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 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
66 /** @defgroup internal MDB Internals
69 /** @defgroup compat Windows Compatibility Macros
70 * A bunch of macros to minimize the amount of platform-specific ifdefs
71 * needed throughout the rest of the code. When the features this library
72 * needs are similar enough to POSIX to be hidden in a one-or-two line
73 * replacement, this macro approach is used.
77 #define pthread_t DWORD
78 #define pthread_mutex_t HANDLE
79 #define pthread_key_t DWORD
80 #define pthread_self() GetCurrentThreadId()
81 #define pthread_key_create(x,y) *(x) = TlsAlloc()
82 #define pthread_key_delete(x) TlsFree(x)
83 #define pthread_getspecific(x) TlsGetValue(x)
84 #define pthread_setspecific(x,y) TlsSetValue(x,y)
85 #define pthread_mutex_unlock(x) ReleaseMutex(x)
86 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
87 #define LOCK_MUTEX_R(env) pthread_mutex_lock(env->me_rmutex)
88 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(env->me_rmutex)
89 #define LOCK_MUTEX_W(env) pthread_mutex_lock(env->me_wmutex)
90 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(env->me_wmutex)
91 #define getpid() GetCurrentProcessId()
92 #define fdatasync(fd) !FlushFileBuffers(fd)
93 #define ErrCode() GetLastError()
94 #define GetPageSize(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
95 #define close(fd) CloseHandle(fd)
96 #define munmap(ptr,len) UnmapViewOfFile(ptr)
98 /** Lock the reader mutex.
100 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&env->me_txns->mti_mutex)
101 /** Unlock the reader mutex.
103 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&env->me_txns->mti_mutex)
105 /** Lock the writer mutex.
106 * Only a single write transaction is allowed at a time. Other writers
107 * will block waiting for this mutex.
109 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&env->me_txns->mti_wmutex)
110 /** Unlock the writer mutex.
112 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&env->me_txns->mti_wmutex)
114 /** Get the error code for the last failed system function.
116 #define ErrCode() errno
118 /** An abstraction for a file handle.
119 * On POSIX systems file handles are small integers. On Windows
120 * they're opaque pointers.
124 /** A value for an invalid file handle.
125 * Mainly used to initialize file variables and signify that they are
128 #define INVALID_HANDLE_VALUE -1
130 /** Get the size of a memory page for the system.
131 * This is the basic size that the platform's memory manager uses, and is
132 * fundamental to the use of memory-mapped files.
134 #define GetPageSize(x) (x) = sysconf(_SC_PAGE_SIZE)
140 /** A flag for opening a file and requesting synchronous data writes.
141 * This is only used when writing a meta page. It's not strictly needed;
142 * we could just do a normal write and then immediately perform a flush.
143 * But if this flag is available it saves us an extra system call.
145 * @note If O_DSYNC is undefined but exists in /usr/include,
146 * preferably set some compiler flag to get the definition.
147 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
150 # define MDB_DSYNC O_DSYNC
154 /** A page number in the database.
155 * Note that 64 bit page numbers are overkill, since pages themselves
156 * already represent 12-13 bits of addressable memory, and the OS will
157 * always limit applications to a maximum of 63 bits of address space.
159 * @note In the #MDB_node structure, we only store 48 bits of this value,
160 * which thus limits us to only 60 bits of addressable data.
162 typedef ULONG pgno_t;
164 /** @defgroup debug Debug Macros
168 /** Enable debug output.
169 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
170 * read from and written to the database (used for free space management).
175 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
176 # define DPRINTF (void) /* Vararg macros may be unsupported */
178 /** Print a debug message with printf formatting. */
179 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
180 fprintf(stderr, "%s:%d:(%p) " fmt "\n", __func__, __LINE__, pthread_self(), __VA_ARGS__)
182 # define DPRINTF(fmt, ...) ((void) 0)
184 /** Print a debug string.
185 * The string is printed literally, with no format processing.
187 #define DPUTS(arg) DPRINTF("%s", arg)
190 /** A default memory page size.
191 * The actual size is platform-dependent, but we use this for
192 * boot-strapping. We probably should not be using this any more.
193 * The #GetPageSize() macro is used to get the actual size.
