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);
1298 last = txn->mt_env->me_pglast + 1;
1300 key.mv_data = &last;
1301 key.mv_size = sizeof(last);
1302 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET_RANGE, NULL);
1305 last = *(txnid_t *)key.mv_data;
1311 oldest = txn->mt_txnid - 1;
1312 nr = txn->mt_env->me_txns->mti_numreaders;
1313 r = txn->mt_env->me_txns->mti_readers;
1314 for (i=0; i<nr; i++) {
1315 if (!r[i].mr_pid) continue;
1322 if (oldest > last) {
1323 /* It's usable, grab it.
1327 if (!txn->mt_env->me_pglast) {
1328 mdb_node_read(txn, leaf, &data);
1330 idl = (MDB_ID *) data.mv_data;
1331 /* We might have a zero-length IDL due to freelist growth
1332 * during a prior commit
1335 txn->mt_env->me_pglast = last;
1338 mop = malloc(MDB_IDL_SIZEOF(idl));
1341 txn->mt_env->me_pglast = last;
1342 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop;
1343 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1348 DPRINTF("IDL read txn %zu root %zu num %zu",
1349 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1350 for (i=0; i<idl[0]; i++) {
1351 DPRINTF("IDL %zu", idl[i+1]);
1358 if (txn->mt_env->me_pghead) {
1359 pgno_t *mop = txn->mt_env->me_pghead;
1362 int retry = 1, readit = 0, n2 = num-1;
1363 unsigned int i, j, k;
1365 /* If current list is too short, must fetch more and coalesce */
1366 if (mop[0] < (unsigned)num)
1369 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1371 /* If on freelist, don't try to read more. If what we have
1372 * right now isn't enough just use new pages.
1373 * TODO: get all of this working. Many circular dependencies...
1375 if (mc->mc_dbi == FREE_DBI) {
1383 last = txn->mt_env->me_pglast + 1;
1385 /* We haven't hit the readers list yet? */
1391 oldest = txn->mt_txnid - 1;
1392 nr = txn->mt_env->me_txns->mti_numreaders;
1393 r = txn->mt_env->me_txns->mti_readers;
1394 for (i=0; i<nr; i++) {
1395 if (!r[i].mr_pid) continue;
1402 /* There's nothing we can use on the freelist */
1403 if (oldest - last < 1)
1406 key.mv_data = &last;
1407 key.mv_size = sizeof(last);
1408 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1410 if (rc == MDB_NOTFOUND)
1414 last = *(txnid_t*)key.mv_data;
1417 idl = (MDB_ID *) data.mv_data;
1418 mop2 = malloc(MDB_IDL_SIZEOF(idl) + MDB_IDL_SIZEOF(mop));
1421 /* merge in sorted order */
1422 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1424 while (i>0 || j>0) {
1425 if (i && idl[i] < mop[j])
1426 mop2[k--] = idl[i--];
1428 mop2[k--] = mop[j--];
1430 txn->mt_env->me_pglast = last;
1431 free(txn->mt_env->me_pgfree);
1432 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop2;
1434 /* Keep trying to read until we have enough */
1435 if (mop[0] < (unsigned)num) {
1440 /* current list has enough pages, but are they contiguous? */
1441 for (i=mop[0]; i>=(unsigned)num; i--) {
1442 if (mop[i-n2] == mop[i] + n2) {
1445 /* move any stragglers down */
1446 for (j=i+num; j<=mop[0]; j++)
1453 /* Stop if we succeeded, or no retries */
1454 if (!retry || pgno != P_INVALID)
1460 /* peel pages off tail, so we only have to truncate the list */
1461 pgno = MDB_IDL_LAST(mop);
1464 if (MDB_IDL_IS_ZERO(mop)) {
1465 free(txn->mt_env->me_pgfree);
1466 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1471 if (pgno == P_INVALID) {
1472 /* DB size is maxed out */
1473 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1474 DPUTS("DB size maxed out");
1475 return MDB_MAP_FULL;
1478 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1479 if (pgno == P_INVALID) {
1480 pgno = txn->mt_next_pgno;
1481 txn->mt_next_pgno += num;
1483 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1486 if (txn->mt_env->me_dpages && num == 1) {
1487 np = txn->mt_env->me_dpages;
1488 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1489 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1490 txn->mt_env->me_dpages = np->mp_next;
1492 size_t sz = txn->mt_env->me_psize * num;
1493 if ((np = malloc(sz)) == NULL)
1495 VGMEMP_ALLOC(txn->mt_env, np, sz);
1497 if (pgno == P_INVALID) {
1498 np->mp_pgno = txn->mt_next_pgno;
1499 txn->mt_next_pgno += num;
1504 mid.mid = np->mp_pgno;
1506 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1507 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1509 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1516 /** Copy a page: avoid copying unused portions of the page.
1517 * @param[in] dst page to copy into
1518 * @param[in] src page to copy from
1521 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1523 dst->mp_flags = src->mp_flags | P_DIRTY;
1524 dst->mp_pages = src->mp_pages;
1526 if (IS_LEAF2(src)) {
1527 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1529 unsigned int i, nkeys = NUMKEYS(src);
1530 for (i=0; i<nkeys; i++)
1531 dst->mp_ptrs[i] = src->mp_ptrs[i];
1532 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1533 psize - src->mp_upper);
1537 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1538 * @param[in] mc cursor pointing to the page to be touched
1539 * @return 0 on success, non-zero on failure.
1542 mdb_page_touch(MDB_cursor *mc)
1544 MDB_page *mp = mc->mc_pg[mc->mc_top];
1548 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1550 if ((rc = mdb_page_alloc(mc, 1, &np)))
1552 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1553 assert(mp->mp_pgno != np->mp_pgno);
1554 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1556 /* If page isn't full, just copy the used portion */
1557 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1560 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1562 np->mp_flags |= P_DIRTY;
1567 /* Adjust other cursors pointing to mp */
1568 if (mc->mc_flags & C_SUB) {
1569 MDB_cursor *m2, *m3;
1570 MDB_dbi dbi = mc->mc_dbi-1;
1572 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1573 if (m2 == mc) continue;
1574 m3 = &m2->mc_xcursor->mx_cursor;
1575 if (m3->mc_snum < mc->mc_snum) continue;
1576 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1577 m3->mc_pg[mc->mc_top] = mp;
1583 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1584 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1585 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1586 m2->mc_pg[mc->mc_top] = mp;
1590 mc->mc_pg[mc->mc_top] = mp;
1591 /** If this page has a parent, update the parent to point to
1595 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1597 mc->mc_db->md_root = mp->mp_pgno;
1598 } else if (mc->mc_txn->mt_parent) {
1601 /* If txn has a parent, make sure the page is in our
1604 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1605 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1606 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1607 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1608 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1609 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1610 mc->mc_pg[mc->mc_top] = mp;
1615 if (mc->mc_txn->mt_u.dirty_list[0].mid >= MDB_IDL_UM_MAX)
1616 return MDB_TXN_FULL;
1618 np = mdb_page_malloc(mc);
1621 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1622 mid.mid = np->mp_pgno;
1624 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1632 mdb_env_sync(MDB_env *env, int force)
1635 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1636 if (env->me_flags & MDB_WRITEMAP) {
1637 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1638 ? MS_ASYNC : MS_SYNC;
1639 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1642 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1646 if (MDB_FDATASYNC(env->me_fd))
1653 /** Make shadow copies of all of parent txn's cursors */
1655 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1657 MDB_cursor *mc, *m2;
1658 unsigned int i, j, size;
1660 for (i=0;i<src->mt_numdbs; i++) {
1661 if (src->mt_cursors[i]) {
1662 size = sizeof(MDB_cursor);
1663 if (src->mt_cursors[i]->mc_xcursor)
1664 size += sizeof(MDB_xcursor);
1665 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1672 mc->mc_db = &dst->mt_dbs[i];
1673 mc->mc_dbx = m2->mc_dbx;
1674 mc->mc_dbflag = &dst->mt_dbflags[i];
1675 mc->mc_snum = m2->mc_snum;
1676 mc->mc_top = m2->mc_top;
1677 mc->mc_flags = m2->mc_flags | C_SHADOW;
1678 for (j=0; j<mc->mc_snum; j++) {
1679 mc->mc_pg[j] = m2->mc_pg[j];
1680 mc->mc_ki[j] = m2->mc_ki[j];
1682 if (m2->mc_xcursor) {
1683 MDB_xcursor *mx, *mx2;
1684 mx = (MDB_xcursor *)(mc+1);
1685 mc->mc_xcursor = mx;
1686 mx2 = m2->mc_xcursor;
1687 mx->mx_db = mx2->mx_db;
1688 mx->mx_dbx = mx2->mx_dbx;
1689 mx->mx_dbflag = mx2->mx_dbflag;
1690 mx->mx_cursor.mc_txn = dst;
1691 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1692 mx->mx_cursor.mc_db = &mx->mx_db;
1693 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1694 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1695 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1696 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1697 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1698 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1699 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1700 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1703 mc->mc_xcursor = NULL;
1705 mc->mc_next = dst->mt_cursors[i];
1706 dst->mt_cursors[i] = mc;
1713 /** Merge shadow cursors back into parent's */
1715 mdb_cursor_merge(MDB_txn *txn)
1718 for (i=0; i<txn->mt_numdbs; i++) {
1719 if (txn->mt_cursors[i]) {
1721 while ((mc = txn->mt_cursors[i])) {
1722 txn->mt_cursors[i] = mc->mc_next;
1723 if (mc->mc_flags & C_SHADOW) {
1724 MDB_cursor *m2 = mc->mc_orig;
1726 m2->mc_snum = mc->mc_snum;
1727 m2->mc_top = mc->mc_top;
1728 for (j=0; j<mc->mc_snum; j++) {
1729 m2->mc_pg[j] = mc->mc_pg[j];
1730 m2->mc_ki[j] = mc->mc_ki[j];
1733 if (mc->mc_flags & C_ALLOCD)
1741 mdb_txn_reset0(MDB_txn *txn);
1743 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1744 * @param[in] txn the transaction handle to initialize
1745 * @return 0 on success, non-zero on failure. This can only
1746 * fail for read-only transactions, and then only if the
1747 * reader table is full.
1750 mdb_txn_renew0(MDB_txn *txn)
1752 MDB_env *env = txn->mt_env;
1757 txn->mt_numdbs = env->me_numdbs;
1758 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1760 if (txn->mt_flags & MDB_TXN_RDONLY) {
1761 if (env->me_flags & MDB_ROFS) {
1762 i = mdb_env_pick_meta(env);
1763 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1764 txn->mt_u.reader = NULL;
1766 MDB_reader *r = pthread_getspecific(env->me_txkey);
1768 pid_t pid = env->me_pid;
1769 pthread_t tid = pthread_self();
1772 for (i=0; i<env->me_txns->mti_numreaders; i++)
1773 if (env->me_txns->mti_readers[i].mr_pid == 0)
1775 if (i == env->me_maxreaders) {
1776 UNLOCK_MUTEX_R(env);
1777 return MDB_READERS_FULL;
1779 env->me_txns->mti_readers[i].mr_pid = pid;
1780 env->me_txns->mti_readers[i].mr_tid = tid;
1781 if (i >= env->me_txns->mti_numreaders)
1782 env->me_txns->mti_numreaders = i+1;
1783 /* Save numreaders for un-mutexed mdb_env_close() */
1784 env->me_numreaders = env->me_txns->mti_numreaders;
1785 UNLOCK_MUTEX_R(env);
1786 r = &env->me_txns->mti_readers[i];
1787 if ((rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1788 env->me_txns->mti_readers[i].mr_pid = 0;
1792 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1793 txn->mt_u.reader = r;
1795 txn->mt_toggle = txn->mt_txnid & 1;
1796 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1800 txn->mt_txnid = env->me_txns->mti_txnid;
1801 txn->mt_toggle = txn->mt_txnid & 1;
1802 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1805 if (txn->mt_txnid == mdb_debug_start)
1808 txn->mt_u.dirty_list = env->me_dirty_list;
1809 txn->mt_u.dirty_list[0].mid = 0;
1810 txn->mt_free_pgs = env->me_free_pgs;
1811 txn->mt_free_pgs[0] = 0;
1815 /* Copy the DB info and flags */
1816 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1817 for (i=2; i<txn->mt_numdbs; i++)
1818 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1819 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1820 if (txn->mt_numdbs > 2)
1821 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1823 if (env->me_maxpg < txn->mt_next_pgno) {
1824 mdb_txn_reset0(txn);
1825 return MDB_MAP_RESIZED;
1832 mdb_txn_renew(MDB_txn *txn)
1836 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1839 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1840 DPUTS("environment had fatal error, must shutdown!");
1844 rc = mdb_txn_renew0(txn);
1845 if (rc == MDB_SUCCESS) {
1846 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1847 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1848 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1854 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1859 if (env->me_flags & MDB_FATAL_ERROR) {
1860 DPUTS("environment had fatal error, must shutdown!");
1863 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1866 /* Nested transactions: Max 1 child, write txns only, no writemap */
1867 if (parent->mt_child ||
1868 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1869 (env->me_flags & MDB_WRITEMAP))
1874 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1875 if (!(flags & MDB_RDONLY))
1876 size += env->me_maxdbs * sizeof(MDB_cursor *);
1878 if ((txn = calloc(1, size)) == NULL) {
1879 DPRINTF("calloc: %s", strerror(ErrCode()));
1882 txn->mt_dbs = (MDB_db *)(txn+1);
1883 if (flags & MDB_RDONLY) {
1884 txn->mt_flags |= MDB_TXN_RDONLY;
1885 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1887 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1888 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1893 txn->mt_free_pgs = mdb_midl_alloc();
1894 if (!txn->mt_free_pgs) {
1898 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1899 if (!txn->mt_u.dirty_list) {
1900 free(txn->mt_free_pgs);
1904 txn->mt_txnid = parent->mt_txnid;
1905 txn->mt_toggle = parent->mt_toggle;
1906 txn->mt_u.dirty_list[0].mid = 0;
1907 txn->mt_free_pgs[0] = 0;
1908 txn->mt_next_pgno = parent->mt_next_pgno;
1909 parent->mt_child = txn;
1910 txn->mt_parent = parent;
1911 txn->mt_numdbs = parent->mt_numdbs;
1912 txn->mt_dbxs = parent->mt_dbxs;
1913 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1914 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1915 mdb_cursor_shadow(parent, txn);
1918 rc = mdb_txn_renew0(txn);
1924 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1925 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1926 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1932 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1933 * @param[in] txn the transaction handle to reset
1936 mdb_txn_reset0(MDB_txn *txn)
1938 MDB_env *env = txn->mt_env;
1940 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1941 if (!(env->me_flags & MDB_ROFS))
1942 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1947 /* close(free) all cursors */
1948 for (i=0; i<txn->mt_numdbs; i++) {
1949 if (txn->mt_cursors[i]) {
1951 while ((mc = txn->mt_cursors[i])) {
1952 txn->mt_cursors[i] = mc->mc_next;
1953 if (mc->mc_flags & C_ALLOCD)
1959 if (!(env->me_flags & MDB_WRITEMAP)) {
1960 /* return all dirty pages to dpage list */
1961 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1962 dp = txn->mt_u.dirty_list[i].mptr;
1963 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1964 mdb_page_free(txn->mt_env, dp);
1966 /* large pages just get freed directly */
1967 VGMEMP_FREE(txn->mt_env, dp);
1973 if (txn->mt_parent) {
1974 txn->mt_parent->mt_child = NULL;
1975 mdb_midl_free(txn->mt_free_pgs);
1976 free(txn->mt_u.dirty_list);
1979 if (mdb_midl_shrink(&txn->mt_free_pgs))
1980 env->me_free_pgs = txn->mt_free_pgs;
1983 free(txn->mt_env->me_pgfree);
1984 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1985 txn->mt_env->me_pglast = 0;
1988 /* The writer mutex was locked in mdb_txn_begin. */
1989 UNLOCK_MUTEX_W(env);
1994 mdb_txn_reset(MDB_txn *txn)
1999 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2000 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2001 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2003 mdb_txn_reset0(txn);
2007 mdb_txn_abort(MDB_txn *txn)
2012 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2013 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2014 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2017 mdb_txn_abort(txn->mt_child);
2019 mdb_txn_reset0(txn);
2024 mdb_txn_commit(MDB_txn *txn)
2032 pgno_t next, freecnt;
2033 txnid_t oldpg_txnid, id;
2036 assert(txn != NULL);
2037 assert(txn->mt_env != NULL);
2039 if (txn->mt_child) {
2040 mdb_txn_commit(txn->mt_child);
2041 txn->mt_child = NULL;
2046 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2047 if (txn->mt_numdbs > env->me_numdbs) {
2048 /* update the DB flags */
2050 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2051 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2058 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2059 DPUTS("error flag is set, can't commit");
2061 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2066 if (txn->mt_parent) {
2067 MDB_txn *parent = txn->mt_parent;
2071 parent->mt_next_pgno = txn->mt_next_pgno;
2072 parent->mt_flags = txn->mt_flags;
2074 /* Merge (and close) our cursors with parent's */
2075 mdb_cursor_merge(txn);
2077 /* Update parent's DB table. */
2078 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2079 memcpy(parent->mt_dbflags, txn->mt_dbflags, txn->mt_numdbs);
2080 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2082 /* Append our free list to parent's */
2083 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
2085 mdb_midl_free(txn->mt_free_pgs);
2087 /* Merge our dirty list with parent's */
2088 dst = txn->mt_parent->mt_u.dirty_list;
2089 src = txn->mt_u.dirty_list;
2090 x = mdb_mid2l_search(dst, src[1].mid);
2091 for (y=1; y<=src[0].mid; y++) {
2092 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
2096 dst[x].mptr = src[y].mptr;
2099 for (; y<=src[0].mid; y++) {
2100 if (++x >= MDB_IDL_UM_MAX) {
2102 return MDB_TXN_FULL;
2107 free(txn->mt_u.dirty_list);
2108 txn->mt_parent->mt_child = NULL;
2113 if (txn != env->me_txn) {
2114 DPUTS("attempt to commit unknown transaction");
2119 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2122 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2123 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2125 /* Update DB root pointers */
2126 if (txn->mt_numdbs > 2) {
2129 data.mv_size = sizeof(MDB_db);
2131 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2132 for (i = 2; i < txn->mt_numdbs; i++) {
2133 if (txn->mt_dbflags[i] & DB_DIRTY) {
2134 data.mv_data = &txn->mt_dbs[i];
2135 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2142 /* Save the freelist as of this transaction to the freeDB. This
2143 * can change the freelist, so keep trying until it stabilizes.
