.
5786 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5793 indx_t ptr, i, numkeys, indx;
5796 indx = mc->mc_ki[mc->mc_top];
5797 mp = mc->mc_pg[mc->mc_top];
5798 node = NODEPTR(mp, indx);
5799 ptr = mp->mp_ptrs[indx];
5803 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5804 k2.mv_data = NODEKEY(node);
5805 k2.mv_size = node->mn_ksize;
5806 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5808 mdb_dkey(&k2, kbuf2),
5814 delta0 = delta = key->mv_size - node->mn_ksize;
5816 /* Must be 2-byte aligned. If new key is
5817 * shorter by 1, the shift will be skipped.
5819 delta += (delta & 1);
5821 if (delta > 0 && SIZELEFT(mp) < delta) {
5823 /* not enough space left, do a delete and split */
5824 DPRINTF("Not enough room, delta = %d, splitting...", delta);
5825 pgno = NODEPGNO(node);
5826 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5827 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
5830 numkeys = NUMKEYS(mp);
5831 for (i = 0; i < numkeys; i++) {
5832 if (mp->mp_ptrs[i] <= ptr)
5833 mp->mp_ptrs[i] -= delta;
5836 base = (char *)mp + mp->mp_upper;
5837 len = ptr - mp->mp_upper + NODESIZE;
5838 memmove(base - delta, base, len);
5839 mp->mp_upper -= delta;
5841 node = NODEPTR(mp, indx);
5844 /* But even if no shift was needed, update ksize */
5846 node->mn_ksize = key->mv_size;
5849 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5855 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
5857 /** Move a node from csrc to cdst.
5860 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5867 unsigned short flags;
5871 /* Mark src and dst as dirty. */
5872 if ((rc = mdb_page_touch(csrc)) ||
5873 (rc = mdb_page_touch(cdst)))
5876 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5877 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5878 key.mv_size = csrc->mc_db->md_pad;
5879 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5881 data.mv_data = NULL;
5885 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5886 assert(!((long)srcnode&1));
5887 srcpg = NODEPGNO(srcnode);
5888 flags = srcnode->mn_flags;
5889 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5890 unsigned int snum = csrc->mc_snum;
5892 /* must find the lowest key below src */
5893 mdb_page_search_root(csrc, NULL, 0);
5894 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5895 key.mv_size = csrc->mc_db->md_pad;
5896 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5898 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5899 key.mv_size = NODEKSZ(s2);
5900 key.mv_data = NODEKEY(s2);
5902 csrc->mc_snum = snum--;
5903 csrc->mc_top = snum;
5905 key.mv_size = NODEKSZ(srcnode);
5906 key.mv_data = NODEKEY(srcnode);
5908 data.mv_size = NODEDSZ(srcnode);
5909 data.mv_data = NODEDATA(srcnode);
5911 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5912 unsigned int snum = cdst->mc_snum;
5915 /* must find the lowest key below dst */
5916 mdb_page_search_root(cdst, NULL, 0);
5917 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5918 bkey.mv_size = cdst->mc_db->md_pad;
5919 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5921 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5922 bkey.mv_size = NODEKSZ(s2);
5923 bkey.mv_data = NODEKEY(s2);
5925 cdst->mc_snum = snum--;
5926 cdst->mc_top = snum;
5927 mdb_cursor_copy(cdst, &mn);
5929 rc = mdb_update_key(&mn, &bkey);
5934 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5935 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5936 csrc->mc_ki[csrc->mc_top],
5938 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5939 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5941 /* Add the node to the destination page.
5943 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5944 if (rc != MDB_SUCCESS)
5947 /* Delete the node from the source page.
5949 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5952 /* Adjust other cursors pointing to mp */
5953 MDB_cursor *m2, *m3;
5954 MDB_dbi dbi = csrc->mc_dbi;
5955 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5957 if (csrc->mc_flags & C_SUB)
5960 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5961 if (m2 == csrc) continue;
5962 if (csrc->mc_flags & C_SUB)
5963 m3 = &m2->mc_xcursor->mx_cursor;
5966 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5967 csrc->mc_ki[csrc->mc_top]) {
5968 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5969 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5974 /* Update the parent separators.
