2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2012 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
35 #include <sys/types.h>
37 #include <sys/param.h>
43 #ifdef HAVE_SYS_FILE_H
60 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
61 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
67 #include <semaphore.h>
72 #include <valgrind/memcheck.h>
73 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
74 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
75 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
76 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
77 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
79 #define VGMEMP_CREATE(h,r,z)
80 #define VGMEMP_ALLOC(h,a,s)
81 #define VGMEMP_FREE(h,a)
82 #define VGMEMP_DESTROY(h)
83 #define VGMEMP_DEFINED(a,s)
87 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
88 /* Solaris just defines one or the other */
89 # define LITTLE_ENDIAN 1234
90 # define BIG_ENDIAN 4321
91 # ifdef _LITTLE_ENDIAN
92 # define BYTE_ORDER LITTLE_ENDIAN
94 # define BYTE_ORDER BIG_ENDIAN
97 # define BYTE_ORDER __BYTE_ORDER
101 #ifndef LITTLE_ENDIAN
102 #define LITTLE_ENDIAN __LITTLE_ENDIAN
105 #define BIG_ENDIAN __BIG_ENDIAN
108 #if defined(__i386) || defined(__x86_64)
109 #define MISALIGNED_OK 1
115 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
116 # error "Unknown or unsupported endianness (BYTE_ORDER)"
117 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
118 # error "Two's complement, reasonably sized integer types, please"
121 /** @defgroup internal MDB Internals
124 /** @defgroup compat Windows Compatibility Macros
125 * A bunch of macros to minimize the amount of platform-specific ifdefs
126 * needed throughout the rest of the code. When the features this library
127 * needs are similar enough to POSIX to be hidden in a one-or-two line
128 * replacement, this macro approach is used.
132 #define pthread_t DWORD
133 #define pthread_mutex_t HANDLE
134 #define pthread_key_t DWORD
135 #define pthread_self() GetCurrentThreadId()
136 #define pthread_key_create(x,y) (*(x) = TlsAlloc())
137 #define pthread_key_delete(x) TlsFree(x)
138 #define pthread_getspecific(x) TlsGetValue(x)
139 #define pthread_setspecific(x,y) TlsSetValue(x,y)
140 #define pthread_mutex_unlock(x) ReleaseMutex(x)
141 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
142 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
143 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
144 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
145 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
146 #define getpid() GetCurrentProcessId()
147 #define fdatasync(fd) (!FlushFileBuffers(fd))
148 #define ErrCode() GetLastError()
149 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
150 #define close(fd) CloseHandle(fd)
151 #define munmap(ptr,len) UnmapViewOfFile(ptr)
154 #define LOCK_MUTEX_R(env) sem_wait((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) sem_wait((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
158 #define fdatasync(fd) fsync(fd)
161 #define fdatasync(fd) fsync(fd)
163 /** Lock the reader mutex.
165 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
166 /** Unlock the reader mutex.
168 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
170 /** Lock the writer mutex.
171 * Only a single write transaction is allowed at a time. Other writers
172 * will block waiting for this mutex.
174 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
175 /** Unlock the writer mutex.
177 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
178 #endif /* __APPLE__ */
180 /** Get the error code for the last failed system function.
182 #define ErrCode() errno
184 /** An abstraction for a file handle.
185 * On POSIX systems file handles are small integers. On Windows
186 * they're opaque pointers.
190 /** A value for an invalid file handle.
191 * Mainly used to initialize file variables and signify that they are
194 #define INVALID_HANDLE_VALUE (-1)
196 /** Get the size of a memory page for the system.
197 * This is the basic size that the platform's memory manager uses, and is
198 * fundamental to the use of memory-mapped files.
200 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
203 #if defined(_WIN32) || defined(__APPLE__)
206 #define MNAME_LEN (sizeof(pthread_mutex_t))
212 /** A flag for opening a file and requesting synchronous data writes.
213 * This is only used when writing a meta page. It's not strictly needed;
214 * we could just do a normal write and then immediately perform a flush.
215 * But if this flag is available it saves us an extra system call.
217 * @note If O_DSYNC is undefined but exists in /usr/include,
218 * preferably set some compiler flag to get the definition.
219 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
222 # define MDB_DSYNC O_DSYNC
226 /** Function for flushing the data of a file. Define this to fsync
227 * if fdatasync() is not supported.
229 #ifndef MDB_FDATASYNC
230 # define MDB_FDATASYNC fdatasync
233 /** A page number in the database.
234 * Note that 64 bit page numbers are overkill, since pages themselves
235 * already represent 12-13 bits of addressable memory, and the OS will
236 * always limit applications to a maximum of 63 bits of address space.
238 * @note In the #MDB_node structure, we only store 48 bits of this value,
239 * which thus limits us to only 60 bits of addressable data.
243 /** A transaction ID.
244 * See struct MDB_txn.mt_txnid for details.
248 /** @defgroup debug Debug Macros
252 /** Enable debug output.
253 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
254 * read from and written to the database (used for free space management).
259 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
260 # define DPRINTF (void) /* Vararg macros may be unsupported */
262 static int mdb_debug;
263 static int mdb_debug_start;
265 /** Print a debug message with printf formatting. */
266 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
267 if (mdb_debug) fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
269 # define DPRINTF(fmt, ...) ((void) 0)
271 /** Print a debug string.
272 * The string is printed literally, with no format processing.
274 #define DPUTS(arg) DPRINTF("%s", arg)
277 /** A default memory page size.
278 * The actual size is platform-dependent, but we use this for
279 * boot-strapping. We probably should not be using this any more.
280 * The #GET_PAGESIZE() macro is used to get the actual size.
282 * Note that we don't currently support Huge pages. On Linux,
283 * regular data files cannot use Huge pages, and in general
284 * Huge pages aren't actually pageable. We rely on the OS
285 * demand-pager to read our data and page it out when memory
286 * pressure from other processes is high. So until OSs have
287 * actual paging support for Huge pages, they're not viable.
289 #define MDB_PAGESIZE 4096
291 /** The minimum number of keys required in a database page.
292 * Setting this to a larger value will place a smaller bound on the
293 * maximum size of a data item. Data items larger than this size will
294 * be pushed into overflow pages instead of being stored directly in
295 * the B-tree node. This value used to default to 4. With a page size
296 * of 4096 bytes that meant that any item larger than 1024 bytes would
297 * go into an overflow page. That also meant that on average 2-3KB of
298 * each overflow page was wasted space. The value cannot be lower than
299 * 2 because then there would no longer be a tree structure. With this
300 * value, items larger than 2KB will go into overflow pages, and on
301 * average only 1KB will be wasted.
303 #define MDB_MINKEYS 2
305 /** A stamp that identifies a file as an MDB file.
306 * There's nothing special about this value other than that it is easily
307 * recognizable, and it will reflect any byte order mismatches.
309 #define MDB_MAGIC 0xBEEFC0DE
311 /** The version number for a database's file format. */
312 #define MDB_VERSION 1
314 /** The maximum size of a key in the database.
315 * While data items have essentially unbounded size, we require that
316 * keys all fit onto a regular page. This limit could be raised a bit
317 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
319 #define MAXKEYSIZE 511
324 * This is used for printing a hex dump of a key's contents.
326 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
327 /** Display a key in hex.
329 * Invoke a function to display a key in hex.
331 #define DKEY(x) mdb_dkey(x, kbuf)
333 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
337 /** @defgroup lazylock Lazy Locking
338 * Macros for locks that aren't actually needed.
339 * The DB view is always consistent because all writes are wrapped in
340 * the wmutex. Finer-grained locks aren't necessary.
344 /** Use lazy locking. I.e., don't lock these accesses at all. */
348 /** Grab the reader lock */
349 #define LAZY_MUTEX_LOCK(x)
350 /** Release the reader lock */
351 #define LAZY_MUTEX_UNLOCK(x)
352 /** Release the DB table reader/writer lock */
353 #define LAZY_RWLOCK_UNLOCK(x)
354 /** Grab the DB table write lock */
355 #define LAZY_RWLOCK_WRLOCK(x)
356 /** Grab the DB table read lock */
357 #define LAZY_RWLOCK_RDLOCK(x)
358 /** Declare the DB table rwlock. Should not be followed by ';'. */
359 #define LAZY_RWLOCK_DEF(x)
360 /** Initialize the DB table rwlock */
361 #define LAZY_RWLOCK_INIT(x,y)
362 /** Destroy the DB table rwlock */
363 #define LAZY_RWLOCK_DESTROY(x)
365 #define LAZY_MUTEX_LOCK(x) pthread_mutex_lock(x)
366 #define LAZY_MUTEX_UNLOCK(x) pthread_mutex_unlock(x)
367 #define LAZY_RWLOCK_UNLOCK(x) pthread_rwlock_unlock(x)
368 #define LAZY_RWLOCK_WRLOCK(x) pthread_rwlock_wrlock(x)
369 #define LAZY_RWLOCK_RDLOCK(x) pthread_rwlock_rdlock(x)
370 #define LAZY_RWLOCK_DEF(x) pthread_rwlock_t x;
371 #define LAZY_RWLOCK_INIT(x,y) pthread_rwlock_init(x,y)
372 #define LAZY_RWLOCK_DESTROY(x) pthread_rwlock_destroy(x)
376 /** An invalid page number.
377 * Mainly used to denote an empty tree.
379 #define P_INVALID (~0UL)
381 /** Test if a flag \b f is set in a flag word \b w. */
382 #define F_ISSET(w, f) (((w) & (f)) == (f))
384 /** Used for offsets within a single page.
385 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
388 typedef uint16_t indx_t;
390 /** Default size of memory map.
391 * This is certainly too small for any actual applications. Apps should always set
392 * the size explicitly using #mdb_env_set_mapsize().
394 #define DEFAULT_MAPSIZE 1048576
396 /** @defgroup readers Reader Lock Table
397 * Readers don't acquire any locks for their data access. Instead, they
398 * simply record their transaction ID in the reader table. The reader
399 * mutex is needed just to find an empty slot in the reader table. The
400 * slot's address is saved in thread-specific data so that subsequent read
401 * transactions started by the same thread need no further locking to proceed.
403 * Since the database uses multi-version concurrency control, readers don't
404 * actually need any locking. This table is used to keep track of which
405 * readers are using data from which old transactions, so that we'll know
406 * when a particular old transaction is no longer in use. Old transactions
407 * that have discarded any data pages can then have those pages reclaimed
408 * for use by a later write transaction.
410 * The lock table is constructed such that reader slots are aligned with the
411 * processor's cache line size. Any slot is only ever used by one thread.
412 * This alignment guarantees that there will be no contention or cache
413 * thrashing as threads update their own slot info, and also eliminates
414 * any need for locking when accessing a slot.
416 * A writer thread will scan every slot in the table to determine the oldest
417 * outstanding reader transaction. Any freed pages older than this will be
418 * reclaimed by the writer. The writer doesn't use any locks when scanning
419 * this table. This means that there's no guarantee that the writer will
420 * see the most up-to-date reader info, but that's not required for correct
421 * operation - all we need is to know the upper bound on the oldest reader,
422 * we don't care at all about the newest reader. So the only consequence of
423 * reading stale information here is that old pages might hang around a
424 * while longer before being reclaimed. That's actually good anyway, because
425 * the longer we delay reclaiming old pages, the more likely it is that a
426 * string of contiguous pages can be found after coalescing old pages from
427 * many old transactions together.
429 * @todo We don't actually do such coalescing yet, we grab pages from one
430 * old transaction at a time.
433 /** Number of slots in the reader table.
434 * This value was chosen somewhat arbitrarily. 126 readers plus a
435 * couple mutexes fit exactly into 8KB on my development machine.
436 * Applications should set the table size using #mdb_env_set_maxreaders().
438 #define DEFAULT_READERS 126
440 /** The size of a CPU cache line in bytes. We want our lock structures
441 * aligned to this size to avoid false cache line sharing in the
443 * This value works for most CPUs. For Itanium this should be 128.
449 /** The information we store in a single slot of the reader table.
450 * In addition to a transaction ID, we also record the process and
451 * thread ID that owns a slot, so that we can detect stale information,
452 * e.g. threads or processes that went away without cleaning up.
453 * @note We currently don't check for stale records. We simply re-init
454 * the table when we know that we're the only process opening the
457 typedef struct MDB_rxbody {
458 /** The current Transaction ID when this transaction began.
459 * Multiple readers that start at the same time will probably have the
460 * same ID here. Again, it's not important to exclude them from
461 * anything; all we need to know is which version of the DB they
462 * started from so we can avoid overwriting any data used in that
463 * particular version.
466 /** The process ID of the process owning this reader txn. */
468 /** The thread ID of the thread owning this txn. */
472 /** The actual reader record, with cacheline padding. */
473 typedef struct MDB_reader {
476 /** shorthand for mrb_txnid */
477 #define mr_txnid mru.mrx.mrb_txnid
478 #define mr_pid mru.mrx.mrb_pid
479 #define mr_tid mru.mrx.mrb_tid
480 /** cache line alignment */
481 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
485 /** The header for the reader table.
486 * The table resides in a memory-mapped file. (This is a different file
487 * than is used for the main database.)
489 * For POSIX the actual mutexes reside in the shared memory of this
490 * mapped file. On Windows, mutexes are named objects allocated by the
491 * kernel; we store the mutex names in this mapped file so that other
492 * processes can grab them. This same approach is also used on
493 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
494 * process-shared POSIX mutexes. For these cases where a named object
495 * is used, the object name is derived from a 64 bit FNV hash of the
496 * environment pathname. As such, naming collisions are extremely
497 * unlikely. If a collision occurs, the results are unpredictable.
499 typedef struct MDB_txbody {
500 /** Stamp identifying this as an MDB file. It must be set
503 /** Version number of this lock file. Must be set to #MDB_VERSION. */
504 uint32_t mtb_version;
505 #if defined(_WIN32) || defined(__APPLE__)
506 char mtb_rmname[MNAME_LEN];
508 /** Mutex protecting access to this table.
509 * This is the reader lock that #LOCK_MUTEX_R acquires.
511 pthread_mutex_t mtb_mutex;
513 /** The ID of the last transaction committed to the database.
514 * This is recorded here only for convenience; the value can always
515 * be determined by reading the main database meta pages.
518 /** The number of slots that have been used in the reader table.
519 * This always records the maximum count, it is not decremented
520 * when readers release their slots.
522 unsigned mtb_numreaders;
523 /** The ID of the most recent meta page in the database.
524 * This is recorded here only for convenience; the value can always
525 * be determined by reading the main database meta pages.
527 uint32_t mtb_me_toggle;
530 /** The actual reader table definition. */
531 typedef struct MDB_txninfo {
534 #define mti_magic mt1.mtb.mtb_magic
535 #define mti_version mt1.mtb.mtb_version
536 #define mti_mutex mt1.mtb.mtb_mutex
537 #define mti_rmname mt1.mtb.mtb_rmname
538 #define mti_txnid mt1.mtb.mtb_txnid
539 #define mti_numreaders mt1.mtb.mtb_numreaders
540 #define mti_me_toggle mt1.mtb.mtb_me_toggle
541 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
544 #if defined(_WIN32) || defined(__APPLE__)
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_SET_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_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 ID2L dirty_list; /**< modified pages */
816 MDB_reader *reader; /**< this thread's slot in the reader table */
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 /** 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 */
845 unsigned int mt_flags; /**< @ref mdb_txn */
846 /** Tracks which of the two meta pages was used at the start
847 * of this transaction.
849 unsigned int mt_toggle;
852 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
853 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
854 * raise this on a 64 bit machine.
856 #define CURSOR_STACK 32
860 /** Cursors are used for all DB operations */
862 /** Next cursor on this DB in this txn */
864 /** Original cursor if this is a shadow */
866 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
867 struct MDB_xcursor *mc_xcursor;
868 /** The transaction that owns this cursor */
870 /** The database handle this cursor operates on */
872 /** The database record for this cursor */
874 /** The database auxiliary record for this cursor */
876 /** The @ref mt_dbflag for this database */
877 unsigned char *mc_dbflag;
878 unsigned short mc_snum; /**< number of pushed pages */
879 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
880 /** @defgroup mdb_cursor Cursor Flags
882 * Cursor state flags.
885 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
886 #define C_EOF 0x02 /**< No more data */
887 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
888 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
889 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
891 unsigned int mc_flags; /**< @ref mdb_cursor */
892 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
893 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
896 /** Context for sorted-dup records.
897 * We could have gone to a fully recursive design, with arbitrarily
898 * deep nesting of sub-databases. But for now we only handle these
899 * levels - main DB, optional sub-DB, sorted-duplicate DB.