2145 * env->me_pglast and the length of txn->mt_free_pgs cannot decrease,
2146 * except the code below can decrease env->me_pglast to split pghead.
2147 * Page numbers cannot disappear from txn->mt_free_pgs. New pages
2148 * can only appear in env->me_pghead when env->me_pglast increases.
2149 * Until then, the me_pghead pointer won't move but can become NULL.
2152 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2153 oldpg_txnid = id = 0;
2156 /* should only be one record now */
2157 if (env->me_pghead || env->me_pglast) {
2158 /* make sure first page of freeDB is touched and on freelist */
2159 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2160 if (rc && rc != MDB_NOTFOUND) {
2167 /* Delete IDLs we used from the free list */
2168 if (env->me_pglast) {
2173 rc = mdb_cursor_first(&mc, &key, NULL);
2176 oldpg_txnid = *(txnid_t *)key.mv_data;
2178 assert(oldpg_txnid <= env->me_pglast);
2180 rc = mdb_cursor_del(&mc, 0);
2183 } while (oldpg_txnid < env->me_pglast);
2186 /* Save IDL of pages freed by this txn, to freeDB */
2188 if (freecnt != txn->mt_free_pgs[0]) {
2191 /* make sure last page of freeDB is touched and on freelist */
2192 key.mv_size = MDB_MAXKEYSIZE+1;
2194 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2195 if (rc && rc != MDB_NOTFOUND)
2201 MDB_IDL idl = txn->mt_free_pgs;
2202 mdb_midl_sort(txn->mt_free_pgs);
2203 DPRINTF("IDL write txn %zu root %zu num %zu",
2204 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2205 for (i=1; i<=idl[0]; i++) {
2206 DPRINTF("IDL %zu", idl[i]);
2210 /* write to last page of freeDB */
2211 key.mv_size = sizeof(pgno_t);
2212 key.mv_data = &txn->mt_txnid;
2213 /* The free list can still grow during this call,
2214 * despite the pre-emptive touches above. So retry
2215 * until the reserved space remains big enough.
2218 assert(freecnt < txn->mt_free_pgs[0]);
2219 freecnt = txn->mt_free_pgs[0];
2220 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2221 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2224 } while (freecnt != txn->mt_free_pgs[0]);
2225 mdb_midl_sort(txn->mt_free_pgs);
2226 memcpy(data.mv_data, txn->mt_free_pgs, data.mv_size);
2227 if (oldpg_txnid < env->me_pglast || (!env->me_pghead && id))
2228 goto free_pgfirst; /* used up freeDB[oldpg_txnid] */
2231 /* Put back page numbers we took from freeDB but did not use */
2232 if (env->me_pghead) {
2237 mop = env->me_pghead;
2238 id = env->me_pglast;
2239 key.mv_size = sizeof(id);
2241 /* These steps may grow the freelist again
2242 * due to freed overflow pages...
2247 if (orig > env->me_maxfree_1pg && id > 4)
2248 orig = env->me_maxfree_1pg; /* Do not use more than 1 page */
2249 data.mv_size = (orig + 1) * sizeof(pgno_t);
2250 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2253 assert(!env->me_pghead || env->me_pglast);
2254 /* mop could have been used again here */
2255 if (id != env->me_pglast || env->me_pghead == NULL)
2256 goto again; /* was completely used up */
2257 assert(mop == env->me_pghead);
2258 } while (mop[0] < orig && --i);
2259 memcpy(data.mv_data, mop, data.mv_size);
2262 *(pgno_t *)data.mv_data = orig;
2263 mop[orig] = mop[0] - orig;
2264 env->me_pghead = mop += orig;
2265 /* Save more oldpages at the previous txnid. */
2266 assert(env->me_pglast == id && id == oldpg_txnid);
2267 env->me_pglast = --oldpg_txnid;
2271 /* Check for growth of freelist again */
2272 if (freecnt != txn->mt_free_pgs[0])
2275 free(env->me_pgfree);
2276 env->me_pghead = env->me_pgfree = NULL;
2278 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2279 if (mdb_midl_shrink(&txn->mt_free_pgs))
2280 env->me_free_pgs = txn->mt_free_pgs;
2287 if (env->me_flags & MDB_WRITEMAP) {
2288 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2289 dp = txn->mt_u.dirty_list[i].mptr;
2290 /* clear dirty flag */
2291 dp->mp_flags &= ~P_DIRTY;
2292 txn->mt_u.dirty_list[i].mid = 0;
2294 txn->mt_u.dirty_list[0].mid = 0;
2298 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2304 /* Windows actually supports scatter/gather I/O, but only on
2305 * unbuffered file handles. Since we're relying on the OS page
2306 * cache for all our data, that's self-defeating. So we just
2307 * write pages one at a time. We use the ov structure to set
2308 * the write offset, to at least save the overhead of a Seek
2312 memset(&ov, 0, sizeof(ov));
2313 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2315 dp = txn->mt_u.dirty_list[i].mptr;
2316 DPRINTF("committing page %zu", dp->mp_pgno);
2317 size = dp->mp_pgno * env->me_psize;
2318 ov.Offset = size & 0xffffffff;
2319 ov.OffsetHigh = size >> 16;
2320 ov.OffsetHigh >>= 16;
2321 /* clear dirty flag */
2322 dp->mp_flags &= ~P_DIRTY;
2323 wsize = env->me_psize;
2324 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2325 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2328 DPRINTF("WriteFile: %d", n);
2335 struct iovec iov[MDB_COMMIT_PAGES];
2339 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2340 dp = txn->mt_u.dirty_list[i].mptr;
2341 if (dp->mp_pgno != next) {
2343 rc = writev(env->me_fd, iov, n);
2347 DPUTS("short write, filesystem full?");
2349 DPRINTF("writev: %s", strerror(n));
2356 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2359 DPRINTF("committing page %zu", dp->mp_pgno);
2360 iov[n].iov_len = env->me_psize;
2361 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2362 iov[n].iov_base = (char *)dp;
2363 size += iov[n].iov_len;
2364 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2365 /* clear dirty flag */
2366 dp->mp_flags &= ~P_DIRTY;
2367 if (++n >= MDB_COMMIT_PAGES) {
2377 rc = writev(env->me_fd, iov, n);
2381 DPUTS("short write, filesystem full?");
2383 DPRINTF("writev: %s", strerror(n));
2390 /* Drop the dirty pages.
2392 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2393 dp = txn->mt_u.dirty_list[i].mptr;
2394 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2395 mdb_page_free(txn->mt_env, dp);
2397 VGMEMP_FREE(txn->mt_env, dp);
2400 txn->mt_u.dirty_list[i].mid = 0;
2402 txn->mt_u.dirty_list[0].mid = 0;
2405 if ((n = mdb_env_sync(env, 0)) != 0 ||
2406 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2414 if (txn->mt_numdbs > env->me_numdbs) {
2415 /* update the DB flags */
2417 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2418 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2422 UNLOCK_MUTEX_W(env);
2428 /** Read the environment parameters of a DB environment before
2429 * mapping it into memory.
2430 * @param[in] env the environment handle
2431 * @param[out] meta address of where to store the meta information
2432 * @return 0 on success, non-zero on failure.
2435 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2442 /* We don't know the page size yet, so use a minimum value.
2443 * Read both meta pages so we can use the latest one.
2446 for (i=0; i<2; i++) {
2448 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2450 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2455 else if (rc != MDB_PAGESIZE) {
2459 DPRINTF("read: %s", strerror(err));
2463 p = (MDB_page *)&pbuf;
2465 if (!F_ISSET(p->mp_flags, P_META)) {
2466 DPRINTF("page %zu not a meta page", p->mp_pgno);
2471 if (m->mm_magic != MDB_MAGIC) {
2472 DPUTS("meta has invalid magic");
2476 if (m->mm_version != MDB_VERSION) {
2477 DPRINTF("database is version %u, expected version %u",
2478 m->mm_version, MDB_VERSION);
2479 return MDB_VERSION_MISMATCH;
2483 if (m->mm_txnid > meta->mm_txnid)
2484 memcpy(meta, m, sizeof(*m));
2486 memcpy(meta, m, sizeof(*m));
2488 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2490 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2498 /** Write the environment parameters of a freshly created DB environment.
2499 * @param[in] env the environment handle
2500 * @param[out] meta address of where to store the meta information
2501 * @return 0 on success, non-zero on failure.
2504 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2511 DPUTS("writing new meta page");
2513 GET_PAGESIZE(psize);
2515 meta->mm_magic = MDB_MAGIC;
2516 meta->mm_version = MDB_VERSION;
2517 meta->mm_mapsize = env->me_mapsize;
2518 meta->mm_psize = psize;
2519 meta->mm_last_pg = 1;
2520 meta->mm_flags = env->me_flags & 0xffff;
2521 meta->mm_flags |= MDB_INTEGERKEY;
2522 meta->mm_dbs[0].md_root = P_INVALID;
2523 meta->mm_dbs[1].md_root = P_INVALID;
2525 p = calloc(2, psize);
2527 p->mp_flags = P_META;
2530 memcpy(m, meta, sizeof(*meta));
2532 q = (MDB_page *)((char *)p + psize);
2535 q->mp_flags = P_META;
2538 memcpy(m, meta, sizeof(*meta));
2543 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2544 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2545 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2548 lseek(env->me_fd, 0, SEEK_SET);
2549 rc = write(env->me_fd, p, psize * 2);
2550 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2556 /** Update the environment info to commit a transaction.
2557 * @param[in] txn the transaction that's being committed
2558 * @return 0 on success, non-zero on failure.
2561 mdb_env_write_meta(MDB_txn *txn)
2564 MDB_meta meta, metab, *mp;
2566 int rc, len, toggle;
2573 assert(txn != NULL);
2574 assert(txn->mt_env != NULL);
2576 toggle = !txn->mt_toggle;
2577 DPRINTF("writing meta page %d for root page %zu",
2578 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2581 mp = env->me_metas[toggle];
2583 if (env->me_flags & MDB_WRITEMAP) {
2584 /* Persist any increases of mapsize config */
2585 if (env->me_mapsize > mp->mm_mapsize)
2586 mp->mm_mapsize = env->me_mapsize;
2587 mp->mm_dbs[0] = txn->mt_dbs[0];
2588 mp->mm_dbs[1] = txn->mt_dbs[1];
2589 mp->mm_last_pg = txn->mt_next_pgno - 1;
2590 mp->mm_txnid = txn->mt_txnid;
2591 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2592 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2595 ptr += env->me_psize;
2596 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2603 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2604 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2606 ptr = (char *)&meta;
2607 if (env->me_mapsize > mp->mm_mapsize) {
2608 /* Persist any increases of mapsize config */
2609 meta.mm_mapsize = env->me_mapsize;
2610 off = offsetof(MDB_meta, mm_mapsize);
2612 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2614 len = sizeof(MDB_meta) - off;
2617 meta.mm_dbs[0] = txn->mt_dbs[0];
2618 meta.mm_dbs[1] = txn->mt_dbs[1];
2619 meta.mm_last_pg = txn->mt_next_pgno - 1;
2620 meta.mm_txnid = txn->mt_txnid;
2623 off += env->me_psize;
2626 /* Write to the SYNC fd */
2627 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2628 env->me_fd : env->me_mfd;
2631 memset(&ov, 0, sizeof(ov));
2633 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2636 rc = pwrite(mfd, ptr, len, off);
2641 DPUTS("write failed, disk error?");
2642 /* On a failure, the pagecache still contains the new data.
2643 * Write some old data back, to prevent it from being used.
2644 * Use the non-SYNC fd; we know it will fail anyway.
2646 meta.mm_last_pg = metab.mm_last_pg;
2647 meta.mm_txnid = metab.mm_txnid;
2649 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2651 r2 = pwrite(env->me_fd, ptr, len, off);
2654 env->me_flags |= MDB_FATAL_ERROR;
2658 /* Memory ordering issues are irrelevant; since the entire writer
2659 * is wrapped by wmutex, all of these changes will become visible
2660 * after the wmutex is unlocked. Since the DB is multi-version,
2661 * readers will get consistent data regardless of how fresh or
2662 * how stale their view of these values is.
2664 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2669 /** Check both meta pages to see which one is newer.
2670 * @param[in] env the environment handle
2671 * @return meta toggle (0 or 1).