5976 if (csrc->mc_ki[csrc->mc_top] == 0) {
5977 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5978 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5979 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5981 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5982 key.mv_size = NODEKSZ(srcnode);
5983 key.mv_data = NODEKEY(srcnode);
5985 DPRINTF("update separator for source page %zu to [%s]",
5986 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5987 mdb_cursor_copy(csrc, &mn);
5990 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
5993 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5995 indx_t ix = csrc->mc_ki[csrc->mc_top];
5996 nullkey.mv_size = 0;
5997 csrc->mc_ki[csrc->mc_top] = 0;
5998 rc = mdb_update_key(csrc, &nullkey);
5999 csrc->mc_ki[csrc->mc_top] = ix;
6000 assert(rc == MDB_SUCCESS);
6004 if (cdst->mc_ki[cdst->mc_top] == 0) {
6005 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6006 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6007 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6009 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6010 key.mv_size = NODEKSZ(srcnode);
6011 key.mv_data = NODEKEY(srcnode);
6013 DPRINTF("update separator for destination page %zu to [%s]",
6014 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6015 mdb_cursor_copy(cdst, &mn);
6018 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6021 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6023 indx_t ix = cdst->mc_ki[cdst->mc_top];
6024 nullkey.mv_size = 0;
6025 cdst->mc_ki[cdst->mc_top] = 0;
6026 rc = mdb_update_key(cdst, &nullkey);
6027 cdst->mc_ki[cdst->mc_top] = ix;
6028 assert(rc == MDB_SUCCESS);
6035 /** Merge one page into another.
6036 * The nodes from the page pointed to by \b csrc will
6037 * be copied to the page pointed to by \b cdst and then
6038 * the \b csrc page will be freed.
6039 * @param[in] csrc Cursor pointing to the source page.
6040 * @param[in] cdst Cursor pointing to the destination page.
6043 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6051 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6052 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6054 assert(csrc->mc_snum > 1); /* can't merge root page */
6055 assert(cdst->mc_snum > 1);
6057 /* Mark dst as dirty. */
6058 if ((rc = mdb_page_touch(cdst)))
6061 /* Move all nodes from src to dst.
6063 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6064 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6065 key.mv_size = csrc->mc_db->md_pad;
6066 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6067 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6068 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6069 if (rc != MDB_SUCCESS)
6071 key.mv_data = (char *)key.mv_data + key.mv_size;
6074 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6075 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6076 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6077 unsigned int snum = csrc->mc_snum;
6079 /* must find the lowest key below src */
6080 mdb_page_search_root(csrc, NULL, 0);
6081 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6082 key.mv_size = csrc->mc_db->md_pad;
6083 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6085 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6086 key.mv_size = NODEKSZ(s2);
6087 key.mv_data = NODEKEY(s2);
6089 csrc->mc_snum = snum--;
6090 csrc->mc_top = snum;
6092 key.mv_size = srcnode->mn_ksize;
6093 key.mv_data = NODEKEY(srcnode);
6096 data.mv_size = NODEDSZ(srcnode);
6097 data.mv_data = NODEDATA(srcnode);
6098 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6099 if (rc != MDB_SUCCESS)
6104 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6105 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);
6107 /* Unlink the src page from parent and add to free list.