195 * Note that we don't currently support Huge pages. On Linux,
196 * regular data files cannot use Huge pages, and in general
197 * Huge pages aren't actually pageable. We rely on the OS
198 * demand-pager to read our data and page it out when memory
199 * pressure from other processes is high. So until OSs have
200 * actual paging support for Huge pages, they're not viable.
202 #define PAGESIZE 4096
204 /** The minimum number of keys required in a database page.
205 * Setting this to a larger value will place a smaller bound on the
206 * maximum size of a data item. Data items larger than this size will
207 * be pushed into overflow pages instead of being stored directly in
208 * the B-tree node. This value used to default to 4. With a page size
209 * of 4096 bytes that meant that any item larger than 1024 bytes would
210 * go into an overflow page. That also meant that on average 2-3KB of
211 * each overflow page was wasted space. The value cannot be lower than
212 * 2 because then there would no longer be a tree structure. With this
213 * value, items larger than 2KB will go into overflow pages, and on
214 * average only 1KB will be wasted.
216 #define MDB_MINKEYS 2
218 /** A stamp that identifies a file as an MDB file.
219 * There's nothing special about this value other than that it is easily
220 * recognizable, and it will reflect any byte order mismatches.
222 #define MDB_MAGIC 0xBEEFC0DE
224 /** The version number for a database's file format. */
225 #define MDB_VERSION 1
227 /** The maximum size of a key in the database.
228 * While data items have essentially unbounded size, we require that
229 * keys all fit onto a regular page. This limit could be raised a bit
230 * further if needed; to something just under #PAGESIZE / #MDB_MINKEYS.
232 #define MAXKEYSIZE 511
237 * This is used for printing a hex dump of a key's contents.
239 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
240 /** Display a key in hex.
242 * Invoke a function to display a key in hex.
244 #define DKEY(x) mdb_dkey(x, kbuf)
250 /** @defgroup lazylock Lazy Locking
251 * Macros for locks that are't actually needed.
252 * The DB view is always consistent because all writes are wrapped in
253 * the wmutex. Finer-grained locks aren't necessary.
257 /** Use lazy locking. I.e., don't lock these accesses at all. */
261 /** Grab the reader lock */
262 #define LAZY_MUTEX_LOCK(x)
263 /** Release the reader lock */
264 #define LAZY_MUTEX_UNLOCK(x)
265 /** Release the DB table reader/writer lock */
266 #define LAZY_RWLOCK_UNLOCK(x)
267 /** Grab the DB table write lock */
268 #define LAZY_RWLOCK_WRLOCK(x)
269 /** Grab the DB table read lock */
270 #define LAZY_RWLOCK_RDLOCK(x)
271 /** Declare the DB table rwlock */
272 #define LAZY_RWLOCK_DEF(x)
273 /** Initialize the DB table rwlock */
274 #define LAZY_RWLOCK_INIT(x,y)
275 /** Destroy the DB table rwlock */
276 #define LAZY_RWLOCK_DESTROY(x)
278 #define LAZY_MUTEX_LOCK(x) pthread_mutex_lock(x)
279 #define LAZY_MUTEX_UNLOCK(x) pthread_mutex_unlock(x)
280 #define LAZY_RWLOCK_UNLOCK(x) pthread_rwlock_unlock(x)
281 #define LAZY_RWLOCK_WRLOCK(x) pthread_rwlock_wrlock(x)
282 #define LAZY_RWLOCK_RDLOCK(x) pthread_rwlock_rdlock(x)
283 #define LAZY_RWLOCK_DEF(x) pthread_rwlock_t x
284 #define LAZY_RWLOCK_INIT(x,y) pthread_rwlock_init(x,y)
285 #define LAZY_RWLOCK_DESTROY(x) pthread_rwlock_destroy(x)
289 /** An invalid page number.
290 * Mainly used to denote an empty tree.
292 #define P_INVALID (~0UL)
294 /** Test if a flag \b f is set in a flag word \b w. */
295 #define F_ISSET(w, f) (((w) & (f)) == (f))
297 /** Used for offsets within a single page.
298 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
301 typedef uint16_t indx_t;
303 /** Default size of memory map.
304 * This is certainly too small for any actual applications. Apps should always set
305 * the size explicitly using #mdb_env_set_mapsize().