901 typedef struct MDB_xcursor {
902 /** A sub-cursor for traversing the Dup DB */
903 MDB_cursor mx_cursor;
904 /** The database record for this Dup DB */
906 /** The auxiliary DB record for this Dup DB */
908 /** The @ref mt_dbflag for this Dup DB */
909 unsigned char mx_dbflag;
912 /** A set of pages freed by an earlier transaction. */
913 typedef struct MDB_oldpages {
914 /** Usually we only read one record from the FREEDB at a time, but
915 * in case we read more, this will chain them together.
917 struct MDB_oldpages *mo_next;
918 /** The ID of the transaction in which these pages were freed. */
920 /** An #IDL of the pages */
921 pgno_t mo_pages[1]; /* dynamic */
924 /** The database environment. */
926 HANDLE me_fd; /**< The main data file */
927 HANDLE me_lfd; /**< The lock file */
928 HANDLE me_mfd; /**< just for writing the meta pages */
929 /** Failed to update the meta page. Probably an I/O error. */
930 #define MDB_FATAL_ERROR 0x80000000U
931 uint32_t me_flags; /**< @ref mdb_env */
932 uint32_t me_extrapad; /**< unused for now */
933 unsigned int me_maxreaders; /**< size of the reader table */
934 MDB_dbi me_numdbs; /**< number of DBs opened */
935 MDB_dbi me_maxdbs; /**< size of the DB table */
936 char *me_path; /**< path to the DB files */
937 char *me_map; /**< the memory map of the data file */
938 MDB_txninfo *me_txns; /**< the memory map of the lock file */
939 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
940 MDB_txn *me_txn; /**< current write transaction */
941 size_t me_mapsize; /**< size of the data memory map */
942 off_t me_size; /**< current file size */
943 pgno_t me_maxpg; /**< me_mapsize / me_psize */
944 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
945 unsigned int me_db_toggle; /**< which DB table is current */
946 txnid_t me_wtxnid; /**< ID of last txn we committed */
947 txnid_t me_pgfirst; /**< ID of first old page record we used */
948 txnid_t me_pglast; /**< ID of last old page record we used */
949 MDB_dbx *me_dbxs; /**< array of static DB info */
950 MDB_db *me_dbs[2]; /**< two arrays of MDB_db info */
951 MDB_oldpages *me_pghead; /**< list of old page records */
952 pthread_key_t me_txkey; /**< thread-key for readers */
953 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
954 /** IDL of pages that became unused in a write txn */
956 /** ID2L of pages that were written during a write txn */
957 ID2 me_dirty_list[MDB_IDL_UM_SIZE];
958 /** rwlock for the DB tables, if #LAZY_LOCKS is false */
959 LAZY_RWLOCK_DEF(me_dblock)
961 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
965 sem_t *me_rmutex; /* Apple doesn't support shared mutexes */
969 /** max number of pages to commit in one writev() call */
970 #define MDB_COMMIT_PAGES 64
971 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
972 #undef MDB_COMMIT_PAGES
973 #define MDB_COMMIT_PAGES IOV_MAX
976 static MDB_page *mdb_page_alloc(MDB_cursor *mc, int num);
977 static MDB_page *mdb_page_new(MDB_cursor *mc, uint32_t flags, int num);
978 static int mdb_page_touch(MDB_cursor *mc);
980 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
981 static int mdb_page_search_root(MDB_cursor *mc,
982 MDB_val *key, int modify);
983 static int mdb_page_search(MDB_cursor *mc,
984 MDB_val *key, int modify);
985 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
986 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
987 pgno_t newpgno, unsigned int nflags);
989 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
990 static int mdb_env_read_meta(MDB_env *env, int *which);
991 static int mdb_env_write_meta(MDB_txn *txn);
993 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
994 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
995 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
996 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
997 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
998 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
999 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1000 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1001 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1003 static int mdb_rebalance(MDB_cursor *mc);
1004 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
1006 static void mdb_cursor_pop(MDB_cursor *mc);
1007 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1009 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1010 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1011 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1012 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1013 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1015 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1016 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1018 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1019 static void mdb_xcursor_init0(MDB_cursor *mc);
1020 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1022 static int mdb_drop0(MDB_cursor *mc, int subs);
1023 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1026 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1030 static SECURITY_DESCRIPTOR mdb_null_sd;
1031 static SECURITY_ATTRIBUTES mdb_all_sa;
1032 static int mdb_sec_inited;
1035 /** Return the library version info. */
1037 mdb_version(int *major, int *minor, int *patch)
1039 if (major) *major = MDB_VERSION_MAJOR;
1040 if (minor) *minor = MDB_VERSION_MINOR;
1041 if (patch) *patch = MDB_VERSION_PATCH;
1042 return MDB_VERSION_STRING;
1045 /** Table of descriptions for MDB @ref errors */
1046 static char *const mdb_errstr[] = {
1047 "MDB_KEYEXIST: Key/data pair already exists",
1048 "MDB_NOTFOUND: No matching key/data pair found",
1049 "MDB_PAGE_NOTFOUND: Requested page not found",
1050 "MDB_CORRUPTED: Located page was wrong type",
1051 "MDB_PANIC: Update of meta page failed",
1052 "MDB_VERSION_MISMATCH: Database environment version mismatch"
1056 mdb_strerror(int err)
1059 return ("Successful return: 0");
1061 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
1062 return mdb_errstr[err - MDB_KEYEXIST];
1064 return strerror(err);
1068 /** Display a key in hexadecimal and return the address of the result.
1069 * @param[in] key the key to display
1070 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1071 * @return The key in hexadecimal form.
1074 mdb_dkey(MDB_val *key, char *buf)
1077 unsigned char *c = key->mv_data;
1079 if (key->mv_size > MAXKEYSIZE)
1080 return "MAXKEYSIZE";
1081 /* may want to make this a dynamic check: if the key is mostly
1082 * printable characters, print it as-is instead of converting to hex.
1086 for (i=0; i<key->mv_size; i++)
1087 ptr += sprintf(ptr, "%02x", *c++);
1089 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1094 /** Display all the keys in the page. */
1096 mdb_page_keys(MDB_page *mp)
1099 unsigned int i, nkeys;
1103 nkeys = NUMKEYS(mp);
1104 DPRINTF("numkeys %d", nkeys);
1105 for (i=0; i<nkeys; i++) {
1106 node = NODEPTR(mp, i);
1107 key.mv_size = node->mn_ksize;
1108 key.mv_data = node->mn_data;
1109 DPRINTF("key %d: %s", i, DKEY(&key));
1115 /** Count all the pages in each DB and in the freelist
1116 * and make sure it matches the actual number of pages
1119 static void mdb_audit(MDB_txn *txn)
1124 ID freecount, count;
1127 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1128 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1129 freecount += *(ID *)data.mv_data;
1130 freecount += txn->mt_dbs[0].md_branch_pages + txn->mt_dbs[0].md_leaf_pages +
1131 txn->mt_dbs[0].md_overflow_pages;
1134 for (i = 0; i<txn->mt_numdbs; i++) {
1135 count += txn->mt_dbs[i].md_branch_pages +
1136 txn->mt_dbs[i].md_leaf_pages +
1137 txn->mt_dbs[i].md_overflow_pages;
1138 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1140 mdb_cursor_init(&mc, txn, i, &mx);
1141 mdb_page_search(&mc, NULL, 0);
1145 mp = mc.mc_pg[mc.mc_top];
1146 for (j=0; j<NUMKEYS(mp); j++) {
1147 MDB_node *leaf = NODEPTR(mp, j);
1148 if (leaf->mn_flags & F_SUBDATA) {
1150 memcpy(&db, NODEDATA(leaf), sizeof(db));
1151 count += db.md_branch_pages + db.md_leaf_pages +
1152 db.md_overflow_pages;
1156 while (mdb_cursor_sibling(&mc, 1) == 0);
1159 assert(freecount + count + 2 >= txn->mt_next_pgno - 1);
1164 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1166 return txn->mt_dbxs[dbi].md_cmp(a, b);
1170 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1172 if (txn->mt_dbxs[dbi].md_dcmp)
1173 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1175 return EINVAL; /* too bad you can't distinguish this from a valid result */
1178 /** Allocate a single page.
1179 * Re-use old malloc'd pages first, otherwise just malloc.
1182 mdb_page_malloc(MDB_cursor *mc) {
1184 size_t sz = mc->mc_txn->mt_env->me_psize;
1185 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1186 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1187 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1188 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1189 } else if ((ret = malloc(sz)) != NULL) {
1190 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1195 /** Allocate pages for writing.
1196 * If there are free pages available from older transactions, they
1197 * will be re-used first. Otherwise a new page will be allocated.
1198 * @param[in] mc cursor A cursor handle identifying the transaction and
1199 * database for which we are allocating.
1200 * @param[in] num the number of pages to allocate.
1201 * @return Address of the allocated page(s). Requests for multiple pages
1202 * will always be satisfied by a single contiguous chunk of memory.
1205 mdb_page_alloc(MDB_cursor *mc, int num)
1207 MDB_txn *txn = mc->mc_txn;
1209 pgno_t pgno = P_INVALID;
1212 if (txn->mt_txnid > 2) {
1214 if (!txn->mt_env->me_pghead &&
1215 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1216 /* See if there's anything in the free DB */
1220 txnid_t *kptr, oldest, last;
1222 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1223 if (!txn->mt_env->me_pgfirst) {
1224 mdb_page_search(&m2, NULL, 0);
1225 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1226 kptr = (txnid_t *)NODEKEY(leaf);
1233 last = txn->mt_env->me_pglast + 1;
1235 key.mv_data = &last;
1236 key.mv_size = sizeof(last);
1237 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1240 last = *(txnid_t *)key.mv_data;
1245 oldest = txn->mt_txnid - 1;
1246 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
1247 txnid_t mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
1248 if (mr && mr < oldest)
1253 if (oldest > last) {
1254 /* It's usable, grab it.
1259 if (!txn->mt_env->me_pgfirst) {
1260 mdb_node_read(txn, leaf, &data);
1262 txn->mt_env->me_pglast = last;
1263 if (!txn->mt_env->me_pgfirst)
1264 txn->mt_env->me_pgfirst = last;
1265 idl = (ID *) data.mv_data;
1266 /* We might have a zero-length IDL due to freelist growth
1267 * during a prior commit
1269 if (!idl[0]) goto again;
1270 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1271 mop->mo_next = txn->mt_env->me_pghead;
1272 mop->mo_txnid = last;
1273 txn->mt_env->me_pghead = mop;
1274 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1279 DPRINTF("IDL read txn %zu root %zu num %zu",
1280 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1281 for (i=0; i<idl[0]; i++) {
1282 DPRINTF("IDL %zu", idl[i+1]);
1289 if (txn->mt_env->me_pghead) {
1290 MDB_oldpages *mop = txn->mt_env->me_pghead;
1292 /* FIXME: For now, always use fresh pages. We
1293 * really ought to search the free list for a
1298 /* peel pages off tail, so we only have to truncate the list */
1299 pgno = MDB_IDL_LAST(mop->mo_pages);
1300 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1302 if (mop->mo_pages[2] > mop->mo_pages[1])
1303 mop->mo_pages[0] = 0;
1307 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1308 txn->mt_env->me_pghead = mop->mo_next;
1315 if (pgno == P_INVALID) {
1316 /* DB size is maxed out */
1317 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1318 DPUTS("DB size maxed out");
1322 if (txn->mt_env->me_dpages && num == 1) {
1323 np = txn->mt_env->me_dpages;
1324 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1325 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1326 txn->mt_env->me_dpages = np->mp_next;
1328 size_t sz = txn->mt_env->me_psize * num;
1329 if ((np = malloc(sz)) == NULL)
1331 VGMEMP_ALLOC(txn->mt_env, np, sz);
1333 if (pgno == P_INVALID) {
1334 np->mp_pgno = txn->mt_next_pgno;
1335 txn->mt_next_pgno += num;
1339 mid.mid = np->mp_pgno;
1341 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1346 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1347 * @param[in] mc cursor pointing to the page to be touched
1348 * @return 0 on success, non-zero on failure.
1351 mdb_page_touch(MDB_cursor *mc)
1353 MDB_page *mp = mc->mc_pg[mc->mc_top];
1356 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1358 if ((np = mdb_page_alloc(mc, 1)) == NULL)
1360 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1361 assert(mp->mp_pgno != np->mp_pgno);
1362 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1364 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1367 mp->mp_flags |= P_DIRTY;
1370 /* Adjust other cursors pointing to mp */
1371 if (mc->mc_flags & C_SUB) {
1372 MDB_cursor *m2, *m3;
1373 MDB_dbi dbi = mc->mc_dbi-1;
1375 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1376 if (m2 == mc) continue;
1377 m3 = &m2->mc_xcursor->mx_cursor;
1378 if (m3->mc_snum < mc->mc_snum) continue;
1379 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1380 m3->mc_pg[mc->mc_top] = mp;
1386 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1387 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1388 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1389 m2->mc_pg[mc->mc_top] = mp;
1393 mc->mc_pg[mc->mc_top] = mp;
1394 /** If this page has a parent, update the parent to point to
1398 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1400 mc->mc_db->md_root = mp->mp_pgno;
1401 } else if (mc->mc_txn->mt_parent) {
1404 /* If txn has a parent, make sure the page is in our
1407 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1408 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1409 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1410 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1411 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1412 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1413 mc->mc_pg[mc->mc_top] = mp;
1419 np = mdb_page_malloc(mc);
1420 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1421 mid.mid = np->mp_pgno;
1423 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1431 mdb_env_sync(MDB_env *env, int force)
1434 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1435 if (MDB_FDATASYNC(env->me_fd))
1441 /** Make shadow copies of all of parent txn's cursors */
1443 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1445 MDB_cursor *mc, *m2;
1446 unsigned int i, j, size;
1448 for (i=0;i<src->mt_numdbs; i++) {
1449 if (src->mt_cursors[i]) {
1450 size = sizeof(MDB_cursor);
1451 if (src->mt_cursors[i]->mc_xcursor)
1452 size += sizeof(MDB_xcursor);
1453 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1460 mc->mc_db = &dst->mt_dbs[i];
1461 mc->mc_dbx = m2->mc_dbx;
1462 mc->mc_dbflag = &dst->mt_dbflags[i];
1463 mc->mc_snum = m2->mc_snum;
1464 mc->mc_top = m2->mc_top;
1465 mc->mc_flags = m2->mc_flags | C_SHADOW;
1466 for (j=0; j<mc->mc_snum; j++) {
1467 mc->mc_pg[j] = m2->mc_pg[j];
1468 mc->mc_ki[j] = m2->mc_ki[j];
1470 if (m2->mc_xcursor) {
1471 MDB_xcursor *mx, *mx2;
1472 mx = (MDB_xcursor *)(mc+1);
1473 mc->mc_xcursor = mx;
1474 mx2 = m2->mc_xcursor;
1475 mx->mx_db = mx2->mx_db;
1476 mx->mx_dbx = mx2->mx_dbx;
1477 mx->mx_dbflag = mx2->mx_dbflag;
1478 mx->mx_cursor.mc_txn = dst;
1479 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1480 mx->mx_cursor.mc_db = &mx->mx_db;
1481 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1482 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1483 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1484 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1485 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1486 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1487 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1488 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1491 mc->mc_xcursor = NULL;
1493 mc->mc_next = dst->mt_cursors[i];
1494 dst->mt_cursors[i] = mc;
1501 /** Merge shadow cursors back into parent's */
1503 mdb_cursor_merge(MDB_txn *txn)
1506 for (i=0; i<txn->mt_numdbs; i++) {
1507 if (txn->mt_cursors[i]) {
1509 while ((mc = txn->mt_cursors[i])) {
1510 txn->mt_cursors[i] = mc->mc_next;
1511 if (mc->mc_flags & C_SHADOW) {
1512 MDB_cursor *m2 = mc->mc_orig;
1514 m2->mc_snum = mc->mc_snum;
1515 m2->mc_top = mc->mc_top;
1516 for (j=0; j<mc->mc_snum; j++) {
1517 m2->mc_pg[j] = mc->mc_pg[j];
1518 m2->mc_ki[j] = mc->mc_ki[j];
1521 if (mc->mc_flags & C_ALLOCD)
1529 mdb_txn_reset0(MDB_txn *txn);
1531 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1532 * @param[in] txn the transaction handle to initialize
1533 * @return 0 on success, non-zero on failure. This can only
1534 * fail for read-only transactions, and then only if the
1535 * reader table is full.