6109 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6110 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6113 rc = mdb_update_key(csrc, &key);
6119 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6120 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6121 csrc->mc_db->md_leaf_pages--;
6123 csrc->mc_db->md_branch_pages--;
6125 /* Adjust other cursors pointing to mp */
6126 MDB_cursor *m2, *m3;
6127 MDB_dbi dbi = csrc->mc_dbi;
6128 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6130 if (csrc->mc_flags & C_SUB)
6133 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6134 if (csrc->mc_flags & C_SUB)
6135 m3 = &m2->mc_xcursor->mx_cursor;
6138 if (m3 == csrc) continue;
6139 if (m3->mc_snum < csrc->mc_snum) continue;
6140 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6141 m3->mc_pg[csrc->mc_top] = mp;
6142 m3->mc_ki[csrc->mc_top] += nkeys;
6146 mdb_cursor_pop(csrc);
6148 return mdb_rebalance(csrc);
6151 /** Copy the contents of a cursor.
6152 * @param[in] csrc The cursor to copy from.
6153 * @param[out] cdst The cursor to copy to.
6156 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6160 cdst->mc_txn = csrc->mc_txn;
6161 cdst->mc_dbi = csrc->mc_dbi;
6162 cdst->mc_db = csrc->mc_db;
6163 cdst->mc_dbx = csrc->mc_dbx;
6164 cdst->mc_snum = csrc->mc_snum;
6165 cdst->mc_top = csrc->mc_top;
6166 cdst->mc_flags = csrc->mc_flags;
6168 for (i=0; i<csrc->mc_snum; i++) {
6169 cdst->mc_pg[i] = csrc->mc_pg[i];
6170 cdst->mc_ki[i] = csrc->mc_ki[i];
6174 /** Rebalance the tree after a delete operation.
6175 * @param[in] mc Cursor pointing to the page where rebalancing
6177 * @return 0 on success, non-zero on failure.
6180 mdb_rebalance(MDB_cursor *mc)
6190 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6191 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6192 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6193 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6197 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
6200 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6201 DPRINTF("no need to rebalance page %zu, above fill threshold",
6207 if (mc->mc_snum < 2) {
6208 MDB_page *mp = mc->mc_pg[0];
6209 if (NUMKEYS(mp) == 0) {
6210 DPUTS("tree is completely empty");
6211 mc->mc_db->md_root = P_INVALID;
6212 mc->mc_db->md_depth = 0;
6213 mc->mc_db->md_leaf_pages = 0;
6214 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6218 /* Adjust other cursors pointing to mp */
6219 MDB_cursor *m2, *m3;
6220 MDB_dbi dbi = mc->mc_dbi;
6222 if (mc->mc_flags & C_SUB)
6225 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6226 if (m2 == mc) continue;
6227 if (mc->mc_flags & C_SUB)
6228 m3 = &m2->mc_xcursor->mx_cursor;
6231 if (m3->mc_snum < mc->mc_snum) continue;
6232 if (m3->mc_pg[0] == mp) {
6238 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6239 DPUTS("collapsing root page!");
6240 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6241 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6242 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6245 mc->mc_db->md_depth--;
6246 mc->mc_db->md_branch_pages--;
6248 /* Adjust other cursors pointing to mp */
6249 MDB_cursor *m2, *m3;
6250 MDB_dbi dbi = mc->mc_dbi;
6252 if (mc->mc_flags & C_SUB)
6255 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6256 if (m2 == mc) continue;
6257 if (mc->mc_flags & C_SUB)
6258 m3 = &m2->mc_xcursor->mx_cursor;
6261 if (m3->mc_snum < mc->mc_snum) continue;
6262 if (m3->mc_pg[0] == mp) {
6263 m3->mc_pg[0] = mc->mc_pg[0];
6268 DPUTS("root page doesn't need rebalancing");
6272 /* The parent (branch page) must have at least 2 pointers,
6273 * otherwise the tree is invalid.
6275 ptop = mc->mc_top-1;
6276 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6278 /* Leaf page fill factor is below the threshold.
6279 * Try to move keys from left or right neighbor, or
6280 * merge with a neighbor page.
6285 mdb_cursor_copy(mc, &mn);
6286 mn.mc_xcursor = NULL;
6288 if (mc->mc_ki[ptop] == 0) {
6289 /* We're the leftmost leaf in our parent.