307 #define DEFAULT_MAPSIZE 1048576
309 /** @defgroup readers Reader Lock Table
310 * Readers don't acquire any locks for their data access. Instead, they
311 * simply record their transaction ID in the reader table. The reader
312 * mutex is needed just to find an empty slot in the reader table. The
313 * slot's address is saved in thread-specific data so that subsequent read
314 * transactions started by the same thread need no further locking to proceed.
316 * Since the database uses multi-version concurrency control, readers don't
317 * actually need any locking. This table is used to keep track of which
318 * readers are using data from which old transactions, so that we'll know
319 * when a particular old transaction is no longer in use, Old transactions
320 * that have freed any data pages can then have their freed pages reclaimed
321 * for use by a later write transaction.
323 * The lock table is constructed such that reader slots are aligned with the
324 * processor's cache line size. Any slot is only ever used by one thread.
325 * This alignment guarantees that there will be no contention or cache
326 * thrashing as threads update their own slot info, and also eliminates
327 * any need for locking when accessing a slot.
329 * A writer thread will scan every slot in the table to determine the oldest
330 * outstanding reader transaction. Any freed pages older than this will be
331 * reclaimed by the writer. The writer doesn't use any locks when scanning
332 * this table. This means that there's no guarantee that the writer will
333 * see the most up-to-date reader info, but that's not required for correct
334 * operation - all we need is to know the upper bound on the oldest reader,
335 * we don't care at all about the newest reader. So the only consequence of
336 * reading stale information here is that old pages might hang around a
337 * while longer before being reclaimed. That's actually good anyway, because
338 * the longer we delay reclaiming old pages, the more likely it is that a
339 * string of contiguous pages can be found after coalescing old pages from
340 * many old transactions together.
342 * @todo We don't actually do such coalescing yet, we grab pages from one
343 * old transaction at a time.
346 /** Number of slots in the reader table.
347 * This value was chosen somewhat arbitrarily. 126 readers plus a
348 * couple mutexes fit exactly into 8KB on my development machine.
349 * Applications should set the table size using #mdb_env_set_maxreaders().
351 #define DEFAULT_READERS 126
353 /** The size of a CPU cache line in bytes. We want our lock structures
354 * aligned to this size to avoid false cache line sharing in the
356 * This value works for most CPUs. For Itanium this should be 128.
362 /** The information we store in a single slot of the reader table.
363 * In addition to a transaction ID, we also record the process and
364 * thread ID that owns a slot, so that we can detect stale information,
365 * e.g. threads or processes that went away without cleaning up.
366 * @note We currently don't check for stale records. We simply re-init
367 * the table when we know that we're the only process opening the
370 typedef struct MDB_rxbody {
371 /** The current Transaction ID when this transaction began.
372 * Multiple readers that start at the same time will probably have the
373 * same ID here. Again, it's not important to exclude them from
374 * anything; all we need to know is which version of the DB they
375 * started from so we can avoid overwriting any data used in that
376 * particular version.
379 /** The process ID of the process owning this reader txn. */
381 /** The thread ID of the thread owning this txn. */
385 /** The actual reader record, with cacheline padding. */
386 typedef struct MDB_reader {
389 /** shorthand for mrb_txnid */
390 #define mr_txnid mru.mrx.mrb_txnid
391 #define mr_pid mru.mrx.mrb_pid
392 #define mr_tid mru.mrx.mrb_tid
393 /** cache line alignment */
394 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
398 /** The header for the reader table.
399 * The table resides in a memory-mapped file. (This is a different file
400 * than is used for the main database.)
402 * For POSIX the actual mutexes reside in the shared memory of this
403 * mapped file. On Windows, mutexes are named objects allocated by the
404 * kernel; we store the mutex names in this mapped file so that other
405 * processes can grab them. This same approach will also be used on
406 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
407 * process-shared POSIX mutexes.
409 typedef struct MDB_txbody {
410 /** Stamp identifying this as an MDB lock file. It must be set
413 /** Version number of this lock file. Must be set to #MDB_VERSION. */
414 uint32_t mtb_version;
418 /** Mutex protecting access to this table.
419 * This is the reader lock that #LOCK_MUTEX_R acquires.
421 pthread_mutex_t mtb_mutex;
423 /** The ID of the last transaction committed to the database.
424 * This is recorded here only for convenience; the value can always
425 * be determined by reading the main database meta pages.
428 /** The number of slots that have been used in the reader table.
429 * This always records the maximum count, it is not decremented
430 * when readers release their slots.