1538 mdb_txn_renew0(MDB_txn *txn)
1540 MDB_env *env = txn->mt_env;
1543 if (txn->mt_flags & MDB_TXN_RDONLY) {
1544 MDB_reader *r = pthread_getspecific(env->me_txkey);
1547 pid_t pid = getpid();
1548 pthread_t tid = pthread_self();
1551 for (i=0; i<env->me_txns->mti_numreaders; i++)
1552 if (env->me_txns->mti_readers[i].mr_pid == 0)
1554 if (i == env->me_maxreaders) {
1555 UNLOCK_MUTEX_R(env);
1558 env->me_txns->mti_readers[i].mr_pid = pid;
1559 env->me_txns->mti_readers[i].mr_tid = tid;
1560 if (i >= env->me_txns->mti_numreaders)
1561 env->me_txns->mti_numreaders = i+1;
1562 UNLOCK_MUTEX_R(env);
1563 r = &env->me_txns->mti_readers[i];
1564 pthread_setspecific(env->me_txkey, r);
1566 txn->mt_toggle = env->me_txns->mti_me_toggle;
1567 txn->mt_txnid = env->me_txns->mti_txnid;
1568 /* This happens if a different process was the
1569 * last writer to the DB.
1571 if (env->me_wtxnid < txn->mt_txnid)
1572 mt_dbflag = DB_STALE;
1573 r->mr_txnid = txn->mt_txnid;
1574 txn->mt_u.reader = r;
1578 txn->mt_txnid = env->me_txns->mti_txnid;
1579 if (env->me_wtxnid < txn->mt_txnid)
1580 mt_dbflag = DB_STALE;
1583 if (txn->mt_txnid == mdb_debug_start)
1586 txn->mt_toggle = env->me_txns->mti_me_toggle;
1587 txn->mt_u.dirty_list = env->me_dirty_list;
1588 txn->mt_u.dirty_list[0].mid = 0;
1589 txn->mt_free_pgs = env->me_free_pgs;
1590 txn->mt_free_pgs[0] = 0;
1591 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1595 /* Copy the DB arrays */
1596 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
1597 txn->mt_numdbs = env->me_numdbs;
1598 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1599 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1600 if (txn->mt_numdbs > 2)
1601 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
1602 (txn->mt_numdbs - 2) * sizeof(MDB_db));
1603 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1605 memset(txn->mt_dbflags, mt_dbflag, env->me_numdbs);
1611 mdb_txn_renew(MDB_txn *txn)
1618 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1619 DPUTS("environment had fatal error, must shutdown!");
1623 rc = mdb_txn_renew0(txn);
1624 if (rc == MDB_SUCCESS) {
1625 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1626 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1627 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1633 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1638 if (env->me_flags & MDB_FATAL_ERROR) {
1639 DPUTS("environment had fatal error, must shutdown!");
1643 /* parent already has an active child txn */
1644 if (parent->mt_child) {
1648 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1649 if (!(flags & MDB_RDONLY))
1650 size += env->me_maxdbs * sizeof(MDB_cursor *);
1652 if ((txn = calloc(1, size)) == NULL) {
1653 DPRINTF("calloc: %s", strerror(ErrCode()));
1656 txn->mt_dbs = (MDB_db *)(txn+1);
1657 if (flags & MDB_RDONLY) {
1658 txn->mt_flags |= MDB_TXN_RDONLY;
1659 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1661 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1662 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1667 txn->mt_free_pgs = mdb_midl_alloc();
1668 if (!txn->mt_free_pgs) {
1672 txn->mt_u.dirty_list = malloc(sizeof(ID2)*MDB_IDL_UM_SIZE);
1673 if (!txn->mt_u.dirty_list) {
1674 free(txn->mt_free_pgs);
1678 txn->mt_txnid = parent->mt_txnid;
1679 txn->mt_toggle = parent->mt_toggle;
1680 txn->mt_u.dirty_list[0].mid = 0;
1681 txn->mt_free_pgs[0] = 0;
1682 txn->mt_next_pgno = parent->mt_next_pgno;
1683 parent->mt_child = txn;
1684 txn->mt_parent = parent;
1685 txn->mt_numdbs = parent->mt_numdbs;
1686 txn->mt_dbxs = parent->mt_dbxs;
1687 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1688 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1689 mdb_cursor_shadow(parent, txn);
1692 rc = mdb_txn_renew0(txn);
1698 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1699 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1700 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1706 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1707 * @param[in] txn the transaction handle to reset
1710 mdb_txn_reset0(MDB_txn *txn)
1712 MDB_env *env = txn->mt_env;
1714 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1715 txn->mt_u.reader->mr_txnid = 0;
1721 /* close(free) all cursors */
1722 for (i=0; i<txn->mt_numdbs; i++) {
1723 if (txn->mt_cursors[i]) {
1725 while ((mc = txn->mt_cursors[i])) {
1726 txn->mt_cursors[i] = mc->mc_next;
1727 if (mc->mc_flags & C_ALLOCD)
1733 /* return all dirty pages to dpage list */
1734 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1735 dp = txn->mt_u.dirty_list[i].mptr;
1736 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1737 dp->mp_next = txn->mt_env->me_dpages;
1738 VGMEMP_FREE(txn->mt_env, dp);
1739 txn->mt_env->me_dpages = dp;
1741 /* large pages just get freed directly */
1742 VGMEMP_FREE(txn->mt_env, dp);
1747 if (txn->mt_parent) {
1748 txn->mt_parent->mt_child = NULL;
1749 free(txn->mt_free_pgs);
1750 free(txn->mt_u.dirty_list);
1753 if (mdb_midl_shrink(&txn->mt_free_pgs))
1754 env->me_free_pgs = txn->mt_free_pgs;
1757 while ((mop = txn->mt_env->me_pghead)) {
1758 txn->mt_env->me_pghead = mop->mo_next;
1761 txn->mt_env->me_pgfirst = 0;
1762 txn->mt_env->me_pglast = 0;
1765 /* The writer mutex was locked in mdb_txn_begin. */
1766 UNLOCK_MUTEX_W(env);
1771 mdb_txn_reset(MDB_txn *txn)
1776 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1777 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1778 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1780 mdb_txn_reset0(txn);
1784 mdb_txn_abort(MDB_txn *txn)
1789 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1790 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1791 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1794 mdb_txn_abort(txn->mt_child);
1796 mdb_txn_reset0(txn);
1801 mdb_txn_commit(MDB_txn *txn)
1809 pgno_t next, freecnt;
1812 assert(txn != NULL);
1813 assert(txn->mt_env != NULL);
1815 if (txn->mt_child) {
1816 mdb_txn_commit(txn->mt_child);
1817 txn->mt_child = NULL;
1822 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1823 if (txn->mt_numdbs > env->me_numdbs) {
1824 /* update the DB tables */
1825 int toggle = !env->me_db_toggle;
1829 ip = &env->me_dbs[toggle][env->me_numdbs];
1830 jp = &txn->mt_dbs[env->me_numdbs];
1831 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1832 for (i = env->me_numdbs; i < txn->mt_numdbs; i++) {
1836 env->me_db_toggle = toggle;
1837 env->me_numdbs = txn->mt_numdbs;
1838 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1844 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1845 DPUTS("error flag is set, can't commit");
1847 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1852 /* Merge (and close) our cursors with parent's */
1853 mdb_cursor_merge(txn);
1855 if (txn->mt_parent) {
1861 /* Update parent's DB table */
1862 ip = &txn->mt_parent->mt_dbs[2];
1863 jp = &txn->mt_dbs[2];
1864 for (i = 2; i < txn->mt_numdbs; i++) {
1865 if (ip->md_root != jp->md_root)
1869 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1871 /* Append our free list to parent's */
1872 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1874 mdb_midl_free(txn->mt_free_pgs);
1876 /* Merge our dirty list with parent's */
1877 dst = txn->mt_parent->mt_u.dirty_list;
1878 src = txn->mt_u.dirty_list;
1879 x = mdb_mid2l_search(dst, src[1].mid);
1880 for (y=1; y<=src[0].mid; y++) {
1881 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1885 dst[x].mptr = src[y].mptr;
1888 for (; y<=src[0].mid; y++) {
1889 if (++x >= MDB_IDL_UM_MAX) {
1896 free(txn->mt_u.dirty_list);
1897 txn->mt_parent->mt_child = NULL;
1902 if (txn != env->me_txn) {
1903 DPUTS("attempt to commit unknown transaction");
1908 if (!txn->mt_u.dirty_list[0].mid)
1911 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1912 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1914 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1916 /* should only be one record now */
1917 if (env->me_pghead) {
1918 /* make sure first page of freeDB is touched and on freelist */
1919 mdb_page_search(&mc, NULL, 1);
1922 /* Delete IDLs we used from the free list */
1923 if (env->me_pgfirst) {
1928 key.mv_size = sizeof(cur);
1929 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
1932 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
1933 mdb_cursor_del(&mc, 0);
1935 env->me_pgfirst = 0;
1939 /* save to free list */
1941 freecnt = txn->mt_free_pgs[0];
1942 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1945 /* make sure last page of freeDB is touched and on freelist */
1946 key.mv_size = MAXKEYSIZE+1;
1948 mdb_page_search(&mc, &key, 1);
1950 mdb_midl_sort(txn->mt_free_pgs);
1954 ID *idl = txn->mt_free_pgs;
1955 DPRINTF("IDL write txn %zu root %zu num %zu",
1956 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1957 for (i=0; i<idl[0]; i++) {
1958 DPRINTF("IDL %zu", idl[i+1]);
1962 /* write to last page of freeDB */
1963 key.mv_size = sizeof(pgno_t);
1964 key.mv_data = &txn->mt_txnid;
1965 data.mv_data = txn->mt_free_pgs;
1966 /* The free list can still grow during this call,
1967 * despite the pre-emptive touches above. So check
1968 * and make sure the entire thing got written.
1971 freecnt = txn->mt_free_pgs[0];
1972 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1973 rc = mdb_cursor_put(&mc, &key, &data, 0);
1978 } while (freecnt != txn->mt_free_pgs[0]);
1980 /* should only be one record now */
1982 if (env->me_pghead) {
1988 mop = env->me_pghead;
1990 key.mv_size = sizeof(id);
1992 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1993 data.mv_data = mop->mo_pages;
1994 orig = mop->mo_pages[0];
1995 /* These steps may grow the freelist again
1996 * due to freed overflow pages...
1998 mdb_cursor_put(&mc, &key, &data, 0);
1999 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2000 /* could have been used again here */
2001 if (mop->mo_pages[0] != orig) {
2002 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2003 data.mv_data = mop->mo_pages;
2005 mdb_cursor_put(&mc, &key, &data, 0);
2007 env->me_pghead = NULL;
2010 /* was completely used up */
2011 mdb_cursor_del(&mc, 0);
2015 env->me_pgfirst = 0;
2018 /* Check for growth of freelist again */
2019 if (freecnt != txn->mt_free_pgs[0])
2022 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2023 if (mdb_midl_shrink(&txn->mt_free_pgs))
2024 env->me_free_pgs = txn->mt_free_pgs;
2027 /* Update DB root pointers. Their pages have already been
2028 * touched so this is all in-place and cannot fail.
2033 data.mv_size = sizeof(MDB_db);
2035 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2036 for (i = 2; i < txn->mt_numdbs; i++) {
2037 if (txn->mt_dbflags[i] & DB_DIRTY) {
2038 data.mv_data = &txn->mt_dbs[i];
2039 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2047 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2053 /* Windows actually supports scatter/gather I/O, but only on
2054 * unbuffered file handles. Since we're relying on the OS page
2055 * cache for all our data, that's self-defeating. So we just
2056 * write pages one at a time. We use the ov structure to set
2057 * the write offset, to at least save the overhead of a Seek
2061 memset(&ov, 0, sizeof(ov));
2062 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2064 dp = txn->mt_u.dirty_list[i].mptr;
2065 DPRINTF("committing page %zu", dp->mp_pgno);
2066 size = dp->mp_pgno * env->me_psize;
2067 ov.Offset = size & 0xffffffff;
2068 ov.OffsetHigh = size >> 16;
2069 ov.OffsetHigh >>= 16;
2070 /* clear dirty flag */
2071 dp->mp_flags &= ~P_DIRTY;
2072 wsize = env->me_psize;
2073 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2074 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2077 DPRINTF("WriteFile: %d", n);
2084 struct iovec iov[MDB_COMMIT_PAGES];
2088 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2089 dp = txn->mt_u.dirty_list[i].mptr;
2090 if (dp->mp_pgno != next) {
2092 rc = writev(env->me_fd, iov, n);
2096 DPUTS("short write, filesystem full?");
2098 DPRINTF("writev: %s", strerror(n));
2105 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2108 DPRINTF("committing page %zu", dp->mp_pgno);
2109 iov[n].iov_len = env->me_psize;
2110 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2111 iov[n].iov_base = (char *)dp;
2112 size += iov[n].iov_len;
2113 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2114 /* clear dirty flag */
2115 dp->mp_flags &= ~P_DIRTY;
2116 if (++n >= MDB_COMMIT_PAGES) {
2126 rc = writev(env->me_fd, iov, n);
2130 DPUTS("short write, filesystem full?");
2132 DPRINTF("writev: %s", strerror(n));
2139 /* Drop the dirty pages.
2141 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2142 dp = txn->mt_u.dirty_list[i].mptr;
2143 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2144 dp->mp_next = txn->mt_env->me_dpages;
2145 VGMEMP_FREE(txn->mt_env, dp);
2146 txn->mt_env->me_dpages = dp;
2148 VGMEMP_FREE(txn->mt_env, dp);
2151 txn->mt_u.dirty_list[i].mid = 0;
2153 txn->mt_u.dirty_list[0].mid = 0;
2155 if ((n = mdb_env_sync(env, 0)) != 0 ||
2156 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2160 env->me_wtxnid = txn->mt_txnid;
2164 /* update the DB tables */
2166 int toggle = !env->me_db_toggle;
2170 ip = &env->me_dbs[toggle][2];
2171 jp = &txn->mt_dbs[2];
2172 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
2173 for (i = 2; i < txn->mt_numdbs; i++) {
2174 if (ip->md_root != jp->md_root)
2179 env->me_db_toggle = toggle;
2180 env->me_numdbs = txn->mt_numdbs;
2181 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
2184 UNLOCK_MUTEX_W(env);
2190 /** Read the environment parameters of a DB environment before
2191 * mapping it into memory.
2192 * @param[in] env the environment handle
2193 * @param[out] meta address of where to store the meta information
2194 * @return 0 on success, non-zero on failure.
2197 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2204 /* We don't know the page size yet, so use a minimum value.
2208 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2210 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2215 else if (rc != MDB_PAGESIZE) {
2219 DPRINTF("read: %s", strerror(err));
2223 p = (MDB_page *)&pbuf;
2225 if (!F_ISSET(p->mp_flags, P_META)) {
2226 DPRINTF("page %zu not a meta page", p->mp_pgno);
2231 if (m->mm_magic != MDB_MAGIC) {
2232 DPUTS("meta has invalid magic");
2236 if (m->mm_version != MDB_VERSION) {
2237 DPRINTF("database is version %u, expected version %u",
2238 m->mm_version, MDB_VERSION);
2239 return MDB_VERSION_MISMATCH;
2242 memcpy(meta, m, sizeof(*m));
2246 /** Write the environment parameters of a freshly created DB environment.
2247 * @param[in] env the environment handle
2248 * @param[out] meta address of where to store the meta information
2249 * @return 0 on success, non-zero on failure.
2252 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2259 DPUTS("writing new meta page");
2261 GET_PAGESIZE(psize);
2263 meta->mm_magic = MDB_MAGIC;
2264 meta->mm_version = MDB_VERSION;
2265 meta->mm_psize = psize;
2266 meta->mm_last_pg = 1;
2267 meta->mm_flags = env->me_flags & 0xffff;
2268 meta->mm_flags |= MDB_INTEGERKEY;
2269 meta->mm_dbs[0].md_root = P_INVALID;
2270 meta->mm_dbs[1].md_root = P_INVALID;
2272 p = calloc(2, psize);
2274 p->mp_flags = P_META;
2277 memcpy(m, meta, sizeof(*meta));
2279 q = (MDB_page *)((char *)p + psize);
2282 q->mp_flags = P_META;
2285 memcpy(m, meta, sizeof(*meta));
2290 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2291 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2294 rc = write(env->me_fd, p, psize * 2);
2295 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2301 /** Update the environment info to commit a transaction.
2302 * @param[in] txn the transaction that's being committed
2303 * @return 0 on success, non-zero on failure.