6291 DPUTS("reading right neighbor");
6293 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6294 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6296 mn.mc_ki[mn.mc_top] = 0;
6297 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6299 /* There is at least one neighbor to the left.
6301 DPUTS("reading left neighbor");
6303 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6304 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6306 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6307 mc->mc_ki[mc->mc_top] = 0;
6310 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6311 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);
6313 /* If the neighbor page is above threshold and has at least two
6314 * keys, move one key from it.
6316 * Otherwise we should try to merge them.
6318 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6319 return mdb_node_move(&mn, mc);
6321 if (mc->mc_ki[ptop] == 0)
6322 rc = mdb_page_merge(&mn, mc);
6324 rc = mdb_page_merge(mc, &mn);
6325 mc->mc_flags &= ~C_INITIALIZED;
6330 /** Complete a delete operation started by #mdb_cursor_del(). */
6332 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6336 /* add overflow pages to free list */
6337 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6341 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6342 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6343 mc->mc_db->md_overflow_pages -= ovpages;
6344 for (i=0; i<ovpages; i++) {
6345 DPRINTF("freed ov page %zu", pg);
6346 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6350 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6351 mc->mc_db->md_entries--;
6352 rc = mdb_rebalance(mc);
6353 if (rc != MDB_SUCCESS)
6354 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6360 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6361 MDB_val *key, MDB_val *data)
6366 MDB_val rdata, *xdata;
6370 assert(key != NULL);
6372 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6374 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6377 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6381 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6385 mdb_cursor_init(&mc, txn, dbi, &mx);
6396 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6398 /* let mdb_page_split know about this cursor if needed:
6399 * delete will trigger a rebalance; if it needs to move
6400 * a node from one page to another, it will have to
6401 * update the parent's separator key(s). If the new sepkey
6402 * is larger than the current one, the parent page may
6403 * run out of space, triggering a split. We need this
6404 * cursor to be consistent until the end of the rebalance.
6406 mc.mc_next = txn->mt_cursors[dbi];
6407 txn->mt_cursors[dbi] = &mc;
6408 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6409 txn->mt_cursors[dbi] = mc.mc_next;
6414 /** Split a page and insert a new node.
6415 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6416 * The cursor will be updated to point to the actual page and index where
6417 * the node got inserted after the split.
6418 * @param[in] newkey The key for the newly inserted node.
6419 * @param[in] newdata The data for the newly inserted node.
6420 * @param[in] newpgno The page number, if the new node is a branch node.
6421 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6422 * @return 0 on success, non-zero on failure.
6425 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6426 unsigned int nflags)
6429 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6432 unsigned int i, j, split_indx, nkeys, pmax;
6434 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6436 MDB_page *mp, *rp, *pp;
6441 mp = mc->mc_pg[mc->mc_top];
6442 newindx = mc->mc_ki[mc->mc_top];
6444 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6445 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6446 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6448 /* Create a right sibling. */
6449 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6451 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6453 if (mc->mc_snum < 2) {
6454 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6456 /* shift current top to make room for new parent */
6457 mc->mc_pg[1] = mc->mc_pg[0];
6458 mc->mc_ki[1] = mc->mc_ki[0];
6461 mc->mc_db->md_root = pp->mp_pgno;