2674 mdb_env_pick_meta(const MDB_env *env)
2676 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2680 mdb_env_create(MDB_env **env)
2684 e = calloc(1, sizeof(MDB_env));
2688 e->me_free_pgs = mdb_midl_alloc();
2689 if (!e->me_free_pgs) {
2693 e->me_maxreaders = DEFAULT_READERS;
2695 e->me_fd = INVALID_HANDLE_VALUE;
2696 e->me_lfd = INVALID_HANDLE_VALUE;
2697 e->me_mfd = INVALID_HANDLE_VALUE;
2698 #ifdef MDB_USE_POSIX_SEM
2699 e->me_rmutex = SEM_FAILED;
2700 e->me_wmutex = SEM_FAILED;
2702 e->me_pid = getpid();
2703 VGMEMP_CREATE(e,0,0);
2709 mdb_env_set_mapsize(MDB_env *env, size_t size)
2713 env->me_mapsize = size;
2715 env->me_maxpg = env->me_mapsize / env->me_psize;
2720 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2724 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2729 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2731 if (env->me_map || readers < 1)
2733 env->me_maxreaders = readers;
2738 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2740 if (!env || !readers)
2742 *readers = env->me_maxreaders;
2746 /** Further setup required for opening an MDB environment
2749 mdb_env_open2(MDB_env *env)
2751 unsigned int flags = env->me_flags;
2752 int i, newenv = 0, prot;
2756 memset(&meta, 0, sizeof(meta));
2758 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2761 DPUTS("new mdbenv");
2765 /* Was a mapsize configured? */
2766 if (!env->me_mapsize) {
2767 /* If this is a new environment, take the default,
2768 * else use the size recorded in the existing env.
2770 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2771 } else if (env->me_mapsize < meta.mm_mapsize) {
2772 /* If the configured size is smaller, make sure it's
2773 * still big enough. Silently round up to minimum if not.
2775 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2776 if (env->me_mapsize < minsize)
2777 env->me_mapsize = minsize;
2783 LONG sizelo, sizehi;
2784 sizelo = env->me_mapsize & 0xffffffff;
2785 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2787 /* Windows won't create mappings for zero length files.
2788 * Just allocate the maxsize right now.
2791 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2792 if (!SetEndOfFile(env->me_fd))
2794 SetFilePointer(env->me_fd, 0, NULL, 0);
2796 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2797 PAGE_READWRITE : PAGE_READONLY,
2798 sizehi, sizelo, NULL);
2801 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2802 FILE_MAP_WRITE : FILE_MAP_READ,
2803 0, 0, env->me_mapsize, meta.mm_address);
2811 if (flags & MDB_WRITEMAP) {
2813 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2816 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2818 if (env->me_map == MAP_FAILED) {
2825 if (flags & MDB_FIXEDMAP)
2826 meta.mm_address = env->me_map;
2827 i = mdb_env_init_meta(env, &meta);
2828 if (i != MDB_SUCCESS) {
2831 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2832 /* Can happen because the address argument to mmap() is just a
2833 * hint. mmap() can pick another, e.g. if the range is in use.
2834 * The MAP_FIXED flag would prevent that, but then mmap could
2835 * instead unmap existing pages to make room for the new map.
2837 return EBUSY; /* TODO: Make a new MDB_* error code? */
2839 env->me_psize = meta.mm_psize;
2840 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2842 env->me_maxpg = env->me_mapsize / env->me_psize;
2844 p = (MDB_page *)env->me_map;
2845 env->me_metas[0] = METADATA(p);
2846 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2850 int toggle = mdb_env_pick_meta(env);
2851 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2853 DPRINTF("opened database version %u, pagesize %u",
2854 env->me_metas[0]->mm_version, env->me_psize);
2855 DPRINTF("using meta page %d", toggle);
2856 DPRINTF("depth: %u", db->md_depth);
2857 DPRINTF("entries: %zu", db->md_entries);
2858 DPRINTF("branch pages: %zu", db->md_branch_pages);
2859 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2860 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2861 DPRINTF("root: %zu", db->md_root);
2869 /** Release a reader thread's slot in the reader lock table.
2870 * This function is called automatically when a thread exits.
2871 * @param[in] ptr This points to the slot in the reader lock table.
2874 mdb_env_reader_dest(void *ptr)
2876 MDB_reader *reader = ptr;
2882 /** Junk for arranging thread-specific callbacks on Windows. This is
2883 * necessarily platform and compiler-specific. Windows supports up
2884 * to 1088 keys. Let's assume nobody opens more than 64 environments
2885 * in a single process, for now. They can override this if needed.
2887 #ifndef MAX_TLS_KEYS
2888 #define MAX_TLS_KEYS 64
2890 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2891 static int mdb_tls_nkeys;
2893 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2897 case DLL_PROCESS_ATTACH: break;
2898 case DLL_THREAD_ATTACH: break;
2899 case DLL_THREAD_DETACH:
2900 for (i=0; i<mdb_tls_nkeys; i++) {
2901 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2902 mdb_env_reader_dest(r);
2905 case DLL_PROCESS_DETACH: break;
2910 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2912 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2916 /* Force some symbol references.
2917 * _tls_used forces the linker to create the TLS directory if not already done
2918 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2920 #pragma comment(linker, "/INCLUDE:_tls_used")
2921 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2922 #pragma const_seg(".CRT$XLB")
2923 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2924 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2927 #pragma comment(linker, "/INCLUDE:__tls_used")
2928 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2929 #pragma data_seg(".CRT$XLB")
2930 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2932 #endif /* WIN 32/64 */
2933 #endif /* !__GNUC__ */
2936 /** Downgrade the exclusive lock on the region back to shared */
2938 mdb_env_share_locks(MDB_env *env, int *excl)
2940 int rc = 0, toggle = mdb_env_pick_meta(env);
2942 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2947 /* First acquire a shared lock. The Unlock will
2948 * then release the existing exclusive lock.
2950 memset(&ov, 0, sizeof(ov));
2951 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2954 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2960 struct flock lock_info;
2961 /* The shared lock replaces the existing lock */
2962 memset((void *)&lock_info, 0, sizeof(lock_info));
2963 lock_info.l_type = F_RDLCK;
2964 lock_info.l_whence = SEEK_SET;
2965 lock_info.l_start = 0;
2966 lock_info.l_len = 1;
2967 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2968 (rc = ErrCode()) == EINTR) ;
2969 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
2976 /** Try to get exlusive lock, otherwise shared.
2977 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
2980 mdb_env_excl_lock(MDB_env *env, int *excl)
2984 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2988 memset(&ov, 0, sizeof(ov));
2989 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2996 struct flock lock_info;
2997 memset((void *)&lock_info, 0, sizeof(lock_info));
2998 lock_info.l_type = F_WRLCK;
2999 lock_info.l_whence = SEEK_SET;
3000 lock_info.l_start = 0;
3001 lock_info.l_len = 1;
3002 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3003 (rc = ErrCode()) == EINTR) ;
3007 # ifdef MDB_USE_POSIX_SEM
3008 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3011 lock_info.l_type = F_RDLCK;
3012 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3013 (rc = ErrCode()) == EINTR) ;
3021 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3023 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3025 * @(#) $Revision: 5.1 $
3026 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3027 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3029 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3033 * Please do not copyright this code. This code is in the public domain.
3035 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3036 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3037 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3038 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3039 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3040 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3041 * PERFORMANCE OF THIS SOFTWARE.
3044 * chongo <Landon Curt Noll> /\oo/\
3045 * http://www.isthe.com/chongo/
3047 * Share and Enjoy! :-)
3050 typedef unsigned long long mdb_hash_t;
3051 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3053 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3054 * @param[in] str string to hash
3055 * @param[in] hval initial value for hash
3056 * @return 64 bit hash
3058 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3059 * hval arg on the first call.
3062 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3064 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3065 unsigned char *end = s + val->mv_size;
3067 * FNV-1a hash each octet of the string
3070 /* xor the bottom with the current octet */
3071 hval ^= (mdb_hash_t)*s++;
3073 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3074 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3075 (hval << 7) + (hval << 8) + (hval << 40);
3077 /* return our new hash value */
3081 /** Hash the string and output the hash in hex.
3082 * @param[in] str string to hash
3083 * @param[out] hexbuf an array of 17 chars to hold the hash
3086 mdb_hash_hex(MDB_val *val, char *hexbuf)
3089 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3090 for (i=0; i<8; i++) {
3091 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3097 /** Open and/or initialize the lock region for the environment.
3098 * @param[in] env The MDB environment.
3099 * @param[in] lpath The pathname of the file used for the lock region.
3100 * @param[in] mode The Unix permissions for the file, if we create it.
3101 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3102 * @return 0 on success, non-zero on failure.
3105 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3113 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3114 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3115 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
3117 if (rc == ERROR_WRITE_PROTECT && (env->me_flags & MDB_RDONLY)) {
3118 env->me_flags |= MDB_ROFS;
3123 /* Try to get exclusive lock. If we succeed, then
3124 * nobody is using the lock region and we should initialize it.
3126 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3127 size = GetFileSize(env->me_lfd, NULL);
3133 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
3135 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3136 env->me_flags |= MDB_ROFS;
3141 /* Lose record locks when exec*() */
3142 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3143 fcntl(env->me_lfd, F_SETFD, fdflags);
3145 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
3146 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1) {
3148 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3149 env->me_flags |= MDB_ROFS;
3156 /* Try to get exclusive lock. If we succeed, then
3157 * nobody is using the lock region and we should initialize it.
3159 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3161 size = lseek(env->me_lfd, 0, SEEK_END);
3163 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3164 if (size < rsize && *excl > 0) {
3166 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3167 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3169 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3173 size = rsize - sizeof(MDB_txninfo);
3174 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3179 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3181 if (!mh) goto fail_errno;
3182 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3184 if (!env->me_txns) goto fail_errno;
3186 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3188 if (m == MAP_FAILED) goto fail_errno;
3194 BY_HANDLE_FILE_INFORMATION stbuf;
3203 if (!mdb_sec_inited) {
3204 InitializeSecurityDescriptor(&mdb_null_sd,
3205 SECURITY_DESCRIPTOR_REVISION);
3206 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3207 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3208 mdb_all_sa.bInheritHandle = FALSE;
3209 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3212 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3213 idbuf.volume = stbuf.dwVolumeSerialNumber;
3214 idbuf.nhigh = stbuf.nFileIndexHigh;
3215 idbuf.nlow = stbuf.nFileIndexLow;
3216 val.mv_data = &idbuf;
3217 val.mv_size = sizeof(idbuf);
3218 mdb_hash_hex(&val, hexbuf);
3219 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3220 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3221 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3222 if (!env->me_rmutex) goto fail_errno;
3223 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3224 if (!env->me_wmutex) goto fail_errno;
3225 #elif defined(MDB_USE_POSIX_SEM)
3234 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3235 idbuf.dev = stbuf.st_dev;
3236 idbuf.ino = stbuf.st_ino;
3237 val.mv_data = &idbuf;
3238 val.mv_size = sizeof(idbuf);
3239 mdb_hash_hex(&val, hexbuf);
3240 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3241 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3242 /* Clean up after a previous run, if needed: Try to
3243 * remove both semaphores before doing anything else.
3245 sem_unlink(env->me_txns->mti_rmname);
3246 sem_unlink(env->me_txns->mti_wmname);
3247 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3248 O_CREAT|O_EXCL, mode, 1);
3249 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3250 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3251 O_CREAT|O_EXCL, mode, 1);
3252 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3253 #else /* MDB_USE_POSIX_SEM */
3254 pthread_mutexattr_t mattr;
3256 if ((rc = pthread_mutexattr_init(&mattr))
3257 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3258 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3259 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3261 pthread_mutexattr_destroy(&mattr);
3262 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3264 env->me_txns->mti_version = MDB_VERSION;
3265 env->me_txns->mti_magic = MDB_MAGIC;
3266 env->me_txns->mti_txnid = 0;
3267 env->me_txns->mti_numreaders = 0;
3270 if (env->me_txns->mti_magic != MDB_MAGIC) {
3271 DPUTS("lock region has invalid magic");
3275 if (env->me_txns->mti_version != MDB_VERSION) {
3276 DPRINTF("lock region is version %u, expected version %u",
3277 env->me_txns->mti_version, MDB_VERSION);
3278 rc = MDB_VERSION_MISMATCH;
3282 if (rc != EACCES && rc != EAGAIN) {
3286 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3287 if (!env->me_rmutex) goto fail_errno;
3288 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3289 if (!env->me_wmutex) goto fail_errno;
3290 #elif defined(MDB_USE_POSIX_SEM)
3291 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3292 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3293 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3294 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3305 /** The name of the lock file in the DB environment */
3306 #define LOCKNAME "/lock.mdb"
3307 /** The name of the data file in the DB environment */
3308 #define DATANAME "/data.mdb"
3309 /** The suffix of the lock file when no subdir is used */
3310 #define LOCKSUFF "-lock"
3311 /** Only a subset of the @ref mdb_env flags can be changed
3312 * at runtime. Changing other flags requires closing the
3313 * environment and re-opening it with the new flags.
3315 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3316 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP)
3319 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3321 int oflags, rc, len, excl;
3322 char *lpath, *dpath;
3324 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3328 if (flags & MDB_NOSUBDIR) {
3329 rc = len + sizeof(LOCKSUFF) + len + 1;
3331 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3336 if (flags & MDB_NOSUBDIR) {
3337 dpath = lpath + len + sizeof(LOCKSUFF);
3338 sprintf(lpath, "%s" LOCKSUFF, path);
3339 strcpy(dpath, path);
3341 dpath = lpath + len + sizeof(LOCKNAME);
3342 sprintf(lpath, "%s" LOCKNAME, path);
3343 sprintf(dpath, "%s" DATANAME, path);
3346 flags |= env->me_flags;
3347 /* silently ignore WRITEMAP if we're only getting read access */
3348 if (F_ISSET(flags, MDB_RDONLY|MDB_WRITEMAP))
3349 flags ^= MDB_WRITEMAP;
3350 env->me_flags = flags |= MDB_ENV_ACTIVE;
3352 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3357 if (F_ISSET(flags, MDB_RDONLY)) {
3358 oflags = GENERIC_READ;
3359 len = OPEN_EXISTING;
3361 oflags = GENERIC_READ|GENERIC_WRITE;
3364 mode = FILE_ATTRIBUTE_NORMAL;
3365 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3366 NULL, len, mode, NULL);
3368 if (F_ISSET(flags, MDB_RDONLY))
3371 oflags = O_RDWR | O_CREAT;
3373 env->me_fd = open(dpath, oflags, mode);
3375 if (env->me_fd == INVALID_HANDLE_VALUE) {
3380 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3381 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3382 env->me_mfd = env->me_fd;
3384 /* Synchronous fd for meta writes. Needed even with
3385 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3388 env->me_mfd = CreateFile(dpath, oflags,
3389 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3390 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3392 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3394 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3399 DPRINTF("opened dbenv %p", (void *) env);
3400 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3403 env->me_numdbs = 2; /* this notes that me_txkey was set */
3405 /* Windows TLS callbacks need help finding their TLS info. */
3406 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3407 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3414 rc = mdb_env_share_locks(env, &excl);
3418 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3419 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3420 env->me_path = strdup(path);
3421 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3427 mdb_env_close0(env, excl);
3433 /** Destroy resources from mdb_env_open() and clear our readers */
3435 mdb_env_close0(MDB_env *env, int excl)
3439 if (!(env->me_flags & MDB_ENV_ACTIVE))
3442 free(env->me_dbflags);
3446 if (env->me_numdbs) {
3447 pthread_key_delete(env->me_txkey);
3449 /* Delete our key from the global list */
3450 for (i=0; i<mdb_tls_nkeys; i++)
3451 if (mdb_tls_keys[i] == env->me_txkey) {
3452 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3460 munmap(env->me_map, env->me_mapsize);
3462 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3464 if (env->me_fd != INVALID_HANDLE_VALUE)
3467 pid_t pid = env->me_pid;
3468 /* Clearing readers is done in this function because
3469 * me_txkey with its destructor must be disabled first.