2306 mdb_env_write_meta(MDB_txn *txn)
2309 MDB_meta meta, metab;
2311 int rc, len, toggle;
2317 assert(txn != NULL);
2318 assert(txn->mt_env != NULL);
2320 toggle = !txn->mt_toggle;
2321 DPRINTF("writing meta page %d for root page %zu",
2322 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2326 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2327 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2329 ptr = (char *)&meta;
2330 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2331 len = sizeof(MDB_meta) - off;
2334 meta.mm_dbs[0] = txn->mt_dbs[0];
2335 meta.mm_dbs[1] = txn->mt_dbs[1];
2336 meta.mm_last_pg = txn->mt_next_pgno - 1;
2337 meta.mm_txnid = txn->mt_txnid;
2340 off += env->me_psize;
2343 /* Write to the SYNC fd */
2346 memset(&ov, 0, sizeof(ov));
2348 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2351 rc = pwrite(env->me_mfd, ptr, len, off);
2356 DPUTS("write failed, disk error?");
2357 /* On a failure, the pagecache still contains the new data.
2358 * Write some old data back, to prevent it from being used.
2359 * Use the non-SYNC fd; we know it will fail anyway.
2361 meta.mm_last_pg = metab.mm_last_pg;
2362 meta.mm_txnid = metab.mm_txnid;
2364 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2366 r2 = pwrite(env->me_fd, ptr, len, off);
2368 env->me_flags |= MDB_FATAL_ERROR;
2371 /* Memory ordering issues are irrelevant; since the entire writer
2372 * is wrapped by wmutex, all of these changes will become visible
2373 * after the wmutex is unlocked. Since the DB is multi-version,
2374 * readers will get consistent data regardless of how fresh or
2375 * how stale their view of these values is.
2377 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
2378 txn->mt_env->me_txns->mti_me_toggle = toggle;
2379 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2380 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
2385 /** Check both meta pages to see which one is newer.
2386 * @param[in] env the environment handle
2387 * @param[out] which address of where to store the meta toggle ID
2388 * @return 0 on success, non-zero on failure.
2391 mdb_env_read_meta(MDB_env *env, int *which)
2395 assert(env != NULL);
2397 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2400 DPRINTF("Using meta page %d", toggle);
2407 mdb_env_create(MDB_env **env)
2411 e = calloc(1, sizeof(MDB_env));
2415 e->me_free_pgs = mdb_midl_alloc();
2416 if (!e->me_free_pgs) {
2420 e->me_maxreaders = DEFAULT_READERS;
2422 e->me_fd = INVALID_HANDLE_VALUE;
2423 e->me_lfd = INVALID_HANDLE_VALUE;
2424 e->me_mfd = INVALID_HANDLE_VALUE;
2425 VGMEMP_CREATE(e,0,0);
2431 mdb_env_set_mapsize(MDB_env *env, size_t size)
2435 env->me_mapsize = size;
2437 env->me_maxpg = env->me_mapsize / env->me_psize;
2442 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2446 env->me_maxdbs = dbs;
2451 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2453 if (env->me_map || readers < 1)
2455 env->me_maxreaders = readers;
2460 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2462 if (!env || !readers)
2464 *readers = env->me_maxreaders;
2468 /** Further setup required for opening an MDB environment
2471 mdb_env_open2(MDB_env *env, unsigned int flags)
2473 int i, newenv = 0, toggle;
2477 env->me_flags = flags;
2479 memset(&meta, 0, sizeof(meta));
2481 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2484 DPUTS("new mdbenv");
2488 if (!env->me_mapsize) {
2489 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2495 LONG sizelo, sizehi;
2496 sizelo = env->me_mapsize & 0xffffffff;
2497 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2499 /* Windows won't create mappings for zero length files.
2500 * Just allocate the maxsize right now.
2503 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2504 if (!SetEndOfFile(env->me_fd))
2506 SetFilePointer(env->me_fd, 0, NULL, 0);
2508 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
2509 sizehi, sizelo, NULL);
2512 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
2520 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2522 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
2524 if (env->me_map == MAP_FAILED) {
2531 meta.mm_mapsize = env->me_mapsize;
2532 if (flags & MDB_FIXEDMAP)
2533 meta.mm_address = env->me_map;
2534 i = mdb_env_init_meta(env, &meta);
2535 if (i != MDB_SUCCESS) {
2536 munmap(env->me_map, env->me_mapsize);
2540 env->me_psize = meta.mm_psize;
2542 env->me_maxpg = env->me_mapsize / env->me_psize;
2544 p = (MDB_page *)env->me_map;
2545 env->me_metas[0] = METADATA(p);
2546 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2548 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
2551 DPRINTF("opened database version %u, pagesize %u",
2552 env->me_metas[toggle]->mm_version, env->me_psize);
2553 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
2554 DPRINTF("entries: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
2555 DPRINTF("branch pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
2556 DPRINTF("leaf pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
2557 DPRINTF("overflow pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
2558 DPRINTF("root: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
2564 /** Release a reader thread's slot in the reader lock table.
2565 * This function is called automatically when a thread exits.
2566 * Windows doesn't support destructor callbacks for thread-specific storage,
2567 * so this function is not compiled there.
2568 * @param[in] ptr This points to the slot in the reader lock table.
2571 mdb_env_reader_dest(void *ptr)
2573 MDB_reader *reader = ptr;
2575 reader->mr_txnid = 0;
2581 /** Downgrade the exclusive lock on the region back to shared */
2583 mdb_env_share_locks(MDB_env *env)
2587 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2589 env->me_txns->mti_me_toggle = toggle;
2590 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2595 /* First acquire a shared lock. The Unlock will
2596 * then release the existing exclusive lock.
2598 memset(&ov, 0, sizeof(ov));
2599 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2600 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2604 struct flock lock_info;
2605 /* The shared lock replaces the existing lock */
2606 memset((void *)&lock_info, 0, sizeof(lock_info));
2607 lock_info.l_type = F_RDLCK;
2608 lock_info.l_whence = SEEK_SET;
2609 lock_info.l_start = 0;
2610 lock_info.l_len = 1;
2611 fcntl(env->me_lfd, F_SETLK, &lock_info);
2615 #if defined(_WIN32) || defined(__APPLE__)
2617 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2619 * @(#) $Revision: 5.1 $
2620 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2621 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2623 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2627 * Please do not copyright this code. This code is in the public domain.
2629 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2630 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2631 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2632 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2633 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2634 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2635 * PERFORMANCE OF THIS SOFTWARE.
2638 * chongo <Landon Curt Noll> /\oo/\
2639 * http://www.isthe.com/chongo/
2641 * Share and Enjoy! :-)
2644 typedef unsigned long long mdb_hash_t;
2645 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2647 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2648 * @param[in] str string to hash
2649 * @param[in] hval initial value for hash
2650 * @return 64 bit hash
2652 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2653 * hval arg on the first call.
2656 mdb_hash_str(char *str, mdb_hash_t hval)
2658 unsigned char *s = (unsigned char *)str; /* unsigned string */
2660 * FNV-1a hash each octet of the string
2663 /* xor the bottom with the current octet */
2664 hval ^= (mdb_hash_t)*s++;
2666 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2667 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2668 (hval << 7) + (hval << 8) + (hval << 40);
2670 /* return our new hash value */
2674 /** Hash the string and output the hash in hex.
2675 * @param[in] str string to hash
2676 * @param[out] hexbuf an array of 17 chars to hold the hash
2679 mdb_hash_hex(char *str, char *hexbuf)
2682 mdb_hash_t h = mdb_hash_str(str, MDB_HASH_INIT);
2683 for (i=0; i<8; i++) {
2684 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2690 /** Open and/or initialize the lock region for the environment.
2691 * @param[in] env The MDB environment.
2692 * @param[in] lpath The pathname of the file used for the lock region.
2693 * @param[in] mode The Unix permissions for the file, if we create it.
2694 * @param[out] excl Set to true if we got an exclusive lock on the region.
2695 * @return 0 on success, non-zero on failure.
2698 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2706 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2707 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2708 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2712 /* Try to get exclusive lock. If we succeed, then
2713 * nobody is using the lock region and we should initialize it.
2716 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2720 memset(&ov, 0, sizeof(ov));
2721 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2727 size = GetFileSize(env->me_lfd, NULL);
2729 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2733 /* Try to get exclusive lock. If we succeed, then
2734 * nobody is using the lock region and we should initialize it.
2737 struct flock lock_info;
2738 memset((void *)&lock_info, 0, sizeof(lock_info));
2739 lock_info.l_type = F_WRLCK;
2740 lock_info.l_whence = SEEK_SET;
2741 lock_info.l_start = 0;
2742 lock_info.l_len = 1;
2743 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2747 lock_info.l_type = F_RDLCK;
2748 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2755 size = lseek(env->me_lfd, 0, SEEK_END);
2757 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2758 if (size < rsize && *excl) {
2760 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2761 if (!SetEndOfFile(env->me_lfd)) {
2766 if (ftruncate(env->me_lfd, rsize) != 0) {
2773 size = rsize - sizeof(MDB_txninfo);
2774 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2779 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2785 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2787 if (!env->me_txns) {
2792 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2794 if (m == MAP_FAILED) {
2795 env->me_txns = NULL;
2805 if (!mdb_sec_inited) {
2806 InitializeSecurityDescriptor(&mdb_null_sd,
2807 SECURITY_DESCRIPTOR_REVISION);
2808 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2809 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2810 mdb_all_sa.bInheritHandle = FALSE;
2811 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2814 mdb_hash_hex(lpath, hexbuf);
2815 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
2816 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2817 if (!env->me_rmutex) {
2821 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
2822 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
2823 if (!env->me_wmutex) {
2830 mdb_hash_hex(lpath, hexbuf);
2831 sprintf(env->me_txns->mti_rmname, "MDBr%s", hexbuf);
2832 if (sem_unlink(env->me_txns->mti_rmname)) {
2834 if (rc != ENOENT && rc != EINVAL)
2837 env->me_rmutex = sem_open(env->me_txns->mti_rmname, O_CREAT, mode, 1);
2838 if (!env->me_rmutex) {
2842 sprintf(env->me_txns->mti_wmname, "MDBw%s", hexbuf);
2843 if (sem_unlink(env->me_txns->mti_wmname)) {
2845 if (rc != ENOENT && rc != EINVAL)
2848 env->me_wmutex = sem_open(env->me_txns->mti_wmname, O_CREAT, mode, 1);
2849 if (!env->me_wmutex) {
2853 #else /* __APPLE__ */
2854 pthread_mutexattr_t mattr;
2856 pthread_mutexattr_init(&mattr);
2857 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2861 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2862 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2863 #endif /* __APPLE__ */
2865 env->me_txns->mti_version = MDB_VERSION;
2866 env->me_txns->mti_magic = MDB_MAGIC;
2867 env->me_txns->mti_txnid = 0;
2868 env->me_txns->mti_numreaders = 0;
2869 env->me_txns->mti_me_toggle = 0;
2872 if (env->me_txns->mti_magic != MDB_MAGIC) {
2873 DPUTS("lock region has invalid magic");
2877 if (env->me_txns->mti_version != MDB_VERSION) {
2878 DPRINTF("lock region is version %u, expected version %u",
2879 env->me_txns->mti_version, MDB_VERSION);
2880 rc = MDB_VERSION_MISMATCH;
2884 if (rc != EACCES && rc != EAGAIN) {
2888 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2889 if (!env->me_rmutex) {
2893 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2894 if (!env->me_wmutex) {
2900 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
2901 if (!env->me_rmutex) {
2905 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
2906 if (!env->me_wmutex) {
2916 env->me_lfd = INVALID_HANDLE_VALUE;
2921 /** The name of the lock file in the DB environment */
2922 #define LOCKNAME "/lock.mdb"
2923 /** The name of the data file in the DB environment */
2924 #define DATANAME "/data.mdb"
2925 /** The suffix of the lock file when no subdir is used */
2926 #define LOCKSUFF "-lock"
2929 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2931 int oflags, rc, len, excl;
2932 char *lpath, *dpath;
2935 if (flags & MDB_NOSUBDIR) {
2936 rc = len + sizeof(LOCKSUFF) + len + 1;
2938 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
2943 if (flags & MDB_NOSUBDIR) {
2944 dpath = lpath + len + sizeof(LOCKSUFF);
2945 sprintf(lpath, "%s" LOCKSUFF, path);
2946 strcpy(dpath, path);
2948 dpath = lpath + len + sizeof(LOCKNAME);
2949 sprintf(lpath, "%s" LOCKNAME, path);
2950 sprintf(dpath, "%s" DATANAME, path);
2953 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
2958 if (F_ISSET(flags, MDB_RDONLY)) {
2959 oflags = GENERIC_READ;
2960 len = OPEN_EXISTING;
2962 oflags = GENERIC_READ|GENERIC_WRITE;
2965 mode = FILE_ATTRIBUTE_NORMAL;
2966 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2967 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2972 if (F_ISSET(flags, MDB_RDONLY))
2975 oflags = O_RDWR | O_CREAT;
2977 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
2983 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
2984 /* synchronous fd for meta writes */
2986 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2987 mode |= FILE_FLAG_WRITE_THROUGH;
2988 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2989 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2994 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2995 oflags |= MDB_DSYNC;
2996 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
3001 env->me_path = strdup(path);
3002 DPRINTF("opened dbenv %p", (void *) env);
3003 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3004 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
3006 mdb_env_share_locks(env);
3007 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3008 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
3009 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
3015 if (env->me_fd != INVALID_HANDLE_VALUE) {
3017 env->me_fd = INVALID_HANDLE_VALUE;
3019 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3021 env->me_lfd = INVALID_HANDLE_VALUE;
3029 mdb_env_close(MDB_env *env)
3036 VGMEMP_DESTROY(env);
3037 while (env->me_dpages) {
3038 dp = env->me_dpages;
3039 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3040 env->me_dpages = dp->mp_next;
3044 free(env->me_dbs[1]);
3045 free(env->me_dbs[0]);
3049 LAZY_RWLOCK_DESTROY(&env->me_dblock);
3050 pthread_key_delete(env->me_txkey);
3053 munmap(env->me_map, env->me_mapsize);
3058 pid_t pid = getpid();
3060 for (i=0; i<env->me_txns->mti_numreaders; i++)
3061 if (env->me_txns->mti_readers[i].mr_pid == pid)
3062 env->me_txns->mti_readers[i].mr_pid = 0;
3063 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3066 mdb_midl_free(env->me_free_pgs);
3070 /** Compare two items pointing at aligned size_t's */
3072 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3074 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3075 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3078 /** Compare two items pointing at aligned int's */
3080 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3082 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3083 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3086 /** Compare two items pointing at ints of unknown alignment.
3087 * Nodes and keys are guaranteed to be 2-byte aligned.
3090 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3092 #if BYTE_ORDER == LITTLE_ENDIAN
3093 unsigned short *u, *c;
3096 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3097 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3100 } while(!x && u > (unsigned short *)a->mv_data);
3103 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3107 /** Compare two items lexically */
3109 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3116 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3122 diff = memcmp(a->mv_data, b->mv_data, len);
3123 return diff ? diff : len_diff<0 ? -1 : len_diff;
3126 /** Compare two items in reverse byte order */
3128 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3130 const unsigned char *p1, *p2, *p1_lim;
3134 p1_lim = (const unsigned char *)a->mv_data;
3135 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3136 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3138 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3144 while (p1 > p1_lim) {
3145 diff = *--p1 - *--p2;
3149 return len_diff<0 ? -1 : len_diff;
3152 /** Search for key within a page, using binary search.
3153 * Returns the smallest entry larger or equal to the key.
3154 * If exactp is non-null, stores whether the found entry was an exact match
3155 * in *exactp (1 or 0).
3156 * Updates the cursor index with the index of the found entry.
3157 * If no entry larger or equal to the key is found, returns NULL.
3160 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3162 unsigned int i = 0, nkeys;
3165 MDB_page *mp = mc->mc_pg[mc->mc_top];
3166 MDB_node *node = NULL;
3171 nkeys = NUMKEYS(mp);
3176 COPY_PGNO(pgno, mp->mp_pgno);
3177 DPRINTF("searching %u keys in %s %spage %zu",
3178 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3185 low = IS_LEAF(mp) ? 0 : 1;
3187 cmp = mc->mc_dbx->md_cmp;
3189 /* Branch pages have no data, so if using integer keys,
3190 * alignment is guaranteed. Use faster mdb_cmp_int.