6462 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6463 mc->mc_db->md_depth++;
6466 /* Add left (implicit) pointer. */
6467 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6468 /* undo the pre-push */
6469 mc->mc_pg[0] = mc->mc_pg[1];
6470 mc->mc_ki[0] = mc->mc_ki[1];
6471 mc->mc_db->md_root = mp->mp_pgno;
6472 mc->mc_db->md_depth--;
6479 ptop = mc->mc_top-1;
6480 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6483 mc->mc_flags |= C_SPLITTING;
6484 mdb_cursor_copy(mc, &mn);
6485 mn.mc_pg[mn.mc_top] = rp;
6486 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6488 if (nflags & MDB_APPEND) {
6489 mn.mc_ki[mn.mc_top] = 0;
6491 split_indx = newindx;
6496 nkeys = NUMKEYS(mp);
6497 split_indx = nkeys / 2;
6498 if (newindx < split_indx)
6504 unsigned int lsize, rsize, ksize;
6505 /* Move half of the keys to the right sibling */
6507 x = mc->mc_ki[mc->mc_top] - split_indx;
6508 ksize = mc->mc_db->md_pad;
6509 split = LEAF2KEY(mp, split_indx, ksize);
6510 rsize = (nkeys - split_indx) * ksize;
6511 lsize = (nkeys - split_indx) * sizeof(indx_t);
6512 mp->mp_lower -= lsize;
6513 rp->mp_lower += lsize;
6514 mp->mp_upper += rsize - lsize;
6515 rp->mp_upper -= rsize - lsize;
6516 sepkey.mv_size = ksize;
6517 if (newindx == split_indx) {
6518 sepkey.mv_data = newkey->mv_data;
6520 sepkey.mv_data = split;
6523 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6524 memcpy(rp->mp_ptrs, split, rsize);
6525 sepkey.mv_data = rp->mp_ptrs;
6526 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6527 memcpy(ins, newkey->mv_data, ksize);
6528 mp->mp_lower += sizeof(indx_t);
6529 mp->mp_upper -= ksize - sizeof(indx_t);
6532 memcpy(rp->mp_ptrs, split, x * ksize);
6533 ins = LEAF2KEY(rp, x, ksize);
6534 memcpy(ins, newkey->mv_data, ksize);
6535 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6536 rp->mp_lower += sizeof(indx_t);
6537 rp->mp_upper -= ksize - sizeof(indx_t);
6538 mc->mc_ki[mc->mc_top] = x;
6539 mc->mc_pg[mc->mc_top] = rp;
6544 /* For leaf pages, check the split point based on what
6545 * fits where, since otherwise mdb_node_add can fail.
6547 * This check is only needed when the data items are
6548 * relatively large, such that being off by one will
6549 * make the difference between success or failure.
6551 * It's also relevant if a page happens to be laid out
6552 * such that one half of its nodes are all "small" and
6553 * the other half of its nodes are "large." If the new
6554 * item is also "large" and falls on the half with
6555 * "large" nodes, it also may not fit.
6558 unsigned int psize, nsize;
6559 /* Maximum free space in an empty page */
6560 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6561 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6562 if ((nkeys < 20) || (nsize > pmax/16)) {
6563 if (newindx <= split_indx) {
6566 for (i=0; i<split_indx; i++) {
6567 node = NODEPTR(mp, i);
6568 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6569 if (F_ISSET(node->mn_flags, F_BIGDATA))
6570 psize += sizeof(pgno_t);
6572 psize += NODEDSZ(node);
6576 split_indx = newindx;
6587 for (i=nkeys-1; i>=split_indx; i--) {
6588 node = NODEPTR(mp, i);
6589 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6590 if (F_ISSET(node->mn_flags, F_BIGDATA))
6591 psize += sizeof(pgno_t);
6593 psize += NODEDSZ(node);
6597 split_indx = newindx;
6608 /* First find the separating key between the split pages.
6609 * The case where newindx == split_indx is ambiguous; the
6610 * new item could go to the new page or stay on the original
6611 * page. If newpos == 1 it goes to the new page.
6613 if (newindx == split_indx && newpos) {
6614 sepkey.mv_size = newkey->mv_size;
6615 sepkey.mv_data = newkey->mv_data;
6617 node = NODEPTR(mp, split_indx);
6618 sepkey.mv_size = node->mn_ksize;
6619 sepkey.mv_data = NODEKEY(node);
6623 DPRINTF("separator is [%s]", DKEY(&sepkey));
6625 /* Copy separator key to the parent.
6627 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6631 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6634 if (mn.mc_snum == mc->mc_snum) {
6635 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6636 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6637 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6638 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6643 /* Right page might now have changed parent.