3471 for (i = env->me_numreaders; --i >= 0; )
3472 if (env->me_txns->mti_readers[i].mr_pid == pid)
3473 env->me_txns->mti_readers[i].mr_pid = 0;
3475 if (env->me_rmutex) {
3476 CloseHandle(env->me_rmutex);
3477 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3479 /* Windows automatically destroys the mutexes when
3480 * the last handle closes.
3482 #elif defined(MDB_USE_POSIX_SEM)
3483 if (env->me_rmutex != SEM_FAILED) {
3484 sem_close(env->me_rmutex);
3485 if (env->me_wmutex != SEM_FAILED)
3486 sem_close(env->me_wmutex);
3487 /* If we have the filelock: If we are the
3488 * only remaining user, clean up semaphores.
3491 mdb_env_excl_lock(env, &excl);
3493 sem_unlink(env->me_txns->mti_rmname);
3494 sem_unlink(env->me_txns->mti_wmname);
3498 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3500 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3503 /* Unlock the lockfile. Windows would have unlocked it
3504 * after closing anyway, but not necessarily at once.
3506 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3512 env->me_flags &= ~MDB_ENV_ACTIVE;
3516 mdb_env_copy(MDB_env *env, const char *path)
3518 MDB_txn *txn = NULL;
3522 HANDLE newfd = INVALID_HANDLE_VALUE;
3524 if (env->me_flags & MDB_NOSUBDIR) {
3525 lpath = (char *)path;
3528 len += sizeof(DATANAME);
3529 lpath = malloc(len);
3532 sprintf(lpath, "%s" DATANAME, path);
3535 /* The destination path must exist, but the destination file must not.
3536 * We don't want the OS to cache the writes, since the source data is
3537 * already in the OS cache.
3540 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3541 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3543 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3549 if (!(env->me_flags & MDB_NOSUBDIR))
3551 if (newfd == INVALID_HANDLE_VALUE) {
3556 #ifdef F_NOCACHE /* __APPLE__ */
3557 rc = fcntl(newfd, F_NOCACHE, 1);
3564 /* Do the lock/unlock of the reader mutex before starting the
3565 * write txn. Otherwise other read txns could block writers.
3567 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3571 if (!(env->me_flags & MDB_ROFS)) {
3572 /* We must start the actual read txn after blocking writers */
3573 mdb_txn_reset0(txn);
3575 /* Temporarily block writers until we snapshot the meta pages */
3578 rc = mdb_txn_renew0(txn);
3580 UNLOCK_MUTEX_W(env);
3585 wsize = env->me_psize * 2;
3589 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3590 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3593 rc = write(newfd, env->me_map, wsize);
3594 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3596 if (! (env->me_flags & MDB_ROFS))
3597 UNLOCK_MUTEX_W(env);
3602 ptr = env->me_map + wsize;
3603 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3604 #define MAX_WRITE 2147483648U
3608 if (wsize > MAX_WRITE)
3612 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3613 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3622 if (wsize > MAX_WRITE)
3626 wres = write(newfd, ptr, w2);
3627 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3636 if (newfd != INVALID_HANDLE_VALUE)
3643 mdb_env_close(MDB_env *env)
3650 VGMEMP_DESTROY(env);
3651 while ((dp = env->me_dpages) != NULL) {
3652 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3653 env->me_dpages = dp->mp_next;
3657 mdb_env_close0(env, 0);
3658 mdb_midl_free(env->me_free_pgs);
3662 /** Compare two items pointing at aligned size_t's */
3664 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3666 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3667 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3670 /** Compare two items pointing at aligned int's */
3672 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3674 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3675 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3678 /** Compare two items pointing at ints of unknown alignment.
3679 * Nodes and keys are guaranteed to be 2-byte aligned.
3682 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3684 #if BYTE_ORDER == LITTLE_ENDIAN
3685 unsigned short *u, *c;
3688 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3689 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3692 } while(!x && u > (unsigned short *)a->mv_data);
3695 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3699 /** Compare two items lexically */
3701 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3708 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3714 diff = memcmp(a->mv_data, b->mv_data, len);
3715 return diff ? diff : len_diff<0 ? -1 : len_diff;
3718 /** Compare two items in reverse byte order */
3720 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3722 const unsigned char *p1, *p2, *p1_lim;
3726 p1_lim = (const unsigned char *)a->mv_data;
3727 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3728 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3730 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3736 while (p1 > p1_lim) {
3737 diff = *--p1 - *--p2;
3741 return len_diff<0 ? -1 : len_diff;
3744 /** Search for key within a page, using binary search.
3745 * Returns the smallest entry larger or equal to the key.
3746 * If exactp is non-null, stores whether the found entry was an exact match
3747 * in *exactp (1 or 0).
3748 * Updates the cursor index with the index of the found entry.
3749 * If no entry larger or equal to the key is found, returns NULL.
3752 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3754 unsigned int i = 0, nkeys;
3757 MDB_page *mp = mc->mc_pg[mc->mc_top];
3758 MDB_node *node = NULL;
3763 nkeys = NUMKEYS(mp);
3768 COPY_PGNO(pgno, mp->mp_pgno);
3769 DPRINTF("searching %u keys in %s %spage %zu",
3770 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3777 low = IS_LEAF(mp) ? 0 : 1;
3779 cmp = mc->mc_dbx->md_cmp;
3781 /* Branch pages have no data, so if using integer keys,
3782 * alignment is guaranteed. Use faster mdb_cmp_int.
3784 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3785 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3792 nodekey.mv_size = mc->mc_db->md_pad;
3793 node = NODEPTR(mp, 0); /* fake */
3794 while (low <= high) {
3795 i = (low + high) >> 1;
3796 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3797 rc = cmp(key, &nodekey);
3798 DPRINTF("found leaf index %u [%s], rc = %i",
3799 i, DKEY(&nodekey), rc);
3808 while (low <= high) {
3809 i = (low + high) >> 1;
3811 node = NODEPTR(mp, i);
3812 nodekey.mv_size = NODEKSZ(node);
3813 nodekey.mv_data = NODEKEY(node);
3815 rc = cmp(key, &nodekey);
3818 DPRINTF("found leaf index %u [%s], rc = %i",
3819 i, DKEY(&nodekey), rc);
3821 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3822 i, DKEY(&nodekey), NODEPGNO(node), rc);
3833 if (rc > 0) { /* Found entry is less than the key. */
3834 i++; /* Skip to get the smallest entry larger than key. */
3836 node = NODEPTR(mp, i);
3839 *exactp = (rc == 0);
3840 /* store the key index */
3841 mc->mc_ki[mc->mc_top] = i;
3843 /* There is no entry larger or equal to the key. */
3846 /* nodeptr is fake for LEAF2 */
3852 mdb_cursor_adjust(MDB_cursor *mc, func)
3856 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3857 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3864 /** Pop a page off the top of the cursor's stack. */
3866 mdb_cursor_pop(MDB_cursor *mc)
3869 #ifndef MDB_DEBUG_SKIP
3870 MDB_page *top = mc->mc_pg[mc->mc_top];
3876 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3877 mc->mc_dbi, (void *) mc);
3881 /** Push a page onto the top of the cursor's stack. */
3883 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3885 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3886 mc->mc_dbi, (void *) mc);
3888 if (mc->mc_snum >= CURSOR_STACK) {
3889 assert(mc->mc_snum < CURSOR_STACK);
3890 return MDB_CURSOR_FULL;
3893 mc->mc_top = mc->mc_snum++;
3894 mc->mc_pg[mc->mc_top] = mp;
3895 mc->mc_ki[mc->mc_top] = 0;
3900 /** Find the address of the page corresponding to a given page number.
3901 * @param[in] txn the transaction for this access.
3902 * @param[in] pgno the page number for the page to retrieve.
3903 * @param[out] ret address of a pointer where the page's address will be stored.
3904 * @return 0 on success, non-zero on failure.
3907 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3911 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
3912 (txn->mt_env->me_flags & MDB_WRITEMAP)))
3916 MDB_ID2L dl = tx2->mt_u.dirty_list;
3918 unsigned x = mdb_mid2l_search(dl, pgno);
3919 if (x <= dl[0].mid && dl[x].mid == pgno) {
3924 } while ((tx2 = tx2->mt_parent) != NULL);
3927 if (pgno < txn->mt_next_pgno) {
3928 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3930 DPRINTF("page %zu not found", pgno);
3936 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3939 /** Search for the page a given key should be in.
3940 * Pushes parent pages on the cursor stack. This function continues a
3941 * search on a cursor that has already been initialized. (Usually by
3942 * #mdb_page_search() but also by #mdb_node_move().)
3943 * @param[in,out] mc the cursor for this operation.
3944 * @param[in] key the key to search for. If NULL, search for the lowest
3945 * page. (This is used by #mdb_cursor_first().)
3946 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3947 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3948 * @return 0 on success, non-zero on failure.
3951 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3953 MDB_page *mp = mc->mc_pg[mc->mc_top];
3958 while (IS_BRANCH(mp)) {
3962 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3963 assert(NUMKEYS(mp) > 1);
3964 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3966 if (key == NULL) /* Initialize cursor to first page. */
3968 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
3969 /* cursor to last page */
3973 node = mdb_node_search(mc, key, &exact);
3975 i = NUMKEYS(mp) - 1;
3977 i = mc->mc_ki[mc->mc_top];
3986 DPRINTF("following index %u for key [%s]",
3988 assert(i < NUMKEYS(mp));
3989 node = NODEPTR(mp, i);
3991 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3994 mc->mc_ki[mc->mc_top] = i;
3995 if ((rc = mdb_cursor_push(mc, mp)))
3999 if ((rc = mdb_page_touch(mc)) != 0)
4001 mp = mc->mc_pg[mc->mc_top];
4006 DPRINTF("internal error, index points to a %02X page!?",
4008 return MDB_CORRUPTED;
4011 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4012 key ? DKEY(key) : NULL);
4017 /** Search for the page a given key should be in.
4018 * Pushes parent pages on the cursor stack. This function just sets up
4019 * the search; it finds the root page for \b mc's database and sets this
4020 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4021 * called to complete the search.
4022 * @param[in,out] mc the cursor for this operation.
4023 * @param[in] key the key to search for. If NULL, search for the lowest
4024 * page. (This is used by #mdb_cursor_first().)
4025 * @param[in] modify If true, visited pages are updated with new page numbers.
4026 * @return 0 on success, non-zero on failure.
4029 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4034 /* Make sure the txn is still viable, then find the root from
4035 * the txn's db table.
4037 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4038 DPUTS("transaction has failed, must abort");
4041 /* Make sure we're using an up-to-date root */
4042 if (mc->mc_dbi > MAIN_DBI) {
4043 if ((*mc->mc_dbflag & DB_STALE) ||
4044 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4046 unsigned char dbflag = 0;
4047 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4048 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4051 if (*mc->mc_dbflag & DB_STALE) {
4054 MDB_node *leaf = mdb_node_search(&mc2,
4055 &mc->mc_dbx->md_name, &exact);
4057 return MDB_NOTFOUND;
4058 mdb_node_read(mc->mc_txn, leaf, &data);
4059 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4061 if (flags & MDB_PS_MODIFY)
4063 *mc->mc_dbflag = dbflag;
4066 root = mc->mc_db->md_root;
4068 if (root == P_INVALID) { /* Tree is empty. */
4069 DPUTS("tree is empty");
4070 return MDB_NOTFOUND;
4075 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4076 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
4082 DPRINTF("db %u root page %zu has flags 0x%X",
4083 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4085 if (flags & MDB_PS_MODIFY) {
4086 if ((rc = mdb_page_touch(mc)))
4090 if (flags & MDB_PS_ROOTONLY)
4093 return mdb_page_search_root(mc, key, flags);
4096 /** Return the data associated with a given node.
4097 * @param[in] txn The transaction for this operation.
4098 * @param[in] leaf The node being read.
4099 * @param[out] data Updated to point to the node's data.
4100 * @return 0 on success, non-zero on failure.
4103 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4105 MDB_page *omp; /* overflow page */
4109 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4110 data->mv_size = NODEDSZ(leaf);
4111 data->mv_data = NODEDATA(leaf);
4115 /* Read overflow data.
4117 data->mv_size = NODEDSZ(leaf);
4118 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4119 if ((rc = mdb_page_get(txn, pgno, &omp))) {
4120 DPRINTF("read overflow page %zu failed", pgno);
4123 data->mv_data = METADATA(omp);
4129 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4130 MDB_val *key, MDB_val *data)
4139 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4141 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4144 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4148 mdb_cursor_init(&mc, txn, dbi, &mx);
4149 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4152 /** Find a sibling for a page.
4153 * Replaces the page at the top of the cursor's stack with the
4154 * specified sibling, if one exists.
4155 * @param[in] mc The cursor for this operation.
4156 * @param[in] move_right Non-zero if the right sibling is requested,
4157 * otherwise the left sibling.
4158 * @return 0 on success, non-zero on failure.
4161 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4167 if (mc->mc_snum < 2) {
4168 return MDB_NOTFOUND; /* root has no siblings */
4172 DPRINTF("parent page is page %zu, index %u",
4173 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4175 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4176 : (mc->mc_ki[mc->mc_top] == 0)) {
4177 DPRINTF("no more keys left, moving to %s sibling",
4178 move_right ? "right" : "left");
4179 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4180 /* undo cursor_pop before returning */
4187 mc->mc_ki[mc->mc_top]++;
4189 mc->mc_ki[mc->mc_top]--;
4190 DPRINTF("just moving to %s index key %u",
4191 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4193 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4195 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4196 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4199 mdb_cursor_push(mc, mp);
4204 /** Move the cursor to the next data item. */
4206 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4212 if (mc->mc_flags & C_EOF) {
4213 return MDB_NOTFOUND;
4216 assert(mc->mc_flags & C_INITIALIZED);
4218 mp = mc->mc_pg[mc->mc_top];
4220 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4221 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4222 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4223 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4224 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4225 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4229 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4230 if (op == MDB_NEXT_DUP)
4231 return MDB_NOTFOUND;
4235 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4237 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4238 DPUTS("=====> move to next sibling page");
4239 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4240 mc->mc_flags |= C_EOF;
4241 mc->mc_flags &= ~C_INITIALIZED;
4242 return MDB_NOTFOUND;
4244 mp = mc->mc_pg[mc->mc_top];
4245 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4247 mc->mc_ki[mc->mc_top]++;
4249 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4250 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4253 key->mv_size = mc->mc_db->md_pad;
4254 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4258 assert(IS_LEAF(mp));
4259 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4261 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4262 mdb_xcursor_init1(mc, leaf);
4265 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4268 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4269 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4270 if (rc != MDB_SUCCESS)
4275 MDB_GET_KEY(leaf, key);
4279 /** Move the cursor to the previous data item. */
4281 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4287 assert(mc->mc_flags & C_INITIALIZED);
4289 mp = mc->mc_pg[mc->mc_top];
4291 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4292 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4293 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4294 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4295 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4296 if (op != MDB_PREV || rc == MDB_SUCCESS)
4299 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4300 if (op == MDB_PREV_DUP)
4301 return MDB_NOTFOUND;
4306 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4308 if (mc->mc_ki[mc->mc_top] == 0) {
4309 DPUTS("=====> move to prev sibling page");
4310 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4311 mc->mc_flags &= ~C_INITIALIZED;
4312 return MDB_NOTFOUND;
4314 mp = mc->mc_pg[mc->mc_top];
4315 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4316 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4318 mc->mc_ki[mc->mc_top]--;
4320 mc->mc_flags &= ~C_EOF;
4322 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4323 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4326 key->mv_size = mc->mc_db->md_pad;
4327 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4331 assert(IS_LEAF(mp));
4332 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4334 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4335 mdb_xcursor_init1(mc, leaf);
4338 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4341 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4342 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4343 if (rc != MDB_SUCCESS)
4348 MDB_GET_KEY(leaf, key);
4352 /** Set the cursor on a specific data item. */
4354 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4355 MDB_cursor_op op, int *exactp)
4359 MDB_node *leaf = NULL;
4364 assert(key->mv_size > 0);
4366 /* See if we're already on the right page */
4367 if (mc->mc_flags & C_INITIALIZED) {
4370 mp = mc->mc_pg[mc->mc_top];
4372 mc->mc_ki[mc->mc_top] = 0;
4373 return MDB_NOTFOUND;
4375 if (mp->mp_flags & P_LEAF2) {
4376 nodekey.mv_size = mc->mc_db->md_pad;
4377 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4379 leaf = NODEPTR(mp, 0);
4380 MDB_GET_KEY(leaf, &nodekey);
4382 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4384 /* Probably happens rarely, but first node on the page
4385 * was the one we wanted.