432 uint32_t mtb_numreaders;
433 /** The ID of the most recent meta page in the database.
434 * This is recorded here only for convenience; the value can always
435 * be determined by reading the main database meta pages.
437 uint32_t mtb_me_toggle;
440 /** The actual reader table definition. */
441 typedef struct MDB_txninfo {
444 #define mti_magic mt1.mtb.mtb_magic
445 #define mti_version mt1.mtb.mtb_version
446 #define mti_mutex mt1.mtb.mtb_mutex
447 #define mti_rmname mt1.mtb.mtb_rmname
448 #define mti_txnid mt1.mtb.mtb_txnid
449 #define mti_numreaders mt1.mtb.mtb_numreaders
450 #define mti_me_toggle mt1.mtb.mtb_me_toggle
451 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
456 #define mti_wmname mt2.mt2_wmname
458 pthread_mutex_t mt2_wmutex;
459 #define mti_wmutex mt2.mt2_wmutex
461 char pad[(sizeof(pthread_mutex_t)+CACHELINE-1) & ~(CACHELINE-1)];
463 MDB_reader mti_readers[1];
467 /** Common header for all page types.
468 * Overflow pages occupy a number of contiguous pages with no
469 * headers on any page after the first.
471 typedef struct MDB_page {
473 pgno_t mp_pgno; /**< page number */
474 void * mp_next; /**< for in-memory list of freed structs */
476 #define P_BRANCH 0x01 /**< branch page */
477 #define P_LEAF 0x02 /**< leaf page */
478 #define P_OVERFLOW 0x04 /**< overflow page */
479 #define P_META 0x08 /**< meta page */
480 #define P_DIRTY 0x10 /**< dirty page */
481 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
485 indx_t mp_lower; /**< lower bound of free space */
486 indx_t mp_upper; /**< upper bound of free space */
488 uint32_t mp_pages; /**< number of overflow pages */
490 indx_t mp_ptrs[1]; /**< dynamic size */
493 /** Size of the page header, excluding dynamic data at the end */
494 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
496 /** Address of first usable data byte in a page, after the header */
497 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
499 /** Number of nodes on a page */
500 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
502 /** The amount of space remaining in the page */
503 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
505 /** The percentage of space used in the page, in tenths of a percent. */
506 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
507 ((env)->me_psize - PAGEHDRSZ))
508 /** The minimum page fill factor, in tenths of a percent.
509 * Pages emptier than this are candidates for merging.
511 #define FILL_THRESHOLD 250
513 /** Test if a page is a leaf page */
514 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
515 /** Test if a page is a LEAF2 page */
516 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
517 /** Test if a page is a branch page */
518 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
519 /** Test if a page is an overflow page */
520 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
522 /** The number of overflow pages needed to store the given size. */
523 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
525 /** Header for a single key/data pair within a page.
526 * We guarantee 2-byte alignment for nodes.
528 typedef struct MDB_node {
529 /** lo and hi are used for data size on leaf nodes and for
530 * child pgno on branch nodes. On 64 bit platforms, flags
531 * is also used for pgno. (branch nodes ignore flags)
533 unsigned short mn_lo;
534 unsigned short mn_hi; /**< part of dsize or pgno */
535 unsigned short mn_flags; /**< flags for special node types */
536 #define F_BIGDATA 0x01 /**< data put on overflow page */
537 #define F_SUBDATA 0x02 /**< data is a sub-database */
538 #define F_DUPDATA 0x04 /**< data has duplicates */
539 unsigned short mn_ksize; /**< key size */
540 char mn_data[1]; /**< key and data are appended here */
543 /** Size of the node header, excluding dynamic data at the end */
544 #define NODESIZE offsetof(MDB_node, mn_data)
546 /** Size of a node in a branch page with a given key.
547 * This is just the node header plus the key, there is no data.
549 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
551 /** Size of a node in a leaf page with a given key and data.
552 * This is node header plus key plus data size.