6644 * Check if left page also changed parent.
6646 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6647 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6648 for (i=0; i<ptop; i++) {
6649 mc->mc_pg[i] = mn.mc_pg[i];
6650 mc->mc_ki[i] = mn.mc_ki[i];
6652 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6653 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6657 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6660 mc->mc_flags ^= C_SPLITTING;
6661 if (rc != MDB_SUCCESS) {
6664 if (nflags & MDB_APPEND) {
6665 mc->mc_pg[mc->mc_top] = rp;
6666 mc->mc_ki[mc->mc_top] = 0;
6667 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6670 for (i=0; i<mc->mc_top; i++)
6671 mc->mc_ki[i] = mn.mc_ki[i];
6678 /* Move half of the keys to the right sibling. */
6680 /* grab a page to hold a temporary copy */
6681 copy = mdb_page_malloc(mc);
6685 copy->mp_pgno = mp->mp_pgno;
6686 copy->mp_flags = mp->mp_flags;
6687 copy->mp_lower = PAGEHDRSZ;
6688 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6689 mc->mc_pg[mc->mc_top] = copy;
6690 for (i = j = 0; i <= nkeys; j++) {
6691 if (i == split_indx) {
6692 /* Insert in right sibling. */
6693 /* Reset insert index for right sibling. */
6694 if (i != newindx || (newpos ^ ins_new)) {
6696 mc->mc_pg[mc->mc_top] = rp;
6700 if (i == newindx && !ins_new) {
6701 /* Insert the original entry that caused the split. */
6702 rkey.mv_data = newkey->mv_data;
6703 rkey.mv_size = newkey->mv_size;
6712 /* Update index for the new key. */
6713 mc->mc_ki[mc->mc_top] = j;
6714 } else if (i == nkeys) {
6717 node = NODEPTR(mp, i);
6718 rkey.mv_data = NODEKEY(node);
6719 rkey.mv_size = node->mn_ksize;
6721 xdata.mv_data = NODEDATA(node);
6722 xdata.mv_size = NODEDSZ(node);
6725 pgno = NODEPGNO(node);
6726 flags = node->mn_flags;
6731 if (!IS_LEAF(mp) && j == 0) {
6732 /* First branch index doesn't need key data. */
6736 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6740 nkeys = NUMKEYS(copy);
6741 for (i=0; i<nkeys; i++)
6742 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6743 mp->mp_lower = copy->mp_lower;
6744 mp->mp_upper = copy->mp_upper;
6745 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6746 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6748 /* reset back to original page */
6749 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6750 mc->mc_pg[mc->mc_top] = mp;
6751 if (nflags & MDB_RESERVE) {
6752 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6753 if (!(node->mn_flags & F_BIGDATA))
6754 newdata->mv_data = NODEDATA(node);
6760 /* return tmp page to freelist */
6761 mdb_page_free(mc->mc_txn->mt_env, copy);
6764 /* Adjust other cursors pointing to mp */
6765 MDB_cursor *m2, *m3;
6766 MDB_dbi dbi = mc->mc_dbi;
6767 int fixup = NUMKEYS(mp);
6769 if (mc->mc_flags & C_SUB)
6772 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6773 if (m2 == mc) continue;
6774 if (mc->mc_flags & C_SUB)
6775 m3 = &m2->mc_xcursor->mx_cursor;
6778 if (!(m3->mc_flags & C_INITIALIZED))
6780 if (m3->mc_flags & C_SPLITTING)
6785 for (k=m3->mc_top; k>=0; k--) {
6786 m3->mc_ki[k+1] = m3->mc_ki[k];
6787 m3->mc_pg[k+1] = m3->mc_pg[k];
6789 if (m3->mc_ki[0] >= split_indx) {
6794 m3->mc_pg[0] = mc->mc_pg[0];
6798 if (m3->mc_pg[mc->mc_top] == mp) {
6799 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6800 m3->mc_ki[mc->mc_top]++;
6801 if (m3->mc_ki[mc->mc_top] >= fixup) {
6802 m3->mc_pg[mc->mc_top] = rp;
6803 m3->mc_ki[mc->mc_top] -= fixup;
6804 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6806 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6807 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6816 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6817 MDB_val *key, MDB_val *data, unsigned int flags)
6822 assert(key != NULL);
6823 assert(data != NULL);
6825 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6828 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6832 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6836 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6839 mdb_cursor_init(&mc, txn, dbi, &mx);
6840 return mdb_cursor_put(&mc, key, data, flags);
6844 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6846 if ((flag & CHANGEABLE) != flag)
6849 env->me_flags |= flag;
6851 env->me_flags &= ~flag;
6856 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6861 *arg = env->me_flags;
6866 mdb_env_get_path(MDB_env *env, const char **arg)
6871 *arg = env->me_path;
6875 /** Common code for #mdb_stat() and #mdb_env_stat().