4387 mc->mc_ki[mc->mc_top] = 0;
4394 unsigned int nkeys = NUMKEYS(mp);
4396 if (mp->mp_flags & P_LEAF2) {
4397 nodekey.mv_data = LEAF2KEY(mp,
4398 nkeys-1, nodekey.mv_size);
4400 leaf = NODEPTR(mp, nkeys-1);
4401 MDB_GET_KEY(leaf, &nodekey);
4403 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4405 /* last node was the one we wanted */
4406 mc->mc_ki[mc->mc_top] = nkeys-1;
4412 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4413 /* This is definitely the right page, skip search_page */
4414 if (mp->mp_flags & P_LEAF2) {
4415 nodekey.mv_data = LEAF2KEY(mp,
4416 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4418 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4419 MDB_GET_KEY(leaf, &nodekey);
4421 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4423 /* current node was the one we wanted */
4433 /* If any parents have right-sibs, search.
4434 * Otherwise, there's nothing further.
4436 for (i=0; i<mc->mc_top; i++)
4438 NUMKEYS(mc->mc_pg[i])-1)
4440 if (i == mc->mc_top) {
4441 /* There are no other pages */
4442 mc->mc_ki[mc->mc_top] = nkeys;
4443 return MDB_NOTFOUND;
4447 /* There are no other pages */
4448 mc->mc_ki[mc->mc_top] = 0;
4449 return MDB_NOTFOUND;
4453 rc = mdb_page_search(mc, key, 0);
4454 if (rc != MDB_SUCCESS)
4457 mp = mc->mc_pg[mc->mc_top];
4458 assert(IS_LEAF(mp));
4461 leaf = mdb_node_search(mc, key, exactp);
4462 if (exactp != NULL && !*exactp) {
4463 /* MDB_SET specified and not an exact match. */
4464 return MDB_NOTFOUND;
4468 DPUTS("===> inexact leaf not found, goto sibling");
4469 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4470 return rc; /* no entries matched */
4471 mp = mc->mc_pg[mc->mc_top];
4472 assert(IS_LEAF(mp));
4473 leaf = NODEPTR(mp, 0);
4477 mc->mc_flags |= C_INITIALIZED;
4478 mc->mc_flags &= ~C_EOF;
4481 key->mv_size = mc->mc_db->md_pad;
4482 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4486 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4487 mdb_xcursor_init1(mc, leaf);
4490 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4491 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4492 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4495 if (op == MDB_GET_BOTH) {
4501 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4502 if (rc != MDB_SUCCESS)
4505 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4507 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4509 rc = mc->mc_dbx->md_dcmp(data, &d2);
4511 if (op == MDB_GET_BOTH || rc > 0)
4512 return MDB_NOTFOUND;
4517 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4518 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4523 /* The key already matches in all other cases */
4524 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4525 MDB_GET_KEY(leaf, key);
4526 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4531 /** Move the cursor to the first item in the database. */
4533 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4538 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4539 rc = mdb_page_search(mc, NULL, 0);
4540 if (rc != MDB_SUCCESS)
4543 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4545 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4546 mc->mc_flags |= C_INITIALIZED;
4547 mc->mc_flags &= ~C_EOF;
4549 mc->mc_ki[mc->mc_top] = 0;
4551 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4552 key->mv_size = mc->mc_db->md_pad;
4553 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4558 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4559 mdb_xcursor_init1(mc, leaf);
4560 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4565 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4566 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4570 MDB_GET_KEY(leaf, key);
4574 /** Move the cursor to the last item in the database. */
4576 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4581 if (!(mc->mc_flags & C_EOF)) {
4583 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4586 lkey.mv_size = MDB_MAXKEYSIZE+1;
4587 lkey.mv_data = NULL;
4588 rc = mdb_page_search(mc, &lkey, 0);
4589 if (rc != MDB_SUCCESS)
4592 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4594 mc->mc_flags |= C_INITIALIZED|C_EOF;
4595 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4597 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4599 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4600 key->mv_size = mc->mc_db->md_pad;
4601 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4606 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4607 mdb_xcursor_init1(mc, leaf);
4608 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4613 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4614 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4619 MDB_GET_KEY(leaf, key);
4624 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4633 case MDB_GET_CURRENT:
4634 if (!(mc->mc_flags & C_INITIALIZED)) {
4637 MDB_page *mp = mc->mc_pg[mc->mc_top];
4639 mc->mc_ki[mc->mc_top] = 0;
4645 key->mv_size = mc->mc_db->md_pad;
4646 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4648 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4649 MDB_GET_KEY(leaf, key);
4651 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4652 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4654 rc = mdb_node_read(mc->mc_txn, leaf, data);
4661 case MDB_GET_BOTH_RANGE:
4662 if (data == NULL || mc->mc_xcursor == NULL) {
4670 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4672 } else if (op == MDB_SET_RANGE)
4673 rc = mdb_cursor_set(mc, key, data, op, NULL);
4675 rc = mdb_cursor_set(mc, key, data, op, &exact);
4677 case MDB_GET_MULTIPLE:
4679 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4680 !(mc->mc_flags & C_INITIALIZED)) {
4685 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4686 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4689 case MDB_NEXT_MULTIPLE:
4691 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4695 if (!(mc->mc_flags & C_INITIALIZED))
4696 rc = mdb_cursor_first(mc, key, data);
4698 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4699 if (rc == MDB_SUCCESS) {
4700 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4703 mx = &mc->mc_xcursor->mx_cursor;
4704 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4706 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4707 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4715 case MDB_NEXT_NODUP:
4716 if (!(mc->mc_flags & C_INITIALIZED))
4717 rc = mdb_cursor_first(mc, key, data);
4719 rc = mdb_cursor_next(mc, key, data, op);
4723 case MDB_PREV_NODUP:
4724 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4725 rc = mdb_cursor_last(mc, key, data);
4726 mc->mc_flags |= C_INITIALIZED;
4727 mc->mc_ki[mc->mc_top]++;
4729 rc = mdb_cursor_prev(mc, key, data, op);
4732 rc = mdb_cursor_first(mc, key, data);
4736 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4737 !(mc->mc_flags & C_INITIALIZED) ||
4738 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4742 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4745 rc = mdb_cursor_last(mc, key, data);
4749 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4750 !(mc->mc_flags & C_INITIALIZED) ||
4751 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4755 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4758 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4766 /** Touch all the pages in the cursor stack.
4767 * Makes sure all the pages are writable, before attempting a write operation.
4768 * @param[in] mc The cursor to operate on.
4771 mdb_cursor_touch(MDB_cursor *mc)
4775 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4778 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4779 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4780 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4783 *mc->mc_dbflag = DB_DIRTY;
4785 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4786 rc = mdb_page_touch(mc);
4790 mc->mc_top = mc->mc_snum-1;
4795 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4798 MDB_node *leaf = NULL;
4799 MDB_val xdata, *rdata, dkey;
4802 int do_sub = 0, insert = 0;
4803 unsigned int mcount = 0;
4807 char dbuf[MDB_MAXKEYSIZE+1];
4808 unsigned int nflags;
4811 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4814 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4817 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4820 #if SIZE_MAX > MAXDATASIZE
4821 if (data->mv_size > MAXDATASIZE)
4825 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4826 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4830 if (flags == MDB_CURRENT) {
4831 if (!(mc->mc_flags & C_INITIALIZED))
4834 } else if (mc->mc_db->md_root == P_INVALID) {
4836 /* new database, write a root leaf page */
4837 DPUTS("allocating new root leaf page");
4838 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4842 mdb_cursor_push(mc, np);
4843 mc->mc_db->md_root = np->mp_pgno;
4844 mc->mc_db->md_depth++;
4845 *mc->mc_dbflag = DB_DIRTY;
4846 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4848 np->mp_flags |= P_LEAF2;
4849 mc->mc_flags |= C_INITIALIZED;
4855 if (flags & MDB_APPEND) {
4857 rc = mdb_cursor_last(mc, &k2, &d2);
4859 rc = mc->mc_dbx->md_cmp(key, &k2);
4862 mc->mc_ki[mc->mc_top]++;
4868 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4870 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4871 DPRINTF("duplicate key [%s]", DKEY(key));
4873 return MDB_KEYEXIST;
4875 if (rc && rc != MDB_NOTFOUND)
4879 /* Cursor is positioned, now make sure all pages are writable */
4880 rc2 = mdb_cursor_touch(mc);
4885 /* The key already exists */
4886 if (rc == MDB_SUCCESS) {
4887 /* there's only a key anyway, so this is a no-op */
4888 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4889 unsigned int ksize = mc->mc_db->md_pad;
4890 if (key->mv_size != ksize)
4892 if (flags == MDB_CURRENT) {
4893 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4894 memcpy(ptr, key->mv_data, ksize);
4899 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4902 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4903 /* Was a single item before, must convert now */
4905 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4906 /* Just overwrite the current item */
4907 if (flags == MDB_CURRENT)
4910 dkey.mv_size = NODEDSZ(leaf);
4911 dkey.mv_data = NODEDATA(leaf);
4912 #if UINT_MAX < SIZE_MAX
4913 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4914 #ifdef MISALIGNED_OK
4915 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4917 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4920 /* if data matches, ignore it */
4921 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4922 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4924 /* create a fake page for the dup items */
4925 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4926 dkey.mv_data = dbuf;
4927 fp = (MDB_page *)&pbuf;
4928 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4929 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4930 fp->mp_lower = PAGEHDRSZ;
4931 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4932 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4933 fp->mp_flags |= P_LEAF2;
4934 fp->mp_pad = data->mv_size;
4935 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4937 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4938 (dkey.mv_size & 1) + (data->mv_size & 1);
4940 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4943 xdata.mv_size = fp->mp_upper;
4948 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4949 /* See if we need to convert from fake page to subDB */
4951 unsigned int offset;
4954 fp = NODEDATA(leaf);
4955 if (flags == MDB_CURRENT) {
4957 fp->mp_flags |= P_DIRTY;
4958 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4959 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4963 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4964 offset = fp->mp_pad;
4965 if (SIZELEFT(fp) >= offset)
4967 offset *= 4; /* space for 4 more */
4969 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4971 offset += offset & 1;
4972 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4973 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4975 /* yes, convert it */
4977 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4978 dummy.md_pad = fp->mp_pad;
4979 dummy.md_flags = MDB_DUPFIXED;
4980 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4981 dummy.md_flags |= MDB_INTEGERKEY;
4984 dummy.md_branch_pages = 0;
4985 dummy.md_leaf_pages = 1;
4986 dummy.md_overflow_pages = 0;
4987 dummy.md_entries = NUMKEYS(fp);
4989 xdata.mv_size = sizeof(MDB_db);
4990 xdata.mv_data = &dummy;
4991 if ((rc = mdb_page_alloc(mc, 1, &mp)))
4993 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4994 flags |= F_DUPDATA|F_SUBDATA;
4995 dummy.md_root = mp->mp_pgno;
4997 /* no, just grow it */
4999 xdata.mv_size = NODEDSZ(leaf) + offset;
5000 xdata.mv_data = &pbuf;
5001 mp = (MDB_page *)&pbuf;
5002 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5005 mp->mp_flags = fp->mp_flags | P_DIRTY;
5006 mp->mp_pad = fp->mp_pad;
5007 mp->mp_lower = fp->mp_lower;
5008 mp->mp_upper = fp->mp_upper + offset;
5010 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5012 nsize = NODEDSZ(leaf) - fp->mp_upper;
5013 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5014 for (i=0; i<NUMKEYS(fp); i++)
5015 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5017 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5021 /* data is on sub-DB, just store it */
5022 flags |= F_DUPDATA|F_SUBDATA;
5026 /* overflow page overwrites need special handling */
5027 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5030 int ovpages, dpages;
5032 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5033 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5034 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5035 mdb_page_get(mc->mc_txn, pg, &omp);
5036 /* Is the ov page writable and large enough? */
5037 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5038 /* yes, overwrite it. Note in this case we don't
5039 * bother to try shrinking the node if the new data
5040 * is smaller than the overflow threshold.
5042 if (F_ISSET(flags, MDB_RESERVE))
5043 data->mv_data = METADATA(omp);
5045 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5048 /* no, free ovpages */
5050 mc->mc_db->md_overflow_pages -= ovpages;
5051 for (i=0; i<ovpages; i++) {
5052 DPRINTF("freed ov page %zu", pg);
5053 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5057 } else if (NODEDSZ(leaf) == data->mv_size) {
5058 /* same size, just replace it. Note that we could
5059 * also reuse this node if the new data is smaller,
5060 * but instead we opt to shrink the node in that case.
5062 if (F_ISSET(flags, MDB_RESERVE))
5063 data->mv_data = NODEDATA(leaf);
5064 else if (data->mv_size)
5065 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5067 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5070 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5071 mc->mc_db->md_entries--;
5073 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5080 nflags = flags & NODE_ADD_FLAGS;
5081 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5082 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5083 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5084 nflags &= ~MDB_APPEND;
5086 nflags |= MDB_SPLIT_REPLACE;
5087 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5089 /* There is room already in this leaf page. */
5090 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5091 if (rc == 0 && !do_sub && insert) {
5092 /* Adjust other cursors pointing to mp */
5093 MDB_cursor *m2, *m3;
5094 MDB_dbi dbi = mc->mc_dbi;
5095 unsigned i = mc->mc_top;
5096 MDB_page *mp = mc->mc_pg[i];
5098 if (mc->mc_flags & C_SUB)
5101 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5102 if (mc->mc_flags & C_SUB)
5103 m3 = &m2->mc_xcursor->mx_cursor;
5106 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5107 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5114 if (rc != MDB_SUCCESS)
5115 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5117 /* Now store the actual data in the child DB. Note that we're
5118 * storing the user data in the keys field, so there are strict
5119 * size limits on dupdata. The actual data fields of the child
5120 * DB are all zero size.