3192 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3193 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3200 nodekey.mv_size = mc->mc_db->md_pad;
3201 node = NODEPTR(mp, 0); /* fake */
3202 while (low <= high) {
3203 i = (low + high) >> 1;
3204 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3205 rc = cmp(key, &nodekey);
3206 DPRINTF("found leaf index %u [%s], rc = %i",
3207 i, DKEY(&nodekey), rc);
3216 while (low <= high) {
3217 i = (low + high) >> 1;
3219 node = NODEPTR(mp, i);
3220 nodekey.mv_size = NODEKSZ(node);
3221 nodekey.mv_data = NODEKEY(node);
3223 rc = cmp(key, &nodekey);
3226 DPRINTF("found leaf index %u [%s], rc = %i",
3227 i, DKEY(&nodekey), rc);
3229 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3230 i, DKEY(&nodekey), NODEPGNO(node), rc);
3241 if (rc > 0) { /* Found entry is less than the key. */
3242 i++; /* Skip to get the smallest entry larger than key. */
3244 node = NODEPTR(mp, i);
3247 *exactp = (rc == 0);
3248 /* store the key index */
3249 mc->mc_ki[mc->mc_top] = i;
3251 /* There is no entry larger or equal to the key. */
3254 /* nodeptr is fake for LEAF2 */
3260 mdb_cursor_adjust(MDB_cursor *mc, func)
3264 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3265 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3272 /** Pop a page off the top of the cursor's stack. */
3274 mdb_cursor_pop(MDB_cursor *mc)
3279 top = mc->mc_pg[mc->mc_top];
3284 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3285 mc->mc_dbi, (void *) mc);
3289 /** Push a page onto the top of the cursor's stack. */
3291 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3293 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3294 mc->mc_dbi, (void *) mc);
3296 if (mc->mc_snum >= CURSOR_STACK) {
3297 assert(mc->mc_snum < CURSOR_STACK);
3301 mc->mc_top = mc->mc_snum++;
3302 mc->mc_pg[mc->mc_top] = mp;
3303 mc->mc_ki[mc->mc_top] = 0;
3308 /** Find the address of the page corresponding to a given page number.
3309 * @param[in] txn the transaction for this access.
3310 * @param[in] pgno the page number for the page to retrieve.
3311 * @param[out] ret address of a pointer where the page's address will be stored.
3312 * @return 0 on success, non-zero on failure.
3315 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3319 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3321 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3322 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3323 p = txn->mt_u.dirty_list[x].mptr;
3327 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
3328 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3332 DPRINTF("page %zu not found", pgno);
3335 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3338 /** Search for the page a given key should be in.
3339 * Pushes parent pages on the cursor stack. This function continues a
3340 * search on a cursor that has already been initialized. (Usually by
3341 * #mdb_page_search() but also by #mdb_node_move().)
3342 * @param[in,out] mc the cursor for this operation.
3343 * @param[in] key the key to search for. If NULL, search for the lowest
3344 * page. (This is used by #mdb_cursor_first().)
3345 * @param[in] modify If true, visited pages are updated with new page numbers.
3346 * @return 0 on success, non-zero on failure.
3349 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3351 MDB_page *mp = mc->mc_pg[mc->mc_top];
3356 while (IS_BRANCH(mp)) {
3360 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3361 assert(NUMKEYS(mp) > 1);
3362 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3364 if (key == NULL) /* Initialize cursor to first page. */
3366 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3367 /* cursor to last page */
3371 node = mdb_node_search(mc, key, &exact);
3373 i = NUMKEYS(mp) - 1;
3375 i = mc->mc_ki[mc->mc_top];
3384 DPRINTF("following index %u for key [%s]",
3386 assert(i < NUMKEYS(mp));
3387 node = NODEPTR(mp, i);
3389 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3392 mc->mc_ki[mc->mc_top] = i;
3393 if ((rc = mdb_cursor_push(mc, mp)))
3397 if ((rc = mdb_page_touch(mc)) != 0)
3399 mp = mc->mc_pg[mc->mc_top];
3404 DPRINTF("internal error, index points to a %02X page!?",
3406 return MDB_CORRUPTED;
3409 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3410 key ? DKEY(key) : NULL);
3415 /** Search for the page a given key should be in.
3416 * Pushes parent pages on the cursor stack. This function just sets up
3417 * the search; it finds the root page for \b mc's database and sets this
3418 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3419 * called to complete the search.
3420 * @param[in,out] mc the cursor for this operation.
3421 * @param[in] key the key to search for. If NULL, search for the lowest
3422 * page. (This is used by #mdb_cursor_first().)
3423 * @param[in] modify If true, visited pages are updated with new page numbers.
3424 * @return 0 on success, non-zero on failure.
3427 mdb_page_search(MDB_cursor *mc, MDB_val *key, int modify)
3432 /* Make sure the txn is still viable, then find the root from
3433 * the txn's db table.
3435 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3436 DPUTS("transaction has failed, must abort");
3439 /* Make sure we're using an up-to-date root */
3440 if (mc->mc_dbi > MAIN_DBI) {
3441 if ((*mc->mc_dbflag & DB_STALE) ||
3442 (modify && !(*mc->mc_dbflag & DB_DIRTY))) {
3444 unsigned char dbflag = 0;
3445 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3446 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, modify);
3449 if (*mc->mc_dbflag & DB_STALE) {
3452 MDB_node *leaf = mdb_node_search(&mc2,
3453 &mc->mc_dbx->md_name, &exact);
3455 return MDB_NOTFOUND;
3456 mdb_node_read(mc->mc_txn, leaf, &data);
3457 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3461 *mc->mc_dbflag = dbflag;
3464 root = mc->mc_db->md_root;
3466 if (root == P_INVALID) { /* Tree is empty. */
3467 DPUTS("tree is empty");
3468 return MDB_NOTFOUND;
3473 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3479 DPRINTF("db %u root page %zu has flags 0x%X",
3480 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3483 if ((rc = mdb_page_touch(mc)))
3487 return mdb_page_search_root(mc, key, modify);
3490 /** Return the data associated with a given node.
3491 * @param[in] txn The transaction for this operation.
3492 * @param[in] leaf The node being read.
3493 * @param[out] data Updated to point to the node's data.
3494 * @return 0 on success, non-zero on failure.
3497 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3499 MDB_page *omp; /* overflow page */
3503 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3504 data->mv_size = NODEDSZ(leaf);
3505 data->mv_data = NODEDATA(leaf);
3509 /* Read overflow data.
3511 data->mv_size = NODEDSZ(leaf);
3512 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3513 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3514 DPRINTF("read overflow page %zu failed", pgno);
3517 data->mv_data = METADATA(omp);
3523 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3524 MDB_val *key, MDB_val *data)
3533 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3535 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3538 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3542 mdb_cursor_init(&mc, txn, dbi, &mx);
3543 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3546 /** Find a sibling for a page.
3547 * Replaces the page at the top of the cursor's stack with the
3548 * specified sibling, if one exists.
3549 * @param[in] mc The cursor for this operation.
3550 * @param[in] move_right Non-zero if the right sibling is requested,
3551 * otherwise the left sibling.
3552 * @return 0 on success, non-zero on failure.
3555 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3561 if (mc->mc_snum < 2) {
3562 return MDB_NOTFOUND; /* root has no siblings */
3566 DPRINTF("parent page is page %zu, index %u",
3567 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3569 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3570 : (mc->mc_ki[mc->mc_top] == 0)) {
3571 DPRINTF("no more keys left, moving to %s sibling",
3572 move_right ? "right" : "left");
3573 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3577 mc->mc_ki[mc->mc_top]++;
3579 mc->mc_ki[mc->mc_top]--;
3580 DPRINTF("just moving to %s index key %u",
3581 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3583 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3585 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3586 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3589 mdb_cursor_push(mc, mp);
3594 /** Move the cursor to the next data item. */
3596 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3602 if (mc->mc_flags & C_EOF) {
3603 return MDB_NOTFOUND;
3606 assert(mc->mc_flags & C_INITIALIZED);
3608 mp = mc->mc_pg[mc->mc_top];
3610 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3611 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3612 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3613 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3614 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3615 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3619 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3620 if (op == MDB_NEXT_DUP)
3621 return MDB_NOTFOUND;
3625 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3627 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3628 DPUTS("=====> move to next sibling page");
3629 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3630 mc->mc_flags |= C_EOF;
3631 mc->mc_flags &= ~C_INITIALIZED;
3632 return MDB_NOTFOUND;
3634 mp = mc->mc_pg[mc->mc_top];
3635 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3637 mc->mc_ki[mc->mc_top]++;
3639 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3640 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3643 key->mv_size = mc->mc_db->md_pad;
3644 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3648 assert(IS_LEAF(mp));
3649 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3651 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3652 mdb_xcursor_init1(mc, leaf);
3655 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3658 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3659 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3660 if (rc != MDB_SUCCESS)
3665 MDB_SET_KEY(leaf, key);
3669 /** Move the cursor to the previous data item. */
3671 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3677 assert(mc->mc_flags & C_INITIALIZED);
3679 mp = mc->mc_pg[mc->mc_top];
3681 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3682 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3683 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3684 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3685 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3686 if (op != MDB_PREV || rc == MDB_SUCCESS)
3689 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3690 if (op == MDB_PREV_DUP)
3691 return MDB_NOTFOUND;
3696 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3698 if (mc->mc_ki[mc->mc_top] == 0) {
3699 DPUTS("=====> move to prev sibling page");
3700 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3701 mc->mc_flags &= ~C_INITIALIZED;
3702 return MDB_NOTFOUND;
3704 mp = mc->mc_pg[mc->mc_top];
3705 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3706 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3708 mc->mc_ki[mc->mc_top]--;
3710 mc->mc_flags &= ~C_EOF;
3712 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3713 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3716 key->mv_size = mc->mc_db->md_pad;
3717 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3721 assert(IS_LEAF(mp));
3722 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3724 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3725 mdb_xcursor_init1(mc, leaf);
3728 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3731 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3732 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3733 if (rc != MDB_SUCCESS)
3738 MDB_SET_KEY(leaf, key);
3742 /** Set the cursor on a specific data item. */
3744 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3745 MDB_cursor_op op, int *exactp)
3754 assert(key->mv_size > 0);
3756 /* See if we're already on the right page */
3757 if (mc->mc_flags & C_INITIALIZED) {
3760 mp = mc->mc_pg[mc->mc_top];
3762 mc->mc_ki[mc->mc_top] = 0;
3763 return MDB_NOTFOUND;
3765 if (mp->mp_flags & P_LEAF2) {
3766 nodekey.mv_size = mc->mc_db->md_pad;
3767 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
3769 leaf = NODEPTR(mp, 0);
3770 MDB_SET_KEY(leaf, &nodekey);
3772 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3774 /* Probably happens rarely, but first node on the page
3775 * was the one we wanted.
3777 mc->mc_ki[mc->mc_top] = 0;
3778 leaf = NODEPTR(mp, 0);
3785 unsigned int nkeys = NUMKEYS(mp);
3787 if (mp->mp_flags & P_LEAF2) {
3788 nodekey.mv_data = LEAF2KEY(mp,
3789 nkeys-1, nodekey.mv_size);
3791 leaf = NODEPTR(mp, nkeys-1);
3792 MDB_SET_KEY(leaf, &nodekey);
3794 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3796 /* last node was the one we wanted */
3797 mc->mc_ki[mc->mc_top] = nkeys-1;
3798 leaf = NODEPTR(mp, nkeys-1);
3804 /* This is definitely the right page, skip search_page */
3809 /* If any parents have right-sibs, search.
3810 * Otherwise, there's nothing further.
3812 for (i=0; i<mc->mc_top; i++)
3814 NUMKEYS(mc->mc_pg[i])-1)
3816 if (i == mc->mc_top) {
3817 /* There are no other pages */
3818 mc->mc_ki[mc->mc_top] = nkeys;
3819 return MDB_NOTFOUND;
3823 /* There are no other pages */
3824 mc->mc_ki[mc->mc_top] = 0;
3825 return MDB_NOTFOUND;
3829 rc = mdb_page_search(mc, key, 0);
3830 if (rc != MDB_SUCCESS)
3833 mp = mc->mc_pg[mc->mc_top];
3834 assert(IS_LEAF(mp));
3837 leaf = mdb_node_search(mc, key, exactp);
3838 if (exactp != NULL && !*exactp) {
3839 /* MDB_SET specified and not an exact match. */
3840 return MDB_NOTFOUND;
3844 DPUTS("===> inexact leaf not found, goto sibling");
3845 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
3846 return rc; /* no entries matched */
3847 mp = mc->mc_pg[mc->mc_top];
3848 assert(IS_LEAF(mp));
3849 leaf = NODEPTR(mp, 0);
3853 mc->mc_flags |= C_INITIALIZED;
3854 mc->mc_flags &= ~C_EOF;
3857 key->mv_size = mc->mc_db->md_pad;
3858 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3862 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3863 mdb_xcursor_init1(mc, leaf);
3866 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3867 if (op == MDB_SET || op == MDB_SET_RANGE) {
3868 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3871 if (op == MDB_GET_BOTH) {
3877 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
3878 if (rc != MDB_SUCCESS)
3881 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
3883 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
3885 rc = mc->mc_dbx->md_dcmp(data, &d2);
3887 if (op == MDB_GET_BOTH || rc > 0)
3888 return MDB_NOTFOUND;
3893 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3894 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3899 /* The key already matches in all other cases */
3900 if (op == MDB_SET_RANGE)
3901 MDB_SET_KEY(leaf, key);
3902 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3907 /** Move the cursor to the first item in the database. */
3909 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3914 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3915 rc = mdb_page_search(mc, NULL, 0);
3916 if (rc != MDB_SUCCESS)
3919 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3921 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3922 mc->mc_flags |= C_INITIALIZED;
3923 mc->mc_flags &= ~C_EOF;
3925 mc->mc_ki[mc->mc_top] = 0;
3927 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3928 key->mv_size = mc->mc_db->md_pad;
3929 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
3934 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3935 mdb_xcursor_init1(mc, leaf);
3936 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3941 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3942 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3946 MDB_SET_KEY(leaf, key);
3950 /** Move the cursor to the last item in the database. */
3952 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3958 lkey.mv_size = MAXKEYSIZE+1;
3959 lkey.mv_data = NULL;
3961 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3962 rc = mdb_page_search(mc, &lkey, 0);
3963 if (rc != MDB_SUCCESS)
3966 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3968 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
3969 mc->mc_flags |= C_INITIALIZED;
3970 mc->mc_flags &= ~C_EOF;
3972 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
3974 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3975 key->mv_size = mc->mc_db->md_pad;
3976 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
3981 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3982 mdb_xcursor_init1(mc, leaf);
3983 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3988 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3989 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3994 MDB_SET_KEY(leaf, key);
3999 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4009 case MDB_GET_BOTH_RANGE:
4010 if (data == NULL || mc->mc_xcursor == NULL) {
4017 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4019 } else if (op == MDB_SET_RANGE)
4020 rc = mdb_cursor_set(mc, key, data, op, NULL);
4022 rc = mdb_cursor_set(mc, key, data, op, &exact);
4024 case MDB_GET_MULTIPLE:
4026 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4027 !(mc->mc_flags & C_INITIALIZED)) {
4032 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4033 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4036 case MDB_NEXT_MULTIPLE:
4038 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4042 if (!(mc->mc_flags & C_INITIALIZED))
4043 rc = mdb_cursor_first(mc, key, data);
4045 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4046 if (rc == MDB_SUCCESS) {
4047 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4050 mx = &mc->mc_xcursor->mx_cursor;
4051 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4053 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4054 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4062 case MDB_NEXT_NODUP:
4063 if (!(mc->mc_flags & C_INITIALIZED))
4064 rc = mdb_cursor_first(mc, key, data);
4066 rc = mdb_cursor_next(mc, key, data, op);
4070 case MDB_PREV_NODUP:
4071 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
4072 rc = mdb_cursor_last(mc, key, data);
4074 rc = mdb_cursor_prev(mc, key, data, op);
4077 rc = mdb_cursor_first(mc, key, data);
4081 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4082 !(mc->mc_flags & C_INITIALIZED) ||
4083 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4087 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4090 rc = mdb_cursor_last(mc, key, data);
4094 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4095 !(mc->mc_flags & C_INITIALIZED) ||
4096 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4100 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4103 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4111 /** Touch all the pages in the cursor stack.
4112 * Makes sure all the pages are writable, before attempting a write operation.
4113 * @param[in] mc The cursor to operate on.