554 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
556 /** Address of node \b i in page \b p */
557 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
559 /** Address of the key for the node */
560 #define NODEKEY(node) (void *)((node)->mn_data)
562 /** Address of the data for a node */
563 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
565 /** Get the page number pointed to by a branch node */
566 #if LONG_MAX == 0x7fffffff
567 #define NODEPGNO(node) ((node)->mn_lo | ((node)->mn_hi << 16))
568 /** Set the page number in a branch node */
569 #define SETPGNO(node,pgno) do { \
570 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16;} while(0)
572 #define NODEPGNO(node) ((node)->mn_lo | ((node)->mn_hi << 16) | ((unsigned long)(node)->mn_flags << 32))
573 /** Set the page number in a branch node */
574 #define SETPGNO(node,pgno) do { \
575 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
576 (node)->mn_flags = (pgno) >> 32; } while(0)
579 /** Get the size of the data in a leaf node */
580 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
581 /** Set the size of the data for a leaf node */
582 #define SETDSZ(node,size) do { \
583 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
584 /** The size of a key in a node */
585 #define NODEKSZ(node) ((node)->mn_ksize)
587 /** The address of a key in a LEAF2 page.
588 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
589 * There are no node headers, keys are stored contiguously.
591 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
593 /** Set the \b node's key into \b key, if requested. */
594 #define MDB_SET_KEY(node, key) if (key!=NULL) {(key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node);}
596 /** Information about a single database in the environment. */
597 typedef struct MDB_db {
598 uint32_t md_pad; /**< also ksize for LEAF2 pages */
599 uint16_t md_flags; /**< @ref mdb_open */
600 uint16_t md_depth; /**< depth of this tree */
601 ULONG md_branch_pages; /**< number of internal pages */
602 ULONG md_leaf_pages; /**< number of leaf pages */
603 ULONG md_overflow_pages; /**< number of overflow pages */
604 ULONG md_entries; /**< number of data items */
605 pgno_t md_root; /**< the root page of this tree */
608 /** Handle for the DB used to track free pages. */
610 /** Handle for the default DB. */
613 /** Meta page content. */
614 typedef struct MDB_meta {
615 /** Stamp identifying this as an MDB data file. It must be set
618 /** Version number of this lock file. Must be set to #MDB_VERSION. */
620 void *mm_address; /**< address for fixed mapping */
621 size_t mm_mapsize; /**< size of mmap region */
622 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
623 /** The size of pages used in this DB */
624 #define mm_psize mm_dbs[0].md_pad
625 /** Any persistent environment flags. @ref mdb_env */
626 #define mm_flags mm_dbs[0].md_flags
627 pgno_t mm_last_pg; /**< last used page in file */
628 ULONG mm_txnid; /**< txnid that committed this page */
631 /** Auxiliary DB info.
632 * The information here is mostly static/read-only. There is
633 * only a single copy of this record in the environment.
634 * The \b md_dirty flag is not read-only, but only a write
635 * transaction can ever update it, and only write transactions
636 * need to worry about it.
638 typedef struct MDB_dbx {
639 MDB_val md_name; /**< name of the database */
640 MDB_cmp_func *md_cmp; /**< function for comparing keys */
641 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
642 MDB_rel_func *md_rel; /**< user relocate function */
643 MDB_dbi md_parent; /**< parent DB of a sub-DB */
644 unsigned int md_dirty; /**< TRUE if DB was written in this txn */
647 /** A database transaction.
648 * Every operation requires a transaction handle.
651 pgno_t mt_next_pgno; /**< next unallocated page */
652 /** The ID of this transaction. IDs are integers incrementing from 1.
653 * Only committed write transactions increment the ID. If a transaction
654 * aborts, the ID may be re-used by the next writer.
657 MDB_env *mt_env; /**< the DB environment */
658 /** The list of pages that became unused during this transaction.
663 ID2L dirty_list; /**< modified pages */
664 MDB_reader *reader; /**< this thread's slot in the reader table */
666 /** Array of records for each DB known in the environment. */
668 /** Array of MDB_db records for each known DB */
670 /** Number of DB records in use. This number only ever increments;
671 * we don't decrement it when individual DB handles are closed.
673 unsigned int mt_numdbs;
675 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
676 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
677 unsigned int mt_flags;
678 /** Tracks which of the two meta pages was used at the start
679 * of this transaction.
681 unsigned int mt_toggle;
684 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
685 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
686 * raise this on a 64 bit machine.