5127 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5128 if (flags & MDB_CURRENT) {
5129 xflags = MDB_CURRENT;
5131 mdb_xcursor_init1(mc, leaf);
5132 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5134 /* converted, write the original data first */
5136 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5140 /* Adjust other cursors pointing to mp */
5142 unsigned i = mc->mc_top;
5143 MDB_page *mp = mc->mc_pg[i];
5145 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5146 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5147 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5148 mdb_xcursor_init1(m2, leaf);
5153 if (flags & MDB_APPENDDUP)
5154 xflags |= MDB_APPEND;
5155 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5156 if (flags & F_SUBDATA) {
5157 void *db = NODEDATA(leaf);
5158 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5161 /* sub-writes might have failed so check rc again.
5162 * Don't increment count if we just replaced an existing item.
5164 if (!rc && !(flags & MDB_CURRENT))
5165 mc->mc_db->md_entries++;
5166 if (flags & MDB_MULTIPLE) {
5168 if (mcount < data[1].mv_size) {
5169 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5170 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5176 /* If we succeeded and the key didn't exist before, make sure
5177 * the cursor is marked valid.
5180 mc->mc_flags |= C_INITIALIZED;
5185 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5190 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5193 if (!(mc->mc_flags & C_INITIALIZED))
5196 rc = mdb_cursor_touch(mc);
5200 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5202 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5203 if (flags != MDB_NODUPDATA) {
5204 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5205 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5207 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5208 /* If sub-DB still has entries, we're done */
5209 if (mc->mc_xcursor->mx_db.md_entries) {
5210 if (leaf->mn_flags & F_SUBDATA) {
5211 /* update subDB info */
5212 void *db = NODEDATA(leaf);
5213 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5215 /* shrink fake page */
5216 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5218 mc->mc_db->md_entries--;
5221 /* otherwise fall thru and delete the sub-DB */
5224 if (leaf->mn_flags & F_SUBDATA) {
5225 /* add all the child DB's pages to the free list */
5226 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5227 if (rc == MDB_SUCCESS) {
5228 mc->mc_db->md_entries -=
5229 mc->mc_xcursor->mx_db.md_entries;
5234 return mdb_cursor_del0(mc, leaf);
5237 /** Allocate and initialize new pages for a database.
5238 * @param[in] mc a cursor on the database being added to.
5239 * @param[in] flags flags defining what type of page is being allocated.
5240 * @param[in] num the number of pages to allocate. This is usually 1,
5241 * unless allocating overflow pages for a large record.
5242 * @param[out] mp Address of a page, or NULL on failure.
5243 * @return 0 on success, non-zero on failure.
5246 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5251 if ((rc = mdb_page_alloc(mc, num, &np)))
5253 DPRINTF("allocated new mpage %zu, page size %u",
5254 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5255 np->mp_flags = flags | P_DIRTY;
5256 np->mp_lower = PAGEHDRSZ;
5257 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5260 mc->mc_db->md_branch_pages++;
5261 else if (IS_LEAF(np))
5262 mc->mc_db->md_leaf_pages++;
5263 else if (IS_OVERFLOW(np)) {
5264 mc->mc_db->md_overflow_pages += num;
5272 /** Calculate the size of a leaf node.
5273 * The size depends on the environment's page size; if a data item
5274 * is too large it will be put onto an overflow page and the node
5275 * size will only include the key and not the data. Sizes are always
5276 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5277 * of the #MDB_node headers.
5278 * @param[in] env The environment handle.
5279 * @param[in] key The key for the node.
5280 * @param[in] data The data for the node.
5281 * @return The number of bytes needed to store the node.
5284 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5288 sz = LEAFSIZE(key, data);
5289 if (sz >= env->me_psize / MDB_MINKEYS) {
5290 /* put on overflow page */
5291 sz -= data->mv_size - sizeof(pgno_t);
5295 return sz + sizeof(indx_t);
5298 /** Calculate the size of a branch node.
5299 * The size should depend on the environment's page size but since
5300 * we currently don't support spilling large keys onto overflow
5301 * pages, it's simply the size of the #MDB_node header plus the
5302 * size of the key. Sizes are always rounded up to an even number
5303 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5304 * @param[in] env The environment handle.
5305 * @param[in] key The key for the node.
5306 * @return The number of bytes needed to store the node.
5309 mdb_branch_size(MDB_env *env, MDB_val *key)
5314 if (sz >= env->me_psize / MDB_MINKEYS) {
5315 /* put on overflow page */
5316 /* not implemented */
5317 /* sz -= key->size - sizeof(pgno_t); */
5320 return sz + sizeof(indx_t);
5323 /** Add a node to the page pointed to by the cursor.
5324 * @param[in] mc The cursor for this operation.
5325 * @param[in] indx The index on the page where the new node should be added.
5326 * @param[in] key The key for the new node.
5327 * @param[in] data The data for the new node, if any.
5328 * @param[in] pgno The page number, if adding a branch node.
5329 * @param[in] flags Flags for the node.
5330 * @return 0 on success, non-zero on failure. Possible errors are:
5332 * <li>ENOMEM - failed to allocate overflow pages for the node.
5333 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5334 * should never happen since all callers already calculate the
5335 * page's free space before calling this function.
5339 mdb_node_add(MDB_cursor *mc, indx_t indx,
5340 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5343 size_t node_size = NODESIZE;
5346 MDB_page *mp = mc->mc_pg[mc->mc_top];
5347 MDB_page *ofp = NULL; /* overflow page */
5350 assert(mp->mp_upper >= mp->mp_lower);
5352 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5353 IS_LEAF(mp) ? "leaf" : "branch",
5354 IS_SUBP(mp) ? "sub-" : "",
5355 mp->mp_pgno, indx, data ? data->mv_size : 0,
5356 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5359 /* Move higher keys up one slot. */
5360 int ksize = mc->mc_db->md_pad, dif;
5361 char *ptr = LEAF2KEY(mp, indx, ksize);
5362 dif = NUMKEYS(mp) - indx;
5364 memmove(ptr+ksize, ptr, dif*ksize);
5365 /* insert new key */
5366 memcpy(ptr, key->mv_data, ksize);
5368 /* Just using these for counting */
5369 mp->mp_lower += sizeof(indx_t);
5370 mp->mp_upper -= ksize - sizeof(indx_t);
5375 node_size += key->mv_size;
5379 if (F_ISSET(flags, F_BIGDATA)) {
5380 /* Data already on overflow page. */
5381 node_size += sizeof(pgno_t);
5382 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5383 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5385 /* Put data on overflow page. */
5386 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5387 data->mv_size, node_size+data->mv_size);
5388 node_size += sizeof(pgno_t);
5389 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5391 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5394 node_size += data->mv_size;
5397 node_size += node_size & 1;
5399 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5400 DPRINTF("not enough room in page %zu, got %u ptrs",
5401 mp->mp_pgno, NUMKEYS(mp));
5402 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5403 mp->mp_upper - mp->mp_lower);
5404 DPRINTF("node size = %zu", node_size);
5405 return MDB_PAGE_FULL;
5408 /* Move higher pointers up one slot. */
5409 for (i = NUMKEYS(mp); i > indx; i--)
5410 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5412 /* Adjust free space offsets. */
5413 ofs = mp->mp_upper - node_size;
5414 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5415 mp->mp_ptrs[indx] = ofs;
5417 mp->mp_lower += sizeof(indx_t);
5419 /* Write the node data. */
5420 node = NODEPTR(mp, indx);
5421 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5422 node->mn_flags = flags;
5424 SETDSZ(node,data->mv_size);
5429 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5434 if (F_ISSET(flags, F_BIGDATA))
5435 memcpy(node->mn_data + key->mv_size, data->mv_data,
5437 else if (F_ISSET(flags, MDB_RESERVE))
5438 data->mv_data = node->mn_data + key->mv_size;
5440 memcpy(node->mn_data + key->mv_size, data->mv_data,
5443 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5445 if (F_ISSET(flags, MDB_RESERVE))
5446 data->mv_data = METADATA(ofp);
5448 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5455 /** Delete the specified node from a page.
5456 * @param[in] mp The page to operate on.
5457 * @param[in] indx The index of the node to delete.
5458 * @param[in] ksize The size of a node. Only used if the page is
5459 * part of a #MDB_DUPFIXED database.
5462 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5465 indx_t i, j, numkeys, ptr;
5472 COPY_PGNO(pgno, mp->mp_pgno);
5473 DPRINTF("delete node %u on %s page %zu", indx,
5474 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5477 assert(indx < NUMKEYS(mp));
5480 int x = NUMKEYS(mp) - 1 - indx;
5481 base = LEAF2KEY(mp, indx, ksize);
5483 memmove(base, base + ksize, x * ksize);
5484 mp->mp_lower -= sizeof(indx_t);
5485 mp->mp_upper += ksize - sizeof(indx_t);
5489 node = NODEPTR(mp, indx);
5490 sz = NODESIZE + node->mn_ksize;
5492 if (F_ISSET(node->mn_flags, F_BIGDATA))
5493 sz += sizeof(pgno_t);
5495 sz += NODEDSZ(node);
5499 ptr = mp->mp_ptrs[indx];
5500 numkeys = NUMKEYS(mp);
5501 for (i = j = 0; i < numkeys; i++) {
5503 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5504 if (mp->mp_ptrs[i] < ptr)
5505 mp->mp_ptrs[j] += sz;
5510 base = (char *)mp + mp->mp_upper;
5511 memmove(base + sz, base, ptr - mp->mp_upper);
5513 mp->mp_lower -= sizeof(indx_t);
5517 /** Compact the main page after deleting a node on a subpage.
5518 * @param[in] mp The main page to operate on.
5519 * @param[in] indx The index of the subpage on the main page.
5522 mdb_node_shrink(MDB_page *mp, indx_t indx)
5529 indx_t i, numkeys, ptr;
5531 node = NODEPTR(mp, indx);
5532 sp = (MDB_page *)NODEDATA(node);
5533 osize = NODEDSZ(node);
5535 delta = sp->mp_upper - sp->mp_lower;
5536 SETDSZ(node, osize - delta);
5537 xp = (MDB_page *)((char *)sp + delta);
5539 /* shift subpage upward */
5541 nsize = NUMKEYS(sp) * sp->mp_pad;
5542 memmove(METADATA(xp), METADATA(sp), nsize);
5545 nsize = osize - sp->mp_upper;
5546 numkeys = NUMKEYS(sp);
5547 for (i=numkeys-1; i>=0; i--)
5548 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5550 xp->mp_upper = sp->mp_lower;
5551 xp->mp_lower = sp->mp_lower;
5552 xp->mp_flags = sp->mp_flags;
5553 xp->mp_pad = sp->mp_pad;
5554 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5556 /* shift lower nodes upward */
5557 ptr = mp->mp_ptrs[indx];
5558 numkeys = NUMKEYS(mp);
5559 for (i = 0; i < numkeys; i++) {
5560 if (mp->mp_ptrs[i] <= ptr)
5561 mp->mp_ptrs[i] += delta;
5564 base = (char *)mp + mp->mp_upper;
5565 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5566 mp->mp_upper += delta;
5569 /** Initial setup of a sorted-dups cursor.
5570 * Sorted duplicates are implemented as a sub-database for the given key.
5571 * The duplicate data items are actually keys of the sub-database.
5572 * Operations on the duplicate data items are performed using a sub-cursor
5573 * initialized when the sub-database is first accessed. This function does
5574 * the preliminary setup of the sub-cursor, filling in the fields that
5575 * depend only on the parent DB.
5576 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5579 mdb_xcursor_init0(MDB_cursor *mc)
5581 MDB_xcursor *mx = mc->mc_xcursor;
5583 mx->mx_cursor.mc_xcursor = NULL;
5584 mx->mx_cursor.mc_txn = mc->mc_txn;
5585 mx->mx_cursor.mc_db = &mx->mx_db;
5586 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5587 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5588 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5589 mx->mx_cursor.mc_snum = 0;
5590 mx->mx_cursor.mc_top = 0;
5591 mx->mx_cursor.mc_flags = C_SUB;
5592 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5593 mx->mx_dbx.md_dcmp = NULL;
5594 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5597 /** Final setup of a sorted-dups cursor.
5598 * Sets up the fields that depend on the data from the main cursor.
5599 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5600 * @param[in] node The data containing the #MDB_db record for the
5601 * sorted-dup database.
5604 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5606 MDB_xcursor *mx = mc->mc_xcursor;
5608 if (node->mn_flags & F_SUBDATA) {
5609 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5610 mx->mx_cursor.mc_pg[0] = 0;
5611 mx->mx_cursor.mc_snum = 0;
5612 mx->mx_cursor.mc_flags = C_SUB;
5614 MDB_page *fp = NODEDATA(node);
5615 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5616 mx->mx_db.md_flags = 0;
5617 mx->mx_db.md_depth = 1;
5618 mx->mx_db.md_branch_pages = 0;
5619 mx->mx_db.md_leaf_pages = 1;
5620 mx->mx_db.md_overflow_pages = 0;
5621 mx->mx_db.md_entries = NUMKEYS(fp);
5622 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5623 mx->mx_cursor.mc_snum = 1;
5624 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5625 mx->mx_cursor.mc_top = 0;
5626 mx->mx_cursor.mc_pg[0] = fp;
5627 mx->mx_cursor.mc_ki[0] = 0;
5628 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5629 mx->mx_db.md_flags = MDB_DUPFIXED;
5630 mx->mx_db.md_pad = fp->mp_pad;
5631 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5632 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5635 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5637 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5639 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5640 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5641 #if UINT_MAX < SIZE_MAX
5642 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5643 #ifdef MISALIGNED_OK
5644 mx->mx_dbx.md_cmp = mdb_cmp_long;
5646 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5651 /** Initialize a cursor for a given transaction and database. */
5653 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5658 mc->mc_db = &txn->mt_dbs[dbi];
5659 mc->mc_dbx = &txn->mt_dbxs[dbi];
5660 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5665 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5667 mc->mc_xcursor = mx;
5668 mdb_xcursor_init0(mc);
5670 mc->mc_xcursor = NULL;
5672 if (*mc->mc_dbflag & DB_STALE) {
5673 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5678 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5681 MDB_xcursor *mx = NULL;
5682 size_t size = sizeof(MDB_cursor);
5684 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5687 /* Allow read access to the freelist */
5688 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5691 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5692 size += sizeof(MDB_xcursor);
5694 if ((mc = malloc(size)) != NULL) {
5695 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5696 mx = (MDB_xcursor *)(mc + 1);
5698 mdb_cursor_init(mc, txn, dbi, mx);
5699 if (txn->mt_cursors) {
5700 mc->mc_next = txn->mt_cursors[dbi];
5701 txn->mt_cursors[dbi] = mc;
5703 mc->mc_flags |= C_ALLOCD;
5714 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5716 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5719 if (txn->mt_cursors)
5722 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5726 /* Return the count of duplicate data items for the current key */
5728 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5732 if (mc == NULL || countp == NULL)
5735 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5738 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5739 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5742 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5745 *countp = mc->mc_xcursor->mx_db.md_entries;
5751 mdb_cursor_close(MDB_cursor *mc)
5754 /* remove from txn, if tracked */
5755 if (mc->mc_txn->mt_cursors) {
5756 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5757 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5759 *prev = mc->mc_next;
5761 if (mc->mc_flags & C_ALLOCD)
5767 mdb_cursor_txn(MDB_cursor *mc)
5769 if (!mc) return NULL;
5774 mdb_cursor_dbi(MDB_cursor *mc)
5780 /** Replace the key for a node with a new key.
5781 * @param[in] mp The page containing the node to operate on.