4116 mdb_cursor_touch(MDB_cursor *mc)
4120 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4122 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4123 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 1);
4126 *mc->mc_dbflag = DB_DIRTY;
4128 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4129 rc = mdb_page_touch(mc);
4133 mc->mc_top = mc->mc_snum-1;
4138 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4141 MDB_node *leaf = NULL;
4142 MDB_val xdata, *rdata, dkey;
4146 unsigned int mcount = 0;
4150 char dbuf[MAXKEYSIZE+1];
4151 unsigned int nflags;
4154 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4157 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4158 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4162 if (flags == MDB_CURRENT) {
4163 if (!(mc->mc_flags & C_INITIALIZED))
4166 } else if (mc->mc_db->md_root == P_INVALID) {
4168 /* new database, write a root leaf page */
4169 DPUTS("allocating new root leaf page");
4170 if ((np = mdb_page_new(mc, P_LEAF, 1)) == NULL) {
4174 mdb_cursor_push(mc, np);
4175 mc->mc_db->md_root = np->mp_pgno;
4176 mc->mc_db->md_depth++;
4177 *mc->mc_dbflag = DB_DIRTY;
4178 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4180 np->mp_flags |= P_LEAF2;
4181 mc->mc_flags |= C_INITIALIZED;
4187 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4188 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4189 DPRINTF("duplicate key [%s]", DKEY(key));
4191 return MDB_KEYEXIST;
4193 if (rc && rc != MDB_NOTFOUND)
4197 /* Cursor is positioned, now make sure all pages are writable */
4198 rc2 = mdb_cursor_touch(mc);
4203 /* The key already exists */
4204 if (rc == MDB_SUCCESS) {
4205 /* there's only a key anyway, so this is a no-op */
4206 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4207 unsigned int ksize = mc->mc_db->md_pad;
4208 if (key->mv_size != ksize)
4210 if (flags == MDB_CURRENT) {
4211 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4212 memcpy(ptr, key->mv_data, ksize);
4217 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4220 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4221 /* Was a single item before, must convert now */
4223 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4224 /* Just overwrite the current item */
4225 if (flags == MDB_CURRENT)
4228 dkey.mv_size = NODEDSZ(leaf);
4229 dkey.mv_data = NODEDATA(leaf);
4230 #if UINT_MAX < SIZE_MAX
4231 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4232 #ifdef MISALIGNED_OK
4233 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4235 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4238 /* if data matches, ignore it */
4239 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4240 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4242 /* create a fake page for the dup items */
4243 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4244 dkey.mv_data = dbuf;
4245 fp = (MDB_page *)&pbuf;
4246 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4247 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4248 fp->mp_lower = PAGEHDRSZ;
4249 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4250 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4251 fp->mp_flags |= P_LEAF2;
4252 fp->mp_pad = data->mv_size;
4254 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4255 (dkey.mv_size & 1) + (data->mv_size & 1);
4257 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4260 xdata.mv_size = fp->mp_upper;
4265 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4266 /* See if we need to convert from fake page to subDB */
4268 unsigned int offset;
4271 fp = NODEDATA(leaf);
4272 if (flags == MDB_CURRENT) {
4273 fp->mp_flags |= P_DIRTY;
4274 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4275 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4279 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4280 offset = fp->mp_pad;
4282 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4284 offset += offset & 1;
4285 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4286 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4288 /* yes, convert it */
4290 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4291 dummy.md_pad = fp->mp_pad;
4292 dummy.md_flags = MDB_DUPFIXED;
4293 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4294 dummy.md_flags |= MDB_INTEGERKEY;
4297 dummy.md_branch_pages = 0;
4298 dummy.md_leaf_pages = 1;
4299 dummy.md_overflow_pages = 0;
4300 dummy.md_entries = NUMKEYS(fp);
4302 xdata.mv_size = sizeof(MDB_db);
4303 xdata.mv_data = &dummy;
4304 mp = mdb_page_alloc(mc, 1);
4307 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4308 flags |= F_DUPDATA|F_SUBDATA;
4309 dummy.md_root = mp->mp_pgno;
4311 /* no, just grow it */
4313 xdata.mv_size = NODEDSZ(leaf) + offset;
4314 xdata.mv_data = &pbuf;
4315 mp = (MDB_page *)&pbuf;
4316 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4319 mp->mp_flags = fp->mp_flags | P_DIRTY;
4320 mp->mp_pad = fp->mp_pad;
4321 mp->mp_lower = fp->mp_lower;
4322 mp->mp_upper = fp->mp_upper + offset;
4324 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4326 nsize = NODEDSZ(leaf) - fp->mp_upper;
4327 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4328 for (i=0; i<NUMKEYS(fp); i++)
4329 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4331 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4335 /* data is on sub-DB, just store it */
4336 flags |= F_DUPDATA|F_SUBDATA;
4340 /* overflow page overwrites need special handling */
4341 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4344 int ovpages, dpages;
4346 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4347 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4348 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4349 mdb_page_get(mc->mc_txn, pg, &omp);
4350 /* Is the ov page writable and large enough? */
4351 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4352 /* yes, overwrite it. Note in this case we don't
4353 * bother to try shrinking the node if the new data
4354 * is smaller than the overflow threshold.
4356 if (F_ISSET(flags, MDB_RESERVE))
4357 data->mv_data = METADATA(omp);
4359 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4362 /* no, free ovpages */
4364 mc->mc_db->md_overflow_pages -= ovpages;
4365 for (i=0; i<ovpages; i++) {
4366 DPRINTF("freed ov page %zu", pg);
4367 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4371 } else if (NODEDSZ(leaf) == data->mv_size) {
4372 /* same size, just replace it. Note that we could
4373 * also reuse this node if the new data is smaller,
4374 * but instead we opt to shrink the node in that case.
4376 if (F_ISSET(flags, MDB_RESERVE))
4377 data->mv_data = NODEDATA(leaf);
4379 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4382 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4383 mc->mc_db->md_entries--;
4385 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4391 nflags = flags & NODE_ADD_FLAGS;
4392 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4393 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4394 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4395 nflags &= ~MDB_APPEND;
4396 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4398 /* There is room already in this leaf page. */
4399 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4400 if (rc == 0 && !do_sub) {
4401 /* Adjust other cursors pointing to mp */
4402 MDB_cursor *m2, *m3;
4403 MDB_dbi dbi = mc->mc_dbi;
4404 unsigned i = mc->mc_top;
4405 MDB_page *mp = mc->mc_pg[i];
4407 if (mc->mc_flags & C_SUB)
4410 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4411 if (mc->mc_flags & C_SUB)
4412 m3 = &m2->mc_xcursor->mx_cursor;
4415 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4416 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4423 if (rc != MDB_SUCCESS)
4424 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4426 /* Now store the actual data in the child DB. Note that we're
4427 * storing the user data in the keys field, so there are strict
4428 * size limits on dupdata. The actual data fields of the child
4429 * DB are all zero size.
4436 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4437 if (flags & MDB_CURRENT) {
4438 xflags = MDB_CURRENT;
4440 mdb_xcursor_init1(mc, leaf);
4441 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4443 /* converted, write the original data first */
4445 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4449 /* Adjust other cursors pointing to mp */
4451 unsigned i = mc->mc_top;
4452 MDB_page *mp = mc->mc_pg[i];
4454 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4455 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4456 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4457 mdb_xcursor_init1(m2, leaf);
4462 xflags |= (flags & MDB_APPEND);
4463 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4464 if (flags & F_SUBDATA) {
4465 void *db = NODEDATA(leaf);
4466 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4469 /* sub-writes might have failed so check rc again.
4470 * Don't increment count if we just replaced an existing item.
4472 if (!rc && !(flags & MDB_CURRENT))
4473 mc->mc_db->md_entries++;
4474 if (flags & MDB_MULTIPLE) {
4476 if (mcount < data[1].mv_size) {
4477 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4478 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4488 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4493 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4496 if (!mc->mc_flags & C_INITIALIZED)
4499 rc = mdb_cursor_touch(mc);
4503 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4505 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4506 if (flags != MDB_NODUPDATA) {
4507 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4508 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4510 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4511 /* If sub-DB still has entries, we're done */
4512 if (mc->mc_xcursor->mx_db.md_entries) {
4513 if (leaf->mn_flags & F_SUBDATA) {
4514 /* update subDB info */
4515 void *db = NODEDATA(leaf);
4516 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4518 /* shrink fake page */
4519 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4521 mc->mc_db->md_entries--;
4524 /* otherwise fall thru and delete the sub-DB */
4527 if (leaf->mn_flags & F_SUBDATA) {
4528 /* add all the child DB's pages to the free list */
4529 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4530 if (rc == MDB_SUCCESS) {
4531 mc->mc_db->md_entries -=
4532 mc->mc_xcursor->mx_db.md_entries;
4537 return mdb_cursor_del0(mc, leaf);
4540 /** Allocate and initialize new pages for a database.
4541 * @param[in] mc a cursor on the database being added to.
4542 * @param[in] flags flags defining what type of page is being allocated.
4543 * @param[in] num the number of pages to allocate. This is usually 1,
4544 * unless allocating overflow pages for a large record.
4545 * @return Address of a page, or NULL on failure.
4548 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num)
4552 if ((np = mdb_page_alloc(mc, num)) == NULL)
4554 DPRINTF("allocated new mpage %zu, page size %u",
4555 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4556 np->mp_flags = flags | P_DIRTY;
4557 np->mp_lower = PAGEHDRSZ;
4558 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4561 mc->mc_db->md_branch_pages++;
4562 else if (IS_LEAF(np))
4563 mc->mc_db->md_leaf_pages++;
4564 else if (IS_OVERFLOW(np)) {
4565 mc->mc_db->md_overflow_pages += num;
4572 /** Calculate the size of a leaf node.
4573 * The size depends on the environment's page size; if a data item
4574 * is too large it will be put onto an overflow page and the node
4575 * size will only include the key and not the data. Sizes are always
4576 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4577 * of the #MDB_node headers.
4578 * @param[in] env The environment handle.
4579 * @param[in] key The key for the node.
4580 * @param[in] data The data for the node.
4581 * @return The number of bytes needed to store the node.
4584 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4588 sz = LEAFSIZE(key, data);
4589 if (sz >= env->me_psize / MDB_MINKEYS) {
4590 /* put on overflow page */
4591 sz -= data->mv_size - sizeof(pgno_t);
4595 return sz + sizeof(indx_t);
4598 /** Calculate the size of a branch node.
4599 * The size should depend on the environment's page size but since
4600 * we currently don't support spilling large keys onto overflow
4601 * pages, it's simply the size of the #MDB_node header plus the
4602 * size of the key. Sizes are always rounded up to an even number
4603 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4604 * @param[in] env The environment handle.
4605 * @param[in] key The key for the node.
4606 * @return The number of bytes needed to store the node.
4609 mdb_branch_size(MDB_env *env, MDB_val *key)
4614 if (sz >= env->me_psize / MDB_MINKEYS) {
4615 /* put on overflow page */
4616 /* not implemented */
4617 /* sz -= key->size - sizeof(pgno_t); */
4620 return sz + sizeof(indx_t);
4623 /** Add a node to the page pointed to by the cursor.
4624 * @param[in] mc The cursor for this operation.
4625 * @param[in] indx The index on the page where the new node should be added.
4626 * @param[in] key The key for the new node.
4627 * @param[in] data The data for the new node, if any.
4628 * @param[in] pgno The page number, if adding a branch node.
4629 * @param[in] flags Flags for the node.
4630 * @return 0 on success, non-zero on failure. Possible errors are:
4632 * <li>ENOMEM - failed to allocate overflow pages for the node.
4633 * <li>ENOSPC - there is insufficient room in the page. This error
4634 * should never happen since all callers already calculate the
4635 * page's free space before calling this function.
4639 mdb_node_add(MDB_cursor *mc, indx_t indx,
4640 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4643 size_t node_size = NODESIZE;
4646 MDB_page *mp = mc->mc_pg[mc->mc_top];
4647 MDB_page *ofp = NULL; /* overflow page */
4650 assert(mp->mp_upper >= mp->mp_lower);
4652 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4653 IS_LEAF(mp) ? "leaf" : "branch",
4654 IS_SUBP(mp) ? "sub-" : "",
4655 mp->mp_pgno, indx, data ? data->mv_size : 0,
4656 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4659 /* Move higher keys up one slot. */
4660 int ksize = mc->mc_db->md_pad, dif;
4661 char *ptr = LEAF2KEY(mp, indx, ksize);
4662 dif = NUMKEYS(mp) - indx;
4664 memmove(ptr+ksize, ptr, dif*ksize);
4665 /* insert new key */
4666 memcpy(ptr, key->mv_data, ksize);
4668 /* Just using these for counting */
4669 mp->mp_lower += sizeof(indx_t);
4670 mp->mp_upper -= ksize - sizeof(indx_t);
4675 node_size += key->mv_size;
4679 if (F_ISSET(flags, F_BIGDATA)) {
4680 /* Data already on overflow page. */
4681 node_size += sizeof(pgno_t);
4682 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4683 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4684 /* Put data on overflow page. */
4685 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4686 data->mv_size, node_size+data->mv_size);
4687 node_size += sizeof(pgno_t);
4688 if ((ofp = mdb_page_new(mc, P_OVERFLOW, ovpages)) == NULL)
4690 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4693 node_size += data->mv_size;
4696 node_size += node_size & 1;
4698 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4699 DPRINTF("not enough room in page %zu, got %u ptrs",
4700 mp->mp_pgno, NUMKEYS(mp));
4701 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4702 mp->mp_upper - mp->mp_lower);
4703 DPRINTF("node size = %zu", node_size);
4707 /* Move higher pointers up one slot. */
4708 for (i = NUMKEYS(mp); i > indx; i--)
4709 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
4711 /* Adjust free space offsets. */
4712 ofs = mp->mp_upper - node_size;
4713 assert(ofs >= mp->mp_lower + sizeof(indx_t));
4714 mp->mp_ptrs[indx] = ofs;
4716 mp->mp_lower += sizeof(indx_t);
4718 /* Write the node data. */
4719 node = NODEPTR(mp, indx);
4720 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
4721 node->mn_flags = flags;
4723 SETDSZ(node,data->mv_size);
4728 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4733 if (F_ISSET(flags, F_BIGDATA))
4734 memcpy(node->mn_data + key->mv_size, data->mv_data,
4736 else if (F_ISSET(flags, MDB_RESERVE))
4737 data->mv_data = node->mn_data + key->mv_size;
4739 memcpy(node->mn_data + key->mv_size, data->mv_data,
4742 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
4744 if (F_ISSET(flags, MDB_RESERVE))
4745 data->mv_data = METADATA(ofp);
4747 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
4754 /** Delete the specified node from a page.
4755 * @param[in] mp The page to operate on.
4756 * @param[in] indx The index of the node to delete.
4757 * @param[in] ksize The size of a node. Only used if the page is
4758 * part of a #MDB_DUPFIXED database.
4761 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
4764 indx_t i, j, numkeys, ptr;
4771 COPY_PGNO(pgno, mp->mp_pgno);
4772 DPRINTF("delete node %u on %s page %zu", indx,
4773 IS_LEAF(mp) ? "leaf" : "branch", pgno);
4776 assert(indx < NUMKEYS(mp));
4779 int x = NUMKEYS(mp) - 1 - indx;
4780 base = LEAF2KEY(mp, indx, ksize);
4782 memmove(base, base + ksize, x * ksize);
4783 mp->mp_lower -= sizeof(indx_t);
4784 mp->mp_upper += ksize - sizeof(indx_t);
4788 node = NODEPTR(mp, indx);
4789 sz = NODESIZE + node->mn_ksize;
4791 if (F_ISSET(node->mn_flags, F_BIGDATA))
4792 sz += sizeof(pgno_t);
4794 sz += NODEDSZ(node);
4798 ptr = mp->mp_ptrs[indx];
4799 numkeys = NUMKEYS(mp);
4800 for (i = j = 0; i < numkeys; i++) {
4802 mp->mp_ptrs[j] = mp->mp_ptrs[i];
4803 if (mp->mp_ptrs[i] < ptr)
4804 mp->mp_ptrs[j] += sz;
4809 base = (char *)mp + mp->mp_upper;
4810 memmove(base + sz, base, ptr - mp->mp_upper);
4812 mp->mp_lower -= sizeof(indx_t);
4816 /** Compact the main page after deleting a node on a subpage.
4817 * @param[in] mp The main page to operate on.
4818 * @param[in] indx The index of the subpage on the main page.