688 #define CURSOR_STACK 32
692 /** Cursors are used for all DB operations */
694 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
695 struct MDB_xcursor *mc_xcursor;
696 /** The transaction that owns this cursor */
698 /** The database handle this cursor operates on */
700 unsigned short mc_snum; /**< number of pushed pages */
701 unsigned short mc_top; /**< index of top page, mc_snum-1 */
702 unsigned int mc_flags;
703 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
704 #define C_EOF 0x02 /**< No more data */
705 #define C_XDIRTY 0x04 /**< @deprecated mc_xcursor needs to be flushed */
706 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
707 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
710 /** Context for sorted-dup records.
711 * We could have gone to a fully recursive design, with arbitrarily
712 * deep nesting of sub-databases. But for now we only handle these
713 * levels - main DB, optional sub-DB, sorted-duplicate DB.
715 typedef struct MDB_xcursor {
716 /** A sub-cursor for traversing the Dup DB */
717 MDB_cursor mx_cursor;
718 /** A fake transaction struct for pointing to our own table
722 /** Our private DB information tables. Slots 0 and 1 are always
723 * copies of the corresponding slots in the main transaction. These
724 * hold the FREEDB and MAINDB, respectively. If the main cursor is
725 * on a sub-database, that will be copied to slot 2, and the duplicate
726 * database info will be in slot 3. If the main cursor is on the MAINDB
727 * then the duplicate DB info will be in slot 2 and slot 3 will be unused.
734 /** A set of pages freed by an earlier transaction. */
735 typedef struct MDB_oldpages {
736 /** Usually we only read one record from the FREEDB at a time, but
737 * in case we read more, this will chain them together.
739 struct MDB_oldpages *mo_next;
740 /** The ID of the transaction in which these pages were freed. */
742 /** An #IDL of the pages */
743 pgno_t mo_pages[1]; /* dynamic */
746 /** The database environment. */
748 HANDLE me_fd; /**< The main data file */
749 HANDLE me_lfd; /**< The lock file */
750 HANDLE me_mfd; /**< just for writing the meta pages */
751 /** Failed to update the meta page. Probably an I/O error. */
752 #define MDB_FATAL_ERROR 0x80000000U
754 uint32_t me_extrapad; /**< unused for now */
755 unsigned int me_maxreaders; /**< size of the reader table */
756 unsigned int me_numdbs; /**< number of DBs opened */
757 unsigned int me_maxdbs; /**< size of the DB table */
758 char *me_path; /**< path to the DB files */
759 char *me_map; /**< the memory map of the data file */
760 MDB_txninfo *me_txns; /**< the memory map of the lock file */
761 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
762 MDB_txn *me_txn; /**< current write transaction */
763 size_t me_mapsize; /**< size of the data memory map */
764 off_t me_size; /**< current file size */
765 pgno_t me_maxpg; /**< me_mapsize / me_psize */
766 unsigned int me_psize; /**< size of a page, from #GetPageSize */
767 unsigned int me_db_toggle; /**< which DB table is current */
768 MDB_dbx *me_dbxs; /**< array of static DB info */
769 MDB_db *me_dbs[2]; /**< two arrays of MDB_db info */
770 MDB_oldpages *me_pghead; /**< list of old page records */
771 pthread_key_t me_txkey; /**< thread-key for readers */
772 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
773 /** IDL of pages that became unused in a write txn */
774 pgno_t me_free_pgs[MDB_IDL_UM_SIZE];
775 /** ID2L of pages that were written during a write txn */
776 ID2 me_dirty_list[MDB_IDL_UM_SIZE];
777 /** rwlock for the DB tables, if #LAZY_LOCKS is false */
778 LAZY_RWLOCK_DEF(me_dblock);
780 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
784 /** max number of pages to commit in one writev() call */
785 #define MDB_COMMIT_PAGES 64
787 static MDB_page *mdb_alloc_page(MDB_cursor *mc, int num);
788 static int mdb_touch(MDB_cursor *mc);
790 static int mdb_search_page_root(MDB_cursor *mc,
791 MDB_val *key, int modify);
792 static int mdb_search_page(MDB_cursor *mc,
793 MDB_val *key, int modify);
795 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
796 static int mdb_env_read_meta(MDB_env *env, int *which);
797 static int mdb_env_write_meta(MDB_txn *txn);
798 static int mdb_get_page(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
800 static MDB_node *mdb_search_node(MDB_cursor *mc, MDB_val *key, int *exactp);
801 static int mdb_add_node(MDB_cursor *mc, indx_t indx,