5782 * @param[in] indx The index of the node to operate on.
5783 * @param[in] key The new key to use.
5784 * @return 0 on success, non-zero on failure.
5787 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5794 indx_t ptr, i, numkeys, indx;
5797 indx = mc->mc_ki[mc->mc_top];
5798 mp = mc->mc_pg[mc->mc_top];
5799 node = NODEPTR(mp, indx);
5800 ptr = mp->mp_ptrs[indx];
5804 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5805 k2.mv_data = NODEKEY(node);
5806 k2.mv_size = node->mn_ksize;
5807 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5809 mdb_dkey(&k2, kbuf2),
5815 delta0 = delta = key->mv_size - node->mn_ksize;
5817 /* Must be 2-byte aligned. If new key is
5818 * shorter by 1, the shift will be skipped.
5820 delta += (delta & 1);
5822 if (delta > 0 && SIZELEFT(mp) < delta) {
5824 /* not enough space left, do a delete and split */
5825 DPRINTF("Not enough room, delta = %d, splitting...", delta);
5826 pgno = NODEPGNO(node);
5827 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5828 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
5831 numkeys = NUMKEYS(mp);
5832 for (i = 0; i < numkeys; i++) {
5833 if (mp->mp_ptrs[i] <= ptr)
5834 mp->mp_ptrs[i] -= delta;
5837 base = (char *)mp + mp->mp_upper;
5838 len = ptr - mp->mp_upper + NODESIZE;
5839 memmove(base - delta, base, len);
5840 mp->mp_upper -= delta;
5842 node = NODEPTR(mp, indx);
5845 /* But even if no shift was needed, update ksize */
5847 node->mn_ksize = key->mv_size;
5850 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5856 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
5858 /** Move a node from csrc to cdst.
5861 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5868 unsigned short flags;
5872 /* Mark src and dst as dirty. */
5873 if ((rc = mdb_page_touch(csrc)) ||
5874 (rc = mdb_page_touch(cdst)))
5877 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5878 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5879 key.mv_size = csrc->mc_db->md_pad;
5880 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5882 data.mv_data = NULL;
5886 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5887 assert(!((long)srcnode&1));
5888 srcpg = NODEPGNO(srcnode);
5889 flags = srcnode->mn_flags;
5890 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5891 unsigned int snum = csrc->mc_snum;
5893 /* must find the lowest key below src */
5894 mdb_page_search_root(csrc, NULL, 0);
5895 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5896 key.mv_size = csrc->mc_db->md_pad;
5897 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5899 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5900 key.mv_size = NODEKSZ(s2);
5901 key.mv_data = NODEKEY(s2);
5903 csrc->mc_snum = snum--;
5904 csrc->mc_top = snum;
5906 key.mv_size = NODEKSZ(srcnode);
5907 key.mv_data = NODEKEY(srcnode);
5909 data.mv_size = NODEDSZ(srcnode);
5910 data.mv_data = NODEDATA(srcnode);
5912 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5913 unsigned int snum = cdst->mc_snum;
5916 /* must find the lowest key below dst */
5917 mdb_page_search_root(cdst, NULL, 0);
5918 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5919 bkey.mv_size = cdst->mc_db->md_pad;
5920 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5922 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5923 bkey.mv_size = NODEKSZ(s2);
5924 bkey.mv_data = NODEKEY(s2);
5926 cdst->mc_snum = snum--;
5927 cdst->mc_top = snum;
5928 mdb_cursor_copy(cdst, &mn);
5930 rc = mdb_update_key(&mn, &bkey);
5935 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5936 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5937 csrc->mc_ki[csrc->mc_top],
5939 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5940 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5942 /* Add the node to the destination page.
5944 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5945 if (rc != MDB_SUCCESS)
5948 /* Delete the node from the source page.
5950 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5953 /* Adjust other cursors pointing to mp */
5954 MDB_cursor *m2, *m3;
5955 MDB_dbi dbi = csrc->mc_dbi;
5956 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5958 if (csrc->mc_flags & C_SUB)
5961 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5962 if (m2 == csrc) continue;
5963 if (csrc->mc_flags & C_SUB)
5964 m3 = &m2->mc_xcursor->mx_cursor;
5967 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5968 csrc->mc_ki[csrc->mc_top]) {
5969 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5970 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5975 /* Update the parent separators.
5977 if (csrc->mc_ki[csrc->mc_top] == 0) {
5978 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5979 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5980 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5982 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5983 key.mv_size = NODEKSZ(srcnode);
5984 key.mv_data = NODEKEY(srcnode);
5986 DPRINTF("update separator for source page %zu to [%s]",
5987 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5988 mdb_cursor_copy(csrc, &mn);
5991 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
5994 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5996 indx_t ix = csrc->mc_ki[csrc->mc_top];
5997 nullkey.mv_size = 0;
5998 csrc->mc_ki[csrc->mc_top] = 0;
5999 rc = mdb_update_key(csrc, &nullkey);
6000 csrc->mc_ki[csrc->mc_top] = ix;
6001 assert(rc == MDB_SUCCESS);
6005 if (cdst->mc_ki[cdst->mc_top] == 0) {
6006 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6007 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6008 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6010 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6011 key.mv_size = NODEKSZ(srcnode);
6012 key.mv_data = NODEKEY(srcnode);
6014 DPRINTF("update separator for destination page %zu to [%s]",
6015 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6016 mdb_cursor_copy(cdst, &mn);
6019 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6022 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6024 indx_t ix = cdst->mc_ki[cdst->mc_top];
6025 nullkey.mv_size = 0;
6026 cdst->mc_ki[cdst->mc_top] = 0;
6027 rc = mdb_update_key(cdst, &nullkey);
6028 cdst->mc_ki[cdst->mc_top] = ix;
6029 assert(rc == MDB_SUCCESS);
6036 /** Merge one page into another.
6037 * The nodes from the page pointed to by \b csrc will
6038 * be copied to the page pointed to by \b cdst and then
6039 * the \b csrc page will be freed.
6040 * @param[in] csrc Cursor pointing to the source page.
6041 * @param[in] cdst Cursor pointing to the destination page.
6044 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6052 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6053 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6055 assert(csrc->mc_snum > 1); /* can't merge root page */
6056 assert(cdst->mc_snum > 1);
6058 /* Mark dst as dirty. */
6059 if ((rc = mdb_page_touch(cdst)))
6062 /* Move all nodes from src to dst.
6064 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6065 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6066 key.mv_size = csrc->mc_db->md_pad;
6067 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6068 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6069 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6070 if (rc != MDB_SUCCESS)
6072 key.mv_data = (char *)key.mv_data + key.mv_size;
6075 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6076 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6077 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6078 unsigned int snum = csrc->mc_snum;
6080 /* must find the lowest key below src */
6081 mdb_page_search_root(csrc, NULL, 0);
6082 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6083 key.mv_size = csrc->mc_db->md_pad;
6084 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6086 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6087 key.mv_size = NODEKSZ(s2);
6088 key.mv_data = NODEKEY(s2);
6090 csrc->mc_snum = snum--;
6091 csrc->mc_top = snum;
6093 key.mv_size = srcnode->mn_ksize;
6094 key.mv_data = NODEKEY(srcnode);
6097 data.mv_size = NODEDSZ(srcnode);
6098 data.mv_data = NODEDATA(srcnode);
6099 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6100 if (rc != MDB_SUCCESS)
6105 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6106 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);
6108 /* Unlink the src page from parent and add to free list.
6110 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6111 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6114 rc = mdb_update_key(csrc, &key);
6120 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6121 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6122 csrc->mc_db->md_leaf_pages--;
6124 csrc->mc_db->md_branch_pages--;
6126 /* Adjust other cursors pointing to mp */
6127 MDB_cursor *m2, *m3;
6128 MDB_dbi dbi = csrc->mc_dbi;
6129 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6131 if (csrc->mc_flags & C_SUB)
6134 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6135 if (csrc->mc_flags & C_SUB)
6136 m3 = &m2->mc_xcursor->mx_cursor;
6139 if (m3 == csrc) continue;
6140 if (m3->mc_snum < csrc->mc_snum) continue;
6141 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6142 m3->mc_pg[csrc->mc_top] = mp;
6143 m3->mc_ki[csrc->mc_top] += nkeys;
6147 mdb_cursor_pop(csrc);
6149 return mdb_rebalance(csrc);
6152 /** Copy the contents of a cursor.
6153 * @param[in] csrc The cursor to copy from.
6154 * @param[out] cdst The cursor to copy to.
6157 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6161 cdst->mc_txn = csrc->mc_txn;
6162 cdst->mc_dbi = csrc->mc_dbi;
6163 cdst->mc_db = csrc->mc_db;
6164 cdst->mc_dbx = csrc->mc_dbx;
6165 cdst->mc_snum = csrc->mc_snum;
6166 cdst->mc_top = csrc->mc_top;
6167 cdst->mc_flags = csrc->mc_flags;
6169 for (i=0; i<csrc->mc_snum; i++) {
6170 cdst->mc_pg[i] = csrc->mc_pg[i];
6171 cdst->mc_ki[i] = csrc->mc_ki[i];
6175 /** Rebalance the tree after a delete operation.
6176 * @param[in] mc Cursor pointing to the page where rebalancing
6178 * @return 0 on success, non-zero on failure.
6181 mdb_rebalance(MDB_cursor *mc)
6191 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6192 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6193 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6194 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6198 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
6201 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6202 DPRINTF("no need to rebalance page %zu, above fill threshold",
6208 if (mc->mc_snum < 2) {
6209 MDB_page *mp = mc->mc_pg[0];
6210 if (NUMKEYS(mp) == 0) {
6211 DPUTS("tree is completely empty");
6212 mc->mc_db->md_root = P_INVALID;
6213 mc->mc_db->md_depth = 0;
6214 mc->mc_db->md_leaf_pages = 0;
6215 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6219 /* Adjust other cursors pointing to mp */
6220 MDB_cursor *m2, *m3;
6221 MDB_dbi dbi = mc->mc_dbi;
6223 if (mc->mc_flags & C_SUB)
6226 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6227 if (m2 == mc) continue;
6228 if (mc->mc_flags & C_SUB)
6229 m3 = &m2->mc_xcursor->mx_cursor;
6232 if (m3->mc_snum < mc->mc_snum) continue;
6233 if (m3->mc_pg[0] == mp) {
6239 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6240 DPUTS("collapsing root page!");
6241 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6242 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6243 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6246 mc->mc_db->md_depth--;
6247 mc->mc_db->md_branch_pages--;
6249 /* Adjust other cursors pointing to mp */
6250 MDB_cursor *m2, *m3;
6251 MDB_dbi dbi = mc->mc_dbi;
6253 if (mc->mc_flags & C_SUB)
6256 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6257 if (m2 == mc) continue;
6258 if (mc->mc_flags & C_SUB)
6259 m3 = &m2->mc_xcursor->mx_cursor;
6262 if (m3->mc_snum < mc->mc_snum) continue;
6263 if (m3->mc_pg[0] == mp) {
6264 m3->mc_pg[0] = mc->mc_pg[0];
6269 DPUTS("root page doesn't need rebalancing");
6273 /* The parent (branch page) must have at least 2 pointers,
6274 * otherwise the tree is invalid.
6276 ptop = mc->mc_top-1;
6277 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6279 /* Leaf page fill factor is below the threshold.
6280 * Try to move keys from left or right neighbor, or
6281 * merge with a neighbor page.
6286 mdb_cursor_copy(mc, &mn);
6287 mn.mc_xcursor = NULL;
6289 if (mc->mc_ki[ptop] == 0) {
6290 /* We're the leftmost leaf in our parent.
6292 DPUTS("reading right neighbor");
6294 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6295 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6297 mn.mc_ki[mn.mc_top] = 0;
6298 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6300 /* There is at least one neighbor to the left.
6302 DPUTS("reading left neighbor");
6304 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6305 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6307 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6308 mc->mc_ki[mc->mc_top] = 0;
6311 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6312 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);
6314 /* If the neighbor page is above threshold and has at least two
6315 * keys, move one key from it.
6317 * Otherwise we should try to merge them.
6319 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6320 return mdb_node_move(&mn, mc);
6322 if (mc->mc_ki[ptop] == 0)
6323 rc = mdb_page_merge(&mn, mc);
6325 rc = mdb_page_merge(mc, &mn);
6326 mc->mc_flags &= ~C_INITIALIZED;
6331 /** Complete a delete operation started by #mdb_cursor_del(). */
6333 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6337 /* add overflow pages to free list */
6338 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6342 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6343 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6344 mc->mc_db->md_overflow_pages -= ovpages;
6345 for (i=0; i<ovpages; i++) {
6346 DPRINTF("freed ov page %zu", pg);
6347 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6351 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6352 mc->mc_db->md_entries--;
6353 rc = mdb_rebalance(mc);
6354 if (rc != MDB_SUCCESS)
6355 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6361 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6362 MDB_val *key, MDB_val *data)
6367 MDB_val rdata, *xdata;
6371 assert(key != NULL);
6373 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6375 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6378 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6382 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6386 mdb_cursor_init(&mc, txn, dbi, &mx);
6397 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6399 /* let mdb_page_split know about this cursor if needed:
6400 * delete will trigger a rebalance; if it needs to move
6401 * a node from one page to another, it will have to
6402 * update the parent's separator key(s). If the new sepkey
6403 * is larger than the current one, the parent page may
6404 * run out of space, triggering a split. We need this
6405 * cursor to be consistent until the end of the rebalance.
6407 mc.mc_next = txn->mt_cursors[dbi];
6408 txn->mt_cursors[dbi] = &mc;
6409 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6410 txn->mt_cursors[dbi] = mc.mc_next;
6415 /** Split a page and insert a new node.
6416 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6417 * The cursor will be updated to point to the actual page and index where
6418 * the node got inserted after the split.
6419 * @param[in] newkey The key for the newly inserted node.
6420 * @param[in] newdata The data for the newly inserted node.
6421 * @param[in] newpgno The page number, if the new node is a branch node.
6422 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6423 * @return 0 on success, non-zero on failure.