4821 mdb_node_shrink(MDB_page *mp, indx_t indx)
4828 indx_t i, numkeys, ptr;
4830 node = NODEPTR(mp, indx);
4831 sp = (MDB_page *)NODEDATA(node);
4832 osize = NODEDSZ(node);
4834 delta = sp->mp_upper - sp->mp_lower;
4835 SETDSZ(node, osize - delta);
4836 xp = (MDB_page *)((char *)sp + delta);
4838 /* shift subpage upward */
4840 nsize = NUMKEYS(sp) * sp->mp_pad;
4841 memmove(METADATA(xp), METADATA(sp), nsize);
4844 nsize = osize - sp->mp_upper;
4845 numkeys = NUMKEYS(sp);
4846 for (i=numkeys-1; i>=0; i--)
4847 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
4849 xp->mp_upper = sp->mp_lower;
4850 xp->mp_lower = sp->mp_lower;
4851 xp->mp_flags = sp->mp_flags;
4852 xp->mp_pad = sp->mp_pad;
4853 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
4855 /* shift lower nodes upward */
4856 ptr = mp->mp_ptrs[indx];
4857 numkeys = NUMKEYS(mp);
4858 for (i = 0; i < numkeys; i++) {
4859 if (mp->mp_ptrs[i] <= ptr)
4860 mp->mp_ptrs[i] += delta;
4863 base = (char *)mp + mp->mp_upper;
4864 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
4865 mp->mp_upper += delta;
4868 /** Initial setup of a sorted-dups cursor.
4869 * Sorted duplicates are implemented as a sub-database for the given key.
4870 * The duplicate data items are actually keys of the sub-database.
4871 * Operations on the duplicate data items are performed using a sub-cursor
4872 * initialized when the sub-database is first accessed. This function does
4873 * the preliminary setup of the sub-cursor, filling in the fields that
4874 * depend only on the parent DB.
4875 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4878 mdb_xcursor_init0(MDB_cursor *mc)
4880 MDB_xcursor *mx = mc->mc_xcursor;
4882 mx->mx_cursor.mc_xcursor = NULL;
4883 mx->mx_cursor.mc_txn = mc->mc_txn;
4884 mx->mx_cursor.mc_db = &mx->mx_db;
4885 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
4886 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
4887 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
4888 mx->mx_cursor.mc_snum = 0;
4889 mx->mx_cursor.mc_top = 0;
4890 mx->mx_cursor.mc_flags = C_SUB;
4891 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
4892 mx->mx_dbx.md_dcmp = NULL;
4893 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
4896 /** Final setup of a sorted-dups cursor.
4897 * Sets up the fields that depend on the data from the main cursor.
4898 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4899 * @param[in] node The data containing the #MDB_db record for the
4900 * sorted-dup database.
4903 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
4905 MDB_xcursor *mx = mc->mc_xcursor;
4907 if (node->mn_flags & F_SUBDATA) {
4908 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
4909 mx->mx_cursor.mc_snum = 0;
4910 mx->mx_cursor.mc_flags = C_SUB;
4912 MDB_page *fp = NODEDATA(node);
4913 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
4914 mx->mx_db.md_flags = 0;
4915 mx->mx_db.md_depth = 1;
4916 mx->mx_db.md_branch_pages = 0;
4917 mx->mx_db.md_leaf_pages = 1;
4918 mx->mx_db.md_overflow_pages = 0;
4919 mx->mx_db.md_entries = NUMKEYS(fp);
4920 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
4921 mx->mx_cursor.mc_snum = 1;
4922 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
4923 mx->mx_cursor.mc_top = 0;
4924 mx->mx_cursor.mc_pg[0] = fp;
4925 mx->mx_cursor.mc_ki[0] = 0;
4926 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4927 mx->mx_db.md_flags = MDB_DUPFIXED;
4928 mx->mx_db.md_pad = fp->mp_pad;
4929 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4930 mx->mx_db.md_flags |= MDB_INTEGERKEY;
4933 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
4935 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
4937 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
4938 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
4939 #if UINT_MAX < SIZE_MAX
4940 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
4941 #ifdef MISALIGNED_OK
4942 mx->mx_dbx.md_cmp = mdb_cmp_long;
4944 mx->mx_dbx.md_cmp = mdb_cmp_cint;
4949 /** Initialize a cursor for a given transaction and database. */
4951 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
4956 mc->mc_db = &txn->mt_dbs[dbi];
4957 mc->mc_dbx = &txn->mt_dbxs[dbi];
4958 mc->mc_dbflag = &txn->mt_dbflags[dbi];
4962 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4964 mc->mc_xcursor = mx;
4965 mdb_xcursor_init0(mc);
4967 mc->mc_xcursor = NULL;
4972 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
4975 MDB_xcursor *mx = NULL;
4976 size_t size = sizeof(MDB_cursor);
4978 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
4981 /* Allow read access to the freelist */
4982 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
4985 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
4986 size += sizeof(MDB_xcursor);
4988 if ((mc = malloc(size)) != NULL) {
4989 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4990 mx = (MDB_xcursor *)(mc + 1);
4992 mdb_cursor_init(mc, txn, dbi, mx);
4993 if (txn->mt_cursors) {
4994 mc->mc_next = txn->mt_cursors[dbi];
4995 txn->mt_cursors[dbi] = mc;
4997 mc->mc_flags |= C_ALLOCD;
5007 /* Return the count of duplicate data items for the current key */
5009 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5013 if (mc == NULL || countp == NULL)
5016 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5019 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5020 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5023 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5026 *countp = mc->mc_xcursor->mx_db.md_entries;
5032 mdb_cursor_close(MDB_cursor *mc)
5035 /* remove from txn, if tracked */
5036 if (mc->mc_txn->mt_cursors) {
5037 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5038 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5040 *prev = mc->mc_next;
5042 if (mc->mc_flags & C_ALLOCD)
5048 mdb_cursor_txn(MDB_cursor *mc)
5050 if (!mc) return NULL;
5055 mdb_cursor_dbi(MDB_cursor *mc)
5061 /** Replace the key for a node with a new key.
5062 * @param[in] mp The page containing the node to operate on.
5063 * @param[in] indx The index of the node to operate on.
5064 * @param[in] key The new key to use.
5065 * @return 0 on success, non-zero on failure.
5068 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5074 indx_t ptr, i, numkeys;
5077 node = NODEPTR(mp, indx);
5078 ptr = mp->mp_ptrs[indx];
5082 char kbuf2[(MAXKEYSIZE*2+1)];
5083 k2.mv_data = NODEKEY(node);
5084 k2.mv_size = node->mn_ksize;
5085 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5087 mdb_dkey(&k2, kbuf2),
5093 delta0 = delta = key->mv_size - node->mn_ksize;
5095 /* Must be 2-byte aligned. If new key is
5096 * shorter by 1, the shift will be skipped.
5098 delta += (delta & 1);
5100 if (delta > 0 && SIZELEFT(mp) < delta) {
5101 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5105 numkeys = NUMKEYS(mp);
5106 for (i = 0; i < numkeys; i++) {
5107 if (mp->mp_ptrs[i] <= ptr)
5108 mp->mp_ptrs[i] -= delta;
5111 base = (char *)mp + mp->mp_upper;
5112 len = ptr - mp->mp_upper + NODESIZE;
5113 memmove(base - delta, base, len);
5114 mp->mp_upper -= delta;
5116 node = NODEPTR(mp, indx);
5119 /* But even if no shift was needed, update ksize */
5121 node->mn_ksize = key->mv_size;
5124 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5129 /** Move a node from csrc to cdst.
5132 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5138 unsigned short flags;
5142 /* Mark src and dst as dirty. */
5143 if ((rc = mdb_page_touch(csrc)) ||
5144 (rc = mdb_page_touch(cdst)))
5147 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5148 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5149 key.mv_size = csrc->mc_db->md_pad;
5150 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5152 data.mv_data = NULL;
5156 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5157 assert(!((long)srcnode&1));
5158 srcpg = NODEPGNO(srcnode);
5159 flags = srcnode->mn_flags;
5160 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5161 unsigned int snum = csrc->mc_snum;
5163 /* must find the lowest key below src */
5164 mdb_page_search_root(csrc, NULL, 0);
5165 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5166 key.mv_size = NODEKSZ(s2);
5167 key.mv_data = NODEKEY(s2);
5168 csrc->mc_snum = snum--;
5169 csrc->mc_top = snum;
5171 key.mv_size = NODEKSZ(srcnode);
5172 key.mv_data = NODEKEY(srcnode);
5174 data.mv_size = NODEDSZ(srcnode);
5175 data.mv_data = NODEDATA(srcnode);
5177 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5178 unsigned int snum = cdst->mc_snum;
5181 /* must find the lowest key below dst */
5182 mdb_page_search_root(cdst, NULL, 0);
5183 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5184 bkey.mv_size = NODEKSZ(s2);
5185 bkey.mv_data = NODEKEY(s2);
5186 cdst->mc_snum = snum--;
5187 cdst->mc_top = snum;
5188 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5191 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5192 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5193 csrc->mc_ki[csrc->mc_top],
5195 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5196 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5198 /* Add the node to the destination page.
5200 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5201 if (rc != MDB_SUCCESS)
5204 /* Delete the node from the source page.
5206 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5209 /* Adjust other cursors pointing to mp */
5210 MDB_cursor *m2, *m3;
5211 MDB_dbi dbi = csrc->mc_dbi;
5212 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5214 if (csrc->mc_flags & C_SUB)
5217 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5218 if (m2 == csrc) continue;
5219 if (csrc->mc_flags & C_SUB)
5220 m3 = &m2->mc_xcursor->mx_cursor;
5223 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5224 csrc->mc_ki[csrc->mc_top]) {
5225 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5226 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5231 /* Update the parent separators.
5233 if (csrc->mc_ki[csrc->mc_top] == 0) {
5234 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5235 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5236 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5238 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5239 key.mv_size = NODEKSZ(srcnode);
5240 key.mv_data = NODEKEY(srcnode);
5242 DPRINTF("update separator for source page %zu to [%s]",
5243 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5244 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5245 &key)) != MDB_SUCCESS)
5248 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5250 nullkey.mv_size = 0;
5251 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5252 assert(rc == MDB_SUCCESS);
5256 if (cdst->mc_ki[cdst->mc_top] == 0) {
5257 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5258 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5259 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5261 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5262 key.mv_size = NODEKSZ(srcnode);
5263 key.mv_data = NODEKEY(srcnode);
5265 DPRINTF("update separator for destination page %zu to [%s]",
5266 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5267 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5268 &key)) != MDB_SUCCESS)
5271 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5273 nullkey.mv_size = 0;
5274 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5275 assert(rc == MDB_SUCCESS);
5282 /** Merge one page into another.
5283 * The nodes from the page pointed to by \b csrc will
5284 * be copied to the page pointed to by \b cdst and then
5285 * the \b csrc page will be freed.
5286 * @param[in] csrc Cursor pointing to the source page.
5287 * @param[in] cdst Cursor pointing to the destination page.
5290 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5298 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5299 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5301 assert(csrc->mc_snum > 1); /* can't merge root page */
5302 assert(cdst->mc_snum > 1);
5304 /* Mark dst as dirty. */
5305 if ((rc = mdb_page_touch(cdst)))
5308 /* Move all nodes from src to dst.
5310 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5311 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5312 key.mv_size = csrc->mc_db->md_pad;
5313 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5314 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5315 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5316 if (rc != MDB_SUCCESS)
5318 key.mv_data = (char *)key.mv_data + key.mv_size;
5321 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5322 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5323 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5324 unsigned int snum = csrc->mc_snum;
5326 /* must find the lowest key below src */
5327 mdb_page_search_root(csrc, NULL, 0);
5328 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5329 key.mv_size = NODEKSZ(s2);
5330 key.mv_data = NODEKEY(s2);
5331 csrc->mc_snum = snum--;
5332 csrc->mc_top = snum;
5334 key.mv_size = srcnode->mn_ksize;
5335 key.mv_data = NODEKEY(srcnode);
5338 data.mv_size = NODEDSZ(srcnode);
5339 data.mv_data = NODEDATA(srcnode);
5340 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5341 if (rc != MDB_SUCCESS)
5346 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5347 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);
5349 /* Unlink the src page from parent and add to free list.
5351 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5352 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5354 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5358 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5359 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5360 csrc->mc_db->md_leaf_pages--;
5362 csrc->mc_db->md_branch_pages--;
5364 /* Adjust other cursors pointing to mp */
5365 MDB_cursor *m2, *m3;
5366 MDB_dbi dbi = csrc->mc_dbi;
5367 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5369 if (csrc->mc_flags & C_SUB)
5372 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5373 if (csrc->mc_flags & C_SUB)
5374 m3 = &m2->mc_xcursor->mx_cursor;
5377 if (m3 == csrc) continue;
5378 if (m3->mc_snum < csrc->mc_snum) continue;
5379 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5380 m3->mc_pg[csrc->mc_top] = mp;
5381 m3->mc_ki[csrc->mc_top] += nkeys;
5385 mdb_cursor_pop(csrc);
5387 return mdb_rebalance(csrc);
5390 /** Copy the contents of a cursor.
5391 * @param[in] csrc The cursor to copy from.
5392 * @param[out] cdst The cursor to copy to.
5395 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5399 cdst->mc_txn = csrc->mc_txn;
5400 cdst->mc_dbi = csrc->mc_dbi;
5401 cdst->mc_db = csrc->mc_db;
5402 cdst->mc_dbx = csrc->mc_dbx;
5403 cdst->mc_snum = csrc->mc_snum;
5404 cdst->mc_top = csrc->mc_top;
5405 cdst->mc_flags = csrc->mc_flags;
5407 for (i=0; i<csrc->mc_snum; i++) {
5408 cdst->mc_pg[i] = csrc->mc_pg[i];
5409 cdst->mc_ki[i] = csrc->mc_ki[i];
5413 /** Rebalance the tree after a delete operation.
5414 * @param[in] mc Cursor pointing to the page where rebalancing
5416 * @return 0 on success, non-zero on failure.
5419 mdb_rebalance(MDB_cursor *mc)
5429 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5430 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5431 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5432 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5436 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5439 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5440 DPRINTF("no need to rebalance page %zu, above fill threshold",
5446 if (mc->mc_snum < 2) {
5447 MDB_page *mp = mc->mc_pg[0];
5448 if (NUMKEYS(mp) == 0) {
5449 DPUTS("tree is completely empty");
5450 mc->mc_db->md_root = P_INVALID;
5451 mc->mc_db->md_depth = 0;
5452 mc->mc_db->md_leaf_pages = 0;
5453 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5457 /* Adjust other cursors pointing to mp */
5458 MDB_cursor *m2, *m3;
5459 MDB_dbi dbi = mc->mc_dbi;
5461 if (mc->mc_flags & C_SUB)
5464 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5465 if (m2 == mc) continue;
5466 if (mc->mc_flags & C_SUB)
5467 m3 = &m2->mc_xcursor->mx_cursor;
5470 if (m3->mc_snum < mc->mc_snum) continue;
5471 if (m3->mc_pg[0] == mp) {
5477 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5478 DPUTS("collapsing root page!");
5479 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5480 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5481 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5484 mc->mc_db->md_depth--;
5485 mc->mc_db->md_branch_pages--;
5487 /* Adjust other cursors pointing to mp */
5488 MDB_cursor *m2, *m3;
5489 MDB_dbi dbi = mc->mc_dbi;
5491 if (mc->mc_flags & C_SUB)
5494 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5495 if (m2 == mc) continue;
5496 if (mc->mc_flags & C_SUB)
5497 m3 = &m2->mc_xcursor->mx_cursor;
5500 if (m3->mc_snum < mc->mc_snum) continue;
5501 if (m3->mc_pg[0] == mp) {
5502 m3->mc_pg[0] = mc->mc_pg[0];
5507 DPUTS("root page doesn't need rebalancing");
5511 /* The parent (branch page) must have at least 2 pointers,
5512 * otherwise the tree is invalid.
5514 ptop = mc->mc_top-1;
5515 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5517 /* Leaf page fill factor is below the threshold.
5518 * Try to move keys from left or right neighbor, or
5519 * merge with a neighbor page.
5524 mdb_cursor_copy(mc, &mn);
5525 mn.mc_xcursor = NULL;
5527 if (mc->mc_ki[ptop] == 0) {
5528 /* We're the leftmost leaf in our parent.
5530 DPUTS("reading right neighbor");
5532 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5533 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5535 mn.mc_ki[mn.mc_top] = 0;
5536 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5538 /* There is at least one neighbor to the left.
5540 DPUTS("reading left neighbor");
5542 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5543 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5545 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5546 mc->mc_ki[mc->mc_top] = 0;
5549 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5550 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);
5552 /* If the neighbor page is above threshold and has at least two
5553 * keys, move one key from it.