6876 * @param[in] env the environment to operate in.
6877 * @param[in] db the #MDB_db record containing the stats to return.
6878 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6879 * @return 0, this function always succeeds.
6882 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6884 arg->ms_psize = env->me_psize;
6885 arg->ms_depth = db->md_depth;
6886 arg->ms_branch_pages = db->md_branch_pages;
6887 arg->ms_leaf_pages = db->md_leaf_pages;
6888 arg->ms_overflow_pages = db->md_overflow_pages;
6889 arg->ms_entries = db->md_entries;
6894 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6898 if (env == NULL || arg == NULL)
6901 toggle = mdb_env_pick_meta(env);
6903 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6907 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
6911 if (env == NULL || arg == NULL)
6914 toggle = mdb_env_pick_meta(env);
6915 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
6916 arg->me_mapsize = env->me_mapsize;
6917 arg->me_maxreaders = env->me_maxreaders;
6918 arg->me_numreaders = env->me_numreaders;
6919 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
6920 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
6924 /** Set the default comparison functions for a database.
6925 * Called immediately after a database is opened to set the defaults.
6926 * The user can then override them with #mdb_set_compare() or
6927 * #mdb_set_dupsort().
6928 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6929 * @param[in] dbi A database handle returned by #mdb_dbi_open()
6932 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6934 uint16_t f = txn->mt_dbs[dbi].md_flags;
6936 txn->mt_dbxs[dbi].md_cmp =
6937 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6938 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6940 txn->mt_dbxs[dbi].md_dcmp =
6941 !(f & MDB_DUPSORT) ? 0 :
6942 ((f & MDB_INTEGERDUP)
6943 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6944 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6947 #define PERSISTENT_FLAGS 0xffff
6948 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
6949 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
6950 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6955 int rc, dbflag, exact;
6956 unsigned int unused = 0;
6959 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6960 mdb_default_cmp(txn, FREE_DBI);
6963 if ((flags & VALID_FLAGS) != flags)
6969 if (flags & PERSISTENT_FLAGS) {
6970 uint16_t f2 = flags & PERSISTENT_FLAGS;
6971 /* make sure flag changes get committed */
6972 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
6973 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
6974 txn->mt_flags |= MDB_TXN_DIRTY;
6977 mdb_default_cmp(txn, MAIN_DBI);
6981 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6982 mdb_default_cmp(txn, MAIN_DBI);
6985 /* Is the DB already open? */
6987 for (i=2; i<txn->mt_numdbs; i++) {
6988 if (!txn->mt_dbxs[i].md_name.mv_size) {
6989 /* Remember this free slot */
6990 if (!unused) unused = i;
6993 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6994 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7000 /* If no free slot and max hit, fail */
7001 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7002 return MDB_DBS_FULL;
7004 /* Find the DB info */
7008 key.mv_data = (void *)name;
7009 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7010 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7011 if (rc == MDB_SUCCESS) {
7012 /* make sure this is actually a DB */
7013 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7014 if (!(node->mn_flags & F_SUBDATA))
7016 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7017 /* Create if requested */
7019 data.mv_size = sizeof(MDB_db);
7020 data.mv_data = &dummy;
7021 memset(&dummy, 0, sizeof(dummy));
7022 dummy.md_root = P_INVALID;
7023 dummy.md_flags = flags & PERSISTENT_FLAGS;
7024 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7028 /* OK, got info, add to table */