6426 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6427 unsigned int nflags)
6430 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6433 unsigned int i, j, split_indx, nkeys, pmax;
6435 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6437 MDB_page *mp, *rp, *pp;
6442 mp = mc->mc_pg[mc->mc_top];
6443 newindx = mc->mc_ki[mc->mc_top];
6445 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6446 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6447 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6449 /* Create a right sibling. */
6450 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6452 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6454 if (mc->mc_snum < 2) {
6455 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6457 /* shift current top to make room for new parent */
6458 mc->mc_pg[1] = mc->mc_pg[0];
6459 mc->mc_ki[1] = mc->mc_ki[0];
6462 mc->mc_db->md_root = pp->mp_pgno;
6463 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6464 mc->mc_db->md_depth++;
6467 /* Add left (implicit) pointer. */
6468 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6469 /* undo the pre-push */
6470 mc->mc_pg[0] = mc->mc_pg[1];
6471 mc->mc_ki[0] = mc->mc_ki[1];
6472 mc->mc_db->md_root = mp->mp_pgno;
6473 mc->mc_db->md_depth--;
6480 ptop = mc->mc_top-1;
6481 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6484 mc->mc_flags |= C_SPLITTING;
6485 mdb_cursor_copy(mc, &mn);
6486 mn.mc_pg[mn.mc_top] = rp;
6487 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6489 if (nflags & MDB_APPEND) {
6490 mn.mc_ki[mn.mc_top] = 0;
6492 split_indx = newindx;
6497 nkeys = NUMKEYS(mp);
6498 split_indx = nkeys / 2;
6499 if (newindx < split_indx)
6505 unsigned int lsize, rsize, ksize;
6506 /* Move half of the keys to the right sibling */
6508 x = mc->mc_ki[mc->mc_top] - split_indx;
6509 ksize = mc->mc_db->md_pad;
6510 split = LEAF2KEY(mp, split_indx, ksize);
6511 rsize = (nkeys - split_indx) * ksize;
6512 lsize = (nkeys - split_indx) * sizeof(indx_t);
6513 mp->mp_lower -= lsize;
6514 rp->mp_lower += lsize;
6515 mp->mp_upper += rsize - lsize;
6516 rp->mp_upper -= rsize - lsize;
6517 sepkey.mv_size = ksize;
6518 if (newindx == split_indx) {
6519 sepkey.mv_data = newkey->mv_data;
6521 sepkey.mv_data = split;
6524 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6525 memcpy(rp->mp_ptrs, split, rsize);
6526 sepkey.mv_data = rp->mp_ptrs;
6527 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6528 memcpy(ins, newkey->mv_data, ksize);
6529 mp->mp_lower += sizeof(indx_t);
6530 mp->mp_upper -= ksize - sizeof(indx_t);
6533 memcpy(rp->mp_ptrs, split, x * ksize);
6534 ins = LEAF2KEY(rp, x, ksize);
6535 memcpy(ins, newkey->mv_data, ksize);
6536 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6537 rp->mp_lower += sizeof(indx_t);
6538 rp->mp_upper -= ksize - sizeof(indx_t);
6539 mc->mc_ki[mc->mc_top] = x;
6540 mc->mc_pg[mc->mc_top] = rp;
6545 /* For leaf pages, check the split point based on what
6546 * fits where, since otherwise mdb_node_add can fail.
6548 * This check is only needed when the data items are
6549 * relatively large, such that being off by one will
6550 * make the difference between success or failure.
6552 * It's also relevant if a page happens to be laid out
6553 * such that one half of its nodes are all "small" and
6554 * the other half of its nodes are "large." If the new
6555 * item is also "large" and falls on the half with
6556 * "large" nodes, it also may not fit.
6559 unsigned int psize, nsize;
6560 /* Maximum free space in an empty page */
6561 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6562 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6563 if ((nkeys < 20) || (nsize > pmax/16)) {
6564 if (newindx <= split_indx) {
6567 for (i=0; i<split_indx; i++) {
6568 node = NODEPTR(mp, i);
6569 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6570 if (F_ISSET(node->mn_flags, F_BIGDATA))
6571 psize += sizeof(pgno_t);
6573 psize += NODEDSZ(node);
6577 split_indx = newindx;
6588 for (i=nkeys-1; i>=split_indx; i--) {
6589 node = NODEPTR(mp, i);
6590 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6591 if (F_ISSET(node->mn_flags, F_BIGDATA))
6592 psize += sizeof(pgno_t);
6594 psize += NODEDSZ(node);
6598 split_indx = newindx;
6609 /* First find the separating key between the split pages.
6610 * The case where newindx == split_indx is ambiguous; the
6611 * new item could go to the new page or stay on the original
6612 * page. If newpos == 1 it goes to the new page.
6614 if (newindx == split_indx && newpos) {
6615 sepkey.mv_size = newkey->mv_size;
6616 sepkey.mv_data = newkey->mv_data;
6618 node = NODEPTR(mp, split_indx);
6619 sepkey.mv_size = node->mn_ksize;
6620 sepkey.mv_data = NODEKEY(node);
6624 DPRINTF("separator is [%s]", DKEY(&sepkey));
6626 /* Copy separator key to the parent.
6628 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6632 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6635 if (mn.mc_snum == mc->mc_snum) {
6636 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6637 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6638 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6639 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6644 /* Right page might now have changed parent.
6645 * Check if left page also changed parent.
6647 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6648 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6649 for (i=0; i<ptop; i++) {
6650 mc->mc_pg[i] = mn.mc_pg[i];
6651 mc->mc_ki[i] = mn.mc_ki[i];
6653 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6654 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6658 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6661 mc->mc_flags ^= C_SPLITTING;
6662 if (rc != MDB_SUCCESS) {
6665 if (nflags & MDB_APPEND) {
6666 mc->mc_pg[mc->mc_top] = rp;
6667 mc->mc_ki[mc->mc_top] = 0;
6668 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6671 for (i=0; i<mc->mc_top; i++)
6672 mc->mc_ki[i] = mn.mc_ki[i];
6679 /* Move half of the keys to the right sibling. */
6681 /* grab a page to hold a temporary copy */
6682 copy = mdb_page_malloc(mc);
6686 copy->mp_pgno = mp->mp_pgno;
6687 copy->mp_flags = mp->mp_flags;
6688 copy->mp_lower = PAGEHDRSZ;
6689 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6690 mc->mc_pg[mc->mc_top] = copy;
6691 for (i = j = 0; i <= nkeys; j++) {
6692 if (i == split_indx) {
6693 /* Insert in right sibling. */
6694 /* Reset insert index for right sibling. */
6695 if (i != newindx || (newpos ^ ins_new)) {
6697 mc->mc_pg[mc->mc_top] = rp;
6701 if (i == newindx && !ins_new) {
6702 /* Insert the original entry that caused the split. */
6703 rkey.mv_data = newkey->mv_data;
6704 rkey.mv_size = newkey->mv_size;
6713 /* Update index for the new key. */
6714 mc->mc_ki[mc->mc_top] = j;
6715 } else if (i == nkeys) {
6718 node = NODEPTR(mp, i);
6719 rkey.mv_data = NODEKEY(node);
6720 rkey.mv_size = node->mn_ksize;
6722 xdata.mv_data = NODEDATA(node);
6723 xdata.mv_size = NODEDSZ(node);
6726 pgno = NODEPGNO(node);
6727 flags = node->mn_flags;
6732 if (!IS_LEAF(mp) && j == 0) {
6733 /* First branch index doesn't need key data. */
6737 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6741 nkeys = NUMKEYS(copy);
6742 for (i=0; i<nkeys; i++)
6743 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6744 mp->mp_lower = copy->mp_lower;
6745 mp->mp_upper = copy->mp_upper;
6746 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6747 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6749 /* reset back to original page */
6750 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6751 mc->mc_pg[mc->mc_top] = mp;
6752 if (nflags & MDB_RESERVE) {
6753 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6754 if (!(node->mn_flags & F_BIGDATA))
6755 newdata->mv_data = NODEDATA(node);
6761 /* return tmp page to freelist */
6762 mdb_page_free(mc->mc_txn->mt_env, copy);
6765 /* Adjust other cursors pointing to mp */
6766 MDB_cursor *m2, *m3;
6767 MDB_dbi dbi = mc->mc_dbi;
6768 int fixup = NUMKEYS(mp);
6770 if (mc->mc_flags & C_SUB)
6773 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6774 if (m2 == mc) continue;
6775 if (mc->mc_flags & C_SUB)
6776 m3 = &m2->mc_xcursor->mx_cursor;
6779 if (!(m3->mc_flags & C_INITIALIZED))
6781 if (m3->mc_flags & C_SPLITTING)
6786 for (k=m3->mc_top; k>=0; k--) {
6787 m3->mc_ki[k+1] = m3->mc_ki[k];
6788 m3->mc_pg[k+1] = m3->mc_pg[k];
6790 if (m3->mc_ki[0] >= split_indx) {
6795 m3->mc_pg[0] = mc->mc_pg[0];
6799 if (m3->mc_pg[mc->mc_top] == mp) {
6800 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6801 m3->mc_ki[mc->mc_top]++;
6802 if (m3->mc_ki[mc->mc_top] >= fixup) {
6803 m3->mc_pg[mc->mc_top] = rp;
6804 m3->mc_ki[mc->mc_top] -= fixup;
6805 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6807 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6808 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6817 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6818 MDB_val *key, MDB_val *data, unsigned int flags)
6823 assert(key != NULL);
6824 assert(data != NULL);
6826 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6829 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6833 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6837 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6840 mdb_cursor_init(&mc, txn, dbi, &mx);
6841 return mdb_cursor_put(&mc, key, data, flags);
6845 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6847 if ((flag & CHANGEABLE) != flag)
6850 env->me_flags |= flag;
6852 env->me_flags &= ~flag;
6857 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6862 *arg = env->me_flags;
6867 mdb_env_get_path(MDB_env *env, const char **arg)
6872 *arg = env->me_path;
6876 /** Common code for #mdb_stat() and #mdb_env_stat().
6877 * @param[in] env the environment to operate in.
6878 * @param[in] db the #MDB_db record containing the stats to return.
6879 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6880 * @return 0, this function always succeeds.
6883 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6885 arg->ms_psize = env->me_psize;
6886 arg->ms_depth = db->md_depth;
6887 arg->ms_branch_pages = db->md_branch_pages;
6888 arg->ms_leaf_pages = db->md_leaf_pages;
6889 arg->ms_overflow_pages = db->md_overflow_pages;
6890 arg->ms_entries = db->md_entries;
6895 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6899 if (env == NULL || arg == NULL)
6902 toggle = mdb_env_pick_meta(env);
6904 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6908 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
6912 if (env == NULL || arg == NULL)
6915 toggle = mdb_env_pick_meta(env);
6916 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
6917 arg->me_mapsize = env->me_mapsize;
6918 arg->me_maxreaders = env->me_maxreaders;
6919 arg->me_numreaders = env->me_numreaders;
6920 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
6921 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
6925 /** Set the default comparison functions for a database.
6926 * Called immediately after a database is opened to set the defaults.
6927 * The user can then override them with #mdb_set_compare() or
6928 * #mdb_set_dupsort().
6929 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6930 * @param[in] dbi A database handle returned by #mdb_dbi_open()
6933 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6935 uint16_t f = txn->mt_dbs[dbi].md_flags;
6937 txn->mt_dbxs[dbi].md_cmp =
6938 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6939 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6941 txn->mt_dbxs[dbi].md_dcmp =
6942 !(f & MDB_DUPSORT) ? 0 :
6943 ((f & MDB_INTEGERDUP)
6944 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6945 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6948 #define PERSISTENT_FLAGS 0xffff
6949 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
6950 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
6951 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6957 int rc, dbflag, exact;
6958 unsigned int unused = 0;
6961 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6962 mdb_default_cmp(txn, FREE_DBI);
6965 if ((flags & VALID_FLAGS) != flags)
6971 if (flags & PERSISTENT_FLAGS) {
6972 uint16_t f2 = flags & PERSISTENT_FLAGS;
6973 /* make sure flag changes get committed */
6974 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
6975 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
6976 txn->mt_flags |= MDB_TXN_DIRTY;
6979 mdb_default_cmp(txn, MAIN_DBI);
6983 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6984 mdb_default_cmp(txn, MAIN_DBI);
6987 /* Is the DB already open? */
6989 for (i=2; i<txn->mt_numdbs; i++) {
6990 if (!txn->mt_dbxs[i].md_name.mv_size) {
6991 /* Remember this free slot */
6992 if (!unused) unused = i;
6995 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6996 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7002 /* If no free slot and max hit, fail */
7003 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7004 return MDB_DBS_FULL;
7006 /* Find the DB info */
7010 key.mv_data = (void *)name;
7011 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7012 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7013 if (rc == MDB_SUCCESS) {
7014 /* make sure this is actually a DB */
7015 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7016 if (!(node->mn_flags & F_SUBDATA))
7018 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7019 /* Create if requested */
7021 data.mv_size = sizeof(MDB_db);
7022 data.mv_data = &dummy;
7023 memset(&dummy, 0, sizeof(dummy));
7024 dummy.md_root = P_INVALID;
7025 dummy.md_flags = flags & PERSISTENT_FLAGS;
7026 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7030 /* OK, got info, add to table */
7031 if (rc == MDB_SUCCESS) {
7032 unsigned int slot = unused ? unused : txn->mt_numdbs;
7033 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7034 txn->mt_dbxs[slot].md_name.mv_size = len;
7035 txn->mt_dbxs[slot].md_rel = NULL;
7036 txn->mt_dbflags[slot] = dbflag;
7037 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7039 txn->mt_env->me_dbflags[slot] = mdflags = txn->mt_dbs[slot].md_flags;
7040 mdb_default_cmp(txn, slot);
7043 txn->mt_env->me_numdbs++;
7045 /* Open the DB in parent txns as well */
7046 while ((txn = txn->mt_parent) != NULL) {
7047 txn->mt_dbflags[slot] = DB_STALE;
7048 txn->mt_dbs[slot].md_flags = mdflags;
7057 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7059 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7062 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7065 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7068 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
7070 ptr = env->me_dbxs[dbi].md_name.mv_data;
7071 env->me_dbxs[dbi].md_name.mv_data = NULL;
7072 env->me_dbxs[dbi].md_name.mv_size = 0;
7076 /** Add all the DB's pages to the free list.
7077 * @param[in] mc Cursor on the DB to free.
7078 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7079 * @return 0 on success, non-zero on failure.
7082 mdb_drop0(MDB_cursor *mc, int subs)
7086 rc = mdb_page_search(mc, NULL, 0);
7087 if (rc == MDB_SUCCESS) {
7092 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7093 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
7096 mdb_cursor_copy(mc, &mx);
7097 while (mc->mc_snum > 0) {
7098 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7099 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7100 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7101 if (ni->mn_flags & F_SUBDATA) {
7102 mdb_xcursor_init1(mc, ni);
7103 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7109 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7111 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7114 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7119 rc = mdb_cursor_sibling(mc, 1);
7121 /* no more siblings, go back to beginning
7122 * of previous level.
7125 for (i=1; i<mc->mc_top; i++)
7126 mc->mc_pg[i] = mx.mc_pg[i];
7130 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7131 mc->mc_db->md_root);
7136 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7141 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1)
7144 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7147 rc = mdb_cursor_open(txn, dbi, &mc);
7151 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7155 /* Can't delete the main DB */
7156 if (del && dbi > MAIN_DBI) {
7157 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7159 txn->mt_dbflags[dbi] = DB_STALE;
7160 mdb_dbi_close(txn->mt_env, dbi);
7163 /* reset the DB record, mark it dirty */
7164 txn->mt_dbflags[dbi] |= DB_DIRTY;
7165 txn->mt_dbs[dbi].md_depth = 0;
7166 txn->mt_dbs[dbi].md_branch_pages = 0;
7167 txn->mt_dbs[dbi].md_leaf_pages = 0;
7168 txn->mt_dbs[dbi].md_overflow_pages = 0;
7169 txn->mt_dbs[dbi].md_entries = 0;
7170 txn->mt_dbs[dbi].md_root = P_INVALID;
7172 if (!txn->mt_u.dirty_list[0].mid) {
7175 /* make sure we have at least one dirty page in this txn
7176 * otherwise these changes will be ignored.
7178 key.mv_size = sizeof(txnid_t);
7179 key.mv_data = &txn->mt_txnid;
7180 data.mv_size = sizeof(MDB_ID);
7181 data.mv_data = txn->mt_free_pgs;
7182 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
7183 rc = mdb_cursor_put(&m2, &key, &data, 0);
7187 mdb_cursor_close(mc);
7191 int mdb_set_compare(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_cmp = cmp;
7200 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7202 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7205 txn->mt_dbxs[dbi].md_dcmp = cmp;
7209 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7211 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7214 txn->mt_dbxs[dbi].md_rel = rel;
7218 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7220 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7223 txn->mt_dbxs[dbi].md_relctx = ctx;