802 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags);
803 static void mdb_del_node(MDB_page *mp, indx_t indx, int ksize);
804 static int mdb_del0(MDB_cursor *mc, MDB_node *leaf);
805 static int mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
807 static int mdb_rebalance(MDB_cursor *mc);
808 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
809 static int mdb_move_node(MDB_cursor *csrc, MDB_cursor *cdst);
810 static int mdb_merge(MDB_cursor *csrc, MDB_cursor *cdst);
811 static int mdb_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
813 static MDB_page *mdb_new_page(MDB_cursor *mc, uint32_t flags, int num);
815 static void cursor_pop_page(MDB_cursor *mc);
816 static int cursor_push_page(MDB_cursor *mc, MDB_page *mp);
818 static int mdb_sibling(MDB_cursor *mc, int move_right);
819 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
820 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
821 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
823 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
824 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
826 static void mdb_xcursor_init0(MDB_cursor *mc);
827 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
828 static void mdb_xcursor_init2(MDB_cursor *mc);
829 static void mdb_xcursor_fini(MDB_cursor *mc);
831 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
832 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
834 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
837 static MDB_cmp_func memncmp, memnrcmp, intcmp, cintcmp;
841 static SECURITY_DESCRIPTOR mdb_null_sd;
842 static SECURITY_ATTRIBUTES mdb_all_sa;
843 static int mdb_sec_inited;
846 /** Return the library version info. */
848 mdb_version(int *major, int *minor, int *patch)
850 if (major) *major = MDB_VERSION_MAJOR;
851 if (minor) *minor = MDB_VERSION_MINOR;
852 if (patch) *patch = MDB_VERSION_PATCH;
853 return MDB_VERSION_STRING;
856 /** Table of descriptions for MDB @ref errors */
857 static char *const mdb_errstr[] = {
858 "MDB_KEYEXIST: Key/data pair already exists",
859 "MDB_NOTFOUND: No matching key/data pair found",
860 "MDB_PAGE_NOTFOUND: Requested page not found",
861 "MDB_CORRUPTED: Located page was wrong type",
862 "MDB_PANIC: Update of meta page failed",
863 "MDB_VERSION_MISMATCH: Database environment version mismatch"
867 mdb_strerror(int err)
870 return ("Successful return: 0");
872 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
873 return mdb_errstr[err - MDB_KEYEXIST];
875 return strerror(err);
879 /** Display a key in hexadecimal and return the address of the result.
880 * @param[in] key the key to display
881 * @param[in] buf the buffer to write into. Should always be #DKBUF.
882 * @return The key in hexadecimal form.
885 mdb_dkey(MDB_val *key, char *buf)
888 unsigned char *c = key->mv_data;
890 if (key->mv_size > MAXKEYSIZE)
892 /* may want to make this a dynamic check: if the key is mostly
893 * printable characters, print it as-is instead of converting to hex.
896 for (i=0; i<key->mv_size; i++)
897 ptr += sprintf(ptr, "%02x", *c++);
899 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
906 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
908 return txn->mt_dbxs[dbi].md_cmp(a, b);
911 /** Compare two data items according to a particular database.
912 * This returns a comparison as if the two items were data items of
913 * a sorted duplicates #MDB_DUPSORT database.
914 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
915 * @param[in] dbi A database handle returned by #mdb_open()
916 * @param[in] a The first item to compare
917 * @param[in] b The second item to compare
918 * @return < 0 if a < b, 0 if a == b, > 0 if a > b
921 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
923 if (txn->mt_dbxs[dbi].md_dcmp)
924 return txn->mt_dbxs[dbi].md_dcmp(a, b);
926 return EINVAL; /* too bad you can't distinguish this from a valid result */
929 /** Allocate pages for writing.
930 * If there are free pages available from older transactions, they
931 * will be re-used first. Otherwise a new page will be allocated.
932 * @param[in] mc cursor A cursor handle identifying the transaction and
933 * database for which we are allocating.
934 * @param[in] num the number of pages to allocate.
935 * @return Address of the allocated page(s). Requests for multiple pages
936 * will always be satisfied by a single contiguous chunk of memory.