7029 if (rc == MDB_SUCCESS) {
7030 unsigned int slot = unused ? unused : txn->mt_numdbs;
7031 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7032 txn->mt_dbxs[slot].md_name.mv_size = len;
7033 txn->mt_dbxs[slot].md_rel = NULL;
7034 txn->mt_dbflags[slot] = dbflag;
7035 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7037 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7038 mdb_default_cmp(txn, slot);
7041 txn->mt_env->me_numdbs++;
7048 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7050 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7053 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7056 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7059 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
7061 ptr = env->me_dbxs[dbi].md_name.mv_data;
7062 env->me_dbxs[dbi].md_name.mv_data = NULL;
7063 env->me_dbxs[dbi].md_name.mv_size = 0;
7067 /** Add all the DB's pages to the free list.
7068 * @param[in] mc Cursor on the DB to free.
7069 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7070 * @return 0 on success, non-zero on failure.
7073 mdb_drop0(MDB_cursor *mc, int subs)
7077 rc = mdb_page_search(mc, NULL, 0);
7078 if (rc == MDB_SUCCESS) {
7083 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7084 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
7087 mdb_cursor_copy(mc, &mx);
7088 while (mc->mc_snum > 0) {
7089 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7090 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7091 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7092 if (ni->mn_flags & F_SUBDATA) {
7093 mdb_xcursor_init1(mc, ni);
7094 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7100 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7102 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7105 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7110 rc = mdb_cursor_sibling(mc, 1);
7112 /* no more siblings, go back to beginning
7113 * of previous level.
7116 for (i=1; i<mc->mc_top; i++)
7117 mc->mc_pg[i] = mx.mc_pg[i];
7121 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7122 mc->mc_db->md_root);
7127 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7132 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1)
7135 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7138 rc = mdb_cursor_open(txn, dbi, &mc);
7142 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7146 /* Can't delete the main DB */
7147 if (del && dbi > MAIN_DBI) {
7148 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7150 txn->mt_dbflags[dbi] = DB_STALE;
7151 mdb_dbi_close(txn->mt_env, dbi);
7154 /* reset the DB record, mark it dirty */
7155 txn->mt_dbflags[dbi] |= DB_DIRTY;
7156 txn->mt_dbs[dbi].md_depth = 0;
7157 txn->mt_dbs[dbi].md_branch_pages = 0;
7158 txn->mt_dbs[dbi].md_leaf_pages = 0;
7159 txn->mt_dbs[dbi].md_overflow_pages = 0;
7160 txn->mt_dbs[dbi].md_entries = 0;
7161 txn->mt_dbs[dbi].md_root = P_INVALID;
7163 if (!txn->mt_u.dirty_list[0].mid) {
7166 /* make sure we have at least one dirty page in this txn
7167 * otherwise these changes will be ignored.
7169 key.mv_size = sizeof(txnid_t);
7170 key.mv_data = &txn->mt_txnid;
7171 data.mv_size = sizeof(MDB_ID);
7172 data.mv_data = txn->mt_free_pgs;
7173 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
7174 rc = mdb_cursor_put(&m2, &key, &data, 0);
7178 mdb_cursor_close(mc);
7182 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7184 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7187 txn->mt_dbxs[dbi].md_cmp = cmp;
7191 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7193 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7196 txn->mt_dbxs[dbi].md_dcmp = cmp;
7200 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7202 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7205 txn->mt_dbxs[dbi].md_rel = rel;
7209 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7211 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7214 txn->mt_dbxs[dbi].md_relctx = ctx;