5555 * Otherwise we should try to merge them.
5557 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5558 return mdb_node_move(&mn, mc);
5559 else { /* FIXME: if (has_enough_room()) */
5560 mc->mc_flags &= ~C_INITIALIZED;
5561 if (mc->mc_ki[ptop] == 0)
5562 return mdb_page_merge(&mn, mc);
5564 return mdb_page_merge(mc, &mn);
5568 /** Complete a delete operation started by #mdb_cursor_del(). */
5570 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5574 /* add overflow pages to free list */
5575 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5579 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5580 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5581 mc->mc_db->md_overflow_pages -= ovpages;
5582 for (i=0; i<ovpages; i++) {
5583 DPRINTF("freed ov page %zu", pg);
5584 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5588 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5589 mc->mc_db->md_entries--;
5590 rc = mdb_rebalance(mc);
5591 if (rc != MDB_SUCCESS)
5592 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5598 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5599 MDB_val *key, MDB_val *data)
5604 MDB_val rdata, *xdata;
5608 assert(key != NULL);
5610 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5612 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5615 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5619 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5623 mdb_cursor_init(&mc, txn, dbi, &mx);
5634 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5636 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5640 /** Split a page and insert a new node.
5641 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5642 * The cursor will be updated to point to the actual page and index where
5643 * the node got inserted after the split.
5644 * @param[in] newkey The key for the newly inserted node.
5645 * @param[in] newdata The data for the newly inserted node.
5646 * @param[in] newpgno The page number, if the new node is a branch node.
5647 * @return 0 on success, non-zero on failure.
5650 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5651 unsigned int nflags)
5654 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1;
5657 unsigned int i, j, split_indx, nkeys, pmax;
5659 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5661 MDB_page *mp, *rp, *pp;
5666 mp = mc->mc_pg[mc->mc_top];
5667 newindx = mc->mc_ki[mc->mc_top];
5669 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
5670 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
5671 DKEY(newkey), mc->mc_ki[mc->mc_top]);
5673 /* Create a right sibling. */
5674 if ((rp = mdb_page_new(mc, mp->mp_flags, 1)) == NULL)
5676 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
5678 if (mc->mc_snum < 2) {
5679 if ((pp = mdb_page_new(mc, P_BRANCH, 1)) == NULL)
5681 /* shift current top to make room for new parent */
5682 mc->mc_pg[1] = mc->mc_pg[0];
5683 mc->mc_ki[1] = mc->mc_ki[0];
5686 mc->mc_db->md_root = pp->mp_pgno;
5687 DPRINTF("root split! new root = %zu", pp->mp_pgno);
5688 mc->mc_db->md_depth++;
5691 /* Add left (implicit) pointer. */
5692 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
5693 /* undo the pre-push */
5694 mc->mc_pg[0] = mc->mc_pg[1];
5695 mc->mc_ki[0] = mc->mc_ki[1];
5696 mc->mc_db->md_root = mp->mp_pgno;
5697 mc->mc_db->md_depth--;
5704 ptop = mc->mc_top-1;
5705 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
5708 mdb_cursor_copy(mc, &mn);
5709 mn.mc_pg[mn.mc_top] = rp;
5710 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
5712 if (nflags & MDB_APPEND) {
5713 mn.mc_ki[mn.mc_top] = 0;
5720 nkeys = NUMKEYS(mp);
5721 split_indx = (nkeys + 1) / 2;
5726 unsigned int lsize, rsize, ksize;
5727 /* Move half of the keys to the right sibling */
5729 x = mc->mc_ki[mc->mc_top] - split_indx;
5730 ksize = mc->mc_db->md_pad;
5731 split = LEAF2KEY(mp, split_indx, ksize);
5732 rsize = (nkeys - split_indx) * ksize;
5733 lsize = (nkeys - split_indx) * sizeof(indx_t);
5734 mp->mp_lower -= lsize;
5735 rp->mp_lower += lsize;
5736 mp->mp_upper += rsize - lsize;
5737 rp->mp_upper -= rsize - lsize;
5738 sepkey.mv_size = ksize;
5739 if (newindx == split_indx) {
5740 sepkey.mv_data = newkey->mv_data;
5742 sepkey.mv_data = split;
5745 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
5746 memcpy(rp->mp_ptrs, split, rsize);
5747 sepkey.mv_data = rp->mp_ptrs;
5748 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
5749 memcpy(ins, newkey->mv_data, ksize);
5750 mp->mp_lower += sizeof(indx_t);
5751 mp->mp_upper -= ksize - sizeof(indx_t);
5754 memcpy(rp->mp_ptrs, split, x * ksize);
5755 ins = LEAF2KEY(rp, x, ksize);
5756 memcpy(ins, newkey->mv_data, ksize);
5757 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
5758 rp->mp_lower += sizeof(indx_t);
5759 rp->mp_upper -= ksize - sizeof(indx_t);
5760 mc->mc_ki[mc->mc_top] = x;
5761 mc->mc_pg[mc->mc_top] = rp;
5766 /* For leaf pages, check the split point based on what
5767 * fits where, since otherwise mdb_node_add can fail.
5769 * This check is only needed when the data items are
5770 * relatively large, such that being off by one will
5771 * make the difference between success or failure.
5772 * When the size of the data items is much smaller than
5773 * one-half of a page, this check is irrelevant.
5776 unsigned int psize, nsize;
5777 /* Maximum free space in an empty page */
5778 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
5779 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
5780 if ((nkeys < 20) || (nsize > pmax/4)) {
5781 if (newindx <= split_indx) {
5784 for (i=0; i<split_indx; i++) {
5785 node = NODEPTR(mp, i);
5786 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5787 if (F_ISSET(node->mn_flags, F_BIGDATA))
5788 psize += sizeof(pgno_t);
5790 psize += NODEDSZ(node);
5793 if (i == split_indx - 1 && newindx == split_indx)
5802 for (i=nkeys-1; i>=split_indx; i--) {
5803 node = NODEPTR(mp, i);
5804 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5805 if (F_ISSET(node->mn_flags, F_BIGDATA))
5806 psize += sizeof(pgno_t);
5808 psize += NODEDSZ(node);
5819 /* First find the separating key between the split pages.
5820 * The case where newindx == split_indx is ambiguous; the
5821 * new item could go to the new page or stay on the original
5822 * page. If newpos == 1 it goes to the new page.
5824 if (newindx == split_indx && newpos) {
5825 sepkey.mv_size = newkey->mv_size;
5826 sepkey.mv_data = newkey->mv_data;
5828 node = NODEPTR(mp, split_indx);
5829 sepkey.mv_size = node->mn_ksize;
5830 sepkey.mv_data = NODEKEY(node);
5834 DPRINTF("separator is [%s]", DKEY(&sepkey));
5836 /* Copy separator key to the parent.
5838 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
5841 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
5843 /* Right page might now have changed parent.
5844 * Check if left page also changed parent.
5846 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
5847 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
5848 mc->mc_pg[ptop] = mn.mc_pg[ptop];
5849 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
5853 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
5856 if (rc != MDB_SUCCESS) {
5859 if (nflags & MDB_APPEND) {
5860 mc->mc_pg[mc->mc_top] = rp;
5861 mc->mc_ki[mc->mc_top] = 0;
5862 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
5871 /* Move half of the keys to the right sibling. */
5873 /* grab a page to hold a temporary copy */
5874 copy = mdb_page_malloc(mc);
5878 copy->mp_pgno = mp->mp_pgno;
5879 copy->mp_flags = mp->mp_flags;
5880 copy->mp_lower = PAGEHDRSZ;
5881 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
5882 mc->mc_pg[mc->mc_top] = copy;
5883 for (i = j = 0; i <= nkeys; j++) {
5884 if (i == split_indx) {
5885 /* Insert in right sibling. */
5886 /* Reset insert index for right sibling. */
5887 if (i != newindx || (newpos ^ ins_new)) {
5889 mc->mc_pg[mc->mc_top] = rp;
5893 if (i == newindx && !ins_new) {
5894 /* Insert the original entry that caused the split. */
5895 rkey.mv_data = newkey->mv_data;
5896 rkey.mv_size = newkey->mv_size;
5905 /* Update index for the new key. */
5906 mc->mc_ki[mc->mc_top] = j;
5907 } else if (i == nkeys) {
5910 node = NODEPTR(mp, i);
5911 rkey.mv_data = NODEKEY(node);
5912 rkey.mv_size = node->mn_ksize;
5914 xdata.mv_data = NODEDATA(node);
5915 xdata.mv_size = NODEDSZ(node);
5918 pgno = NODEPGNO(node);
5919 flags = node->mn_flags;
5924 if (!IS_LEAF(mp) && j == 0) {
5925 /* First branch index doesn't need key data. */
5929 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
5933 nkeys = NUMKEYS(copy);
5934 for (i=0; i<nkeys; i++)
5935 mp->mp_ptrs[i] = copy->mp_ptrs[i];
5936 mp->mp_lower = copy->mp_lower;
5937 mp->mp_upper = copy->mp_upper;
5938 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
5939 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
5941 /* reset back to original page */
5942 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
5943 mc->mc_pg[mc->mc_top] = mp;
5944 if (nflags & MDB_RESERVE) {
5945 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5946 if (!(node->mn_flags & F_BIGDATA))
5947 newdata->mv_data = NODEDATA(node);
5951 /* return tmp page to freelist */
5952 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
5953 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
5954 mc->mc_txn->mt_env->me_dpages = copy;
5957 /* Adjust other cursors pointing to mp */
5958 MDB_cursor *m2, *m3;
5959 MDB_dbi dbi = mc->mc_dbi;
5961 if (mc->mc_flags & C_SUB)
5964 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5965 if (m2 == mc) continue;
5966 if (mc->mc_flags & C_SUB)
5967 m3 = &m2->mc_xcursor->mx_cursor;
5970 if (!(m3->mc_flags & C_INITIALIZED))
5975 for (k=m3->mc_top; k>=0; k--) {
5976 m3->mc_ki[k+1] = m3->mc_ki[k];
5977 m3->mc_pg[k+1] = m3->mc_pg[k];
5979 m3->mc_ki[0] = mc->mc_ki[0];
5980 m3->mc_pg[0] = mc->mc_pg[0];
5984 if (m3->mc_pg[mc->mc_top] == mp) {
5985 if (m3->mc_ki[m3->mc_top] >= split_indx) {
5986 m3->mc_pg[m3->mc_top] = rp;
5987 m3->mc_ki[m3->mc_top] -= split_indx;
5996 mdb_put(MDB_txn *txn, MDB_dbi dbi,
5997 MDB_val *key, MDB_val *data, unsigned int flags)
6002 assert(key != NULL);
6003 assert(data != NULL);
6005 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6008 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6012 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6016 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6019 mdb_cursor_init(&mc, txn, dbi, &mx);
6020 return mdb_cursor_put(&mc, key, data, flags);
6023 /** Only a subset of the @ref mdb_env flags can be changed
6024 * at runtime. Changing other flags requires closing the environment
6025 * and re-opening it with the new flags.
6027 #define CHANGEABLE (MDB_NOSYNC)
6029 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6031 if ((flag & CHANGEABLE) != flag)
6034 env->me_flags |= flag;
6036 env->me_flags &= ~flag;
6041 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6046 *arg = env->me_flags;
6051 mdb_env_get_path(MDB_env *env, const char **arg)
6056 *arg = env->me_path;
6060 /** Common code for #mdb_stat() and #mdb_env_stat().
6061 * @param[in] env the environment to operate in.
6062 * @param[in] db the #MDB_db record containing the stats to return.
6063 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6064 * @return 0, this function always succeeds.
6067 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6069 arg->ms_psize = env->me_psize;
6070 arg->ms_depth = db->md_depth;
6071 arg->ms_branch_pages = db->md_branch_pages;
6072 arg->ms_leaf_pages = db->md_leaf_pages;
6073 arg->ms_overflow_pages = db->md_overflow_pages;
6074 arg->ms_entries = db->md_entries;
6079 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6083 if (env == NULL || arg == NULL)
6086 mdb_env_read_meta(env, &toggle);
6088 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6091 /** Set the default comparison functions for a database.
6092 * Called immediately after a database is opened to set the defaults.
6093 * The user can then override them with #mdb_set_compare() or
6094 * #mdb_set_dupsort().
6095 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6096 * @param[in] dbi A database handle returned by #mdb_open()
6099 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6101 if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEKEY)
6102 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memnr;
6103 else if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERKEY)
6104 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_cint;
6106 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memn;
6108 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6109 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP) {
6110 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED)
6111 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_int;
6113 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_cint;
6114 } else if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEDUP) {
6115 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memnr;
6117 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memn;
6120 txn->mt_dbxs[dbi].md_dcmp = NULL;
6124 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6129 int rc, dbflag, exact;
6132 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6133 mdb_default_cmp(txn, FREE_DBI);
6139 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
6140 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
6141 mdb_default_cmp(txn, MAIN_DBI);
6145 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6146 mdb_default_cmp(txn, MAIN_DBI);
6149 /* Is the DB already open? */
6151 for (i=2; i<txn->mt_numdbs; i++) {
6152 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6153 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6159 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
6162 /* Find the DB info */
6166 key.mv_data = (void *)name;
6167 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6168 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6169 if (rc == MDB_SUCCESS) {
6170 /* make sure this is actually a DB */
6171 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6172 if (!(node->mn_flags & F_SUBDATA))
6174 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6175 /* Create if requested */
6177 data.mv_size = sizeof(MDB_db);
6178 data.mv_data = &dummy;
6179 memset(&dummy, 0, sizeof(dummy));
6180 dummy.md_root = P_INVALID;
6181 dummy.md_flags = flags & 0xffff;
6182 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6186 /* OK, got info, add to table */
6187 if (rc == MDB_SUCCESS) {
6188 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
6189 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
6190 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
6191 txn->mt_dbflags[txn->mt_numdbs] = dbflag;
6192 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
6193 *dbi = txn->mt_numdbs;
6194 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
6195 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
6196 mdb_default_cmp(txn, txn->mt_numdbs);
6203 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6205 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6208 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6211 void mdb_close(MDB_env *env, MDB_dbi dbi)
6214 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6216 ptr = env->me_dbxs[dbi].md_name.mv_data;
6217 env->me_dbxs[dbi].md_name.mv_data = NULL;
6218 env->me_dbxs[dbi].md_name.mv_size = 0;
6222 /** Add all the DB's pages to the free list.
6223 * @param[in] mc Cursor on the DB to free.
6224 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6225 * @return 0 on success, non-zero on failure.
6228 mdb_drop0(MDB_cursor *mc, int subs)
6232 rc = mdb_page_search(mc, NULL, 0);
6233 if (rc == MDB_SUCCESS) {
6238 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6239 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6242 mdb_cursor_copy(mc, &mx);
6243 while (mc->mc_snum > 0) {
6244 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6245 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6246 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6247 if (ni->mn_flags & F_SUBDATA) {
6248 mdb_xcursor_init1(mc, ni);
6249 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6255 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6257 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6260 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6265 rc = mdb_cursor_sibling(mc, 1);
6267 /* no more siblings, go back to beginning
6268 * of previous level. (stack was already popped
6269 * by mdb_cursor_sibling)
6271 for (i=1; i<mc->mc_top; i++)
6272 mc->mc_pg[i] = mx.mc_pg[i];
6276 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6277 mc->mc_db->md_root);
6282 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6287 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6290 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6293 rc = mdb_cursor_open(txn, dbi, &mc);
6297 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6301 /* Can't delete the main DB */
6302 if (del && dbi > MAIN_DBI) {
6303 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6305 mdb_close(txn->mt_env, dbi);
6307 txn->mt_dbflags[dbi] |= DB_DIRTY;
6308 txn->mt_dbs[dbi].md_depth = 0;
6309 txn->mt_dbs[dbi].md_branch_pages = 0;
6310 txn->mt_dbs[dbi].md_leaf_pages = 0;
6311 txn->mt_dbs[dbi].md_overflow_pages = 0;
6312 txn->mt_dbs[dbi].md_entries = 0;
6313 txn->mt_dbs[dbi].md_root = P_INVALID;
6316 mdb_cursor_close(mc);
6320 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6322 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6325 txn->mt_dbxs[dbi].md_cmp = cmp;
6329 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6331 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6334 txn->mt_dbxs[dbi].md_dcmp = cmp;
6338 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6340 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6343 txn->mt_dbxs[dbi].md_rel = rel;
6347 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6349 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6352 txn->mt_dbxs[dbi].md_relctx = ctx;