939 mdb_alloc_page(MDB_cursor *mc, int num)
941 MDB_txn *txn = mc->mc_txn;
943 pgno_t pgno = P_INVALID;
946 if (txn->mt_txnid > 2) {
948 if (!txn->mt_env->me_pghead && mc->mc_dbi != FREE_DBI &&
949 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
950 /* See if there's anything in the free DB */
956 m2.mc_dbi = FREE_DBI;
959 mdb_search_page(&m2, NULL, 0);
960 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
961 kptr = (ULONG *)NODEKEY(leaf);
965 oldest = txn->mt_txnid - 1;
966 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
967 ULONG mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
968 if (mr && mr < oldest)
973 if (oldest > *kptr) {
974 /* It's usable, grab it.
980 mdb_read_data(txn, leaf, &data);
981 idl = (ULONG *)data.mv_data;
982 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
983 mop->mo_next = txn->mt_env->me_pghead;
984 mop->mo_txnid = *kptr;
985 txn->mt_env->me_pghead = mop;
986 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
991 DPRINTF("IDL read txn %lu root %lu num %lu",
992 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
993 for (i=0; i<idl[0]; i++) {
994 DPRINTF("IDL %lu", idl[i+1]);
998 /* drop this IDL from the DB */
999 m2.mc_ki[m2.mc_top] = 0;
1000 m2.mc_flags = C_INITIALIZED;
1001 mdb_cursor_del(&m2, 0);
1004 if (txn->mt_env->me_pghead) {
1005 MDB_oldpages *mop = txn->mt_env->me_pghead;
1007 /* FIXME: For now, always use fresh pages. We
1008 * really ought to search the free list for a
1013 /* peel pages off tail, so we only have to truncate the list */
1014 pgno = MDB_IDL_LAST(mop->mo_pages);
1015 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1017 if (mop->mo_pages[2] > mop->mo_pages[1])
1018 mop->mo_pages[0] = 0;
1022 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1023 txn->mt_env->me_pghead = mop->mo_next;
1030 if (pgno == P_INVALID) {
1031 /* DB size is maxed out */
1032 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg)
1035 if (txn->mt_env->me_dpages && num == 1) {
1036 np = txn->mt_env->me_dpages;
1037 txn->mt_env->me_dpages = np->mp_next;
1039 if ((np = malloc(txn->mt_env->me_psize * num )) == NULL)
1042 if (pgno == P_INVALID) {
1043 np->mp_pgno = txn->mt_next_pgno;
1044 txn->mt_next_pgno += num;
1048 mid.mid = np->mp_pgno;
1050 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1055 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1056 * @param[in] mc cursor pointing to the page to be touched
1057 * @return 0 on success, non-zero on failure.
1060 mdb_touch(MDB_cursor *mc)
1062 MDB_page *mp = mc->mc_pg[mc->mc_top];
1065 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1067 if ((np = mdb_alloc_page(mc, 1)) == NULL)
1069 DPRINTF("touched db %u page %lu -> %lu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1070 assert(mp->mp_pgno != np->mp_pgno);
1071 mdb_midl_append(mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1073 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1076 mp->mp_flags |= P_DIRTY;
1078 mc->mc_pg[mc->mc_top] = mp;
1079 /** If this page has a parent, update the parent to point to
1083 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1089 mdb_env_sync(MDB_env *env, int force)
1092 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1093 if (fdatasync(env->me_fd))
1100 mdb_txn_reset0(MDB_txn *txn);
1102 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1103 * @param[in] txn the transaction handle to initialize
1104 * @return 0 on success, non-zero on failure. This can only
1105 * fail for read-only transactions, and then only if the
1106 * reader table is full.
1109 mdb_txn_renew0(MDB_txn *txn)
1111 MDB_env *env = txn->mt_env;
1113 if (txn->mt_flags & MDB_TXN_RDONLY) {
1114 MDB_reader *r = pthread_getspecific(env->me_txkey);
1117 pid_t pid = getpid();
1118 pthread_t tid = pthread_self();
1121 for (i=0; i<env->me_txns->mti_numreaders; i++)
1122 if (env->me_txns->mti_readers[i].mr_pid == 0)
1124 if (i == env->me_maxreaders) {
1125 UNLOCK_MUTEX_R(env);
1128 env->me_txns->mti_readers[i].mr_pid = pid;
1129 env->me_txns->mti_readers[i].mr_tid = tid;
1130 if (i >= env->me_txns->mti_numreaders)
1131 env->me_txns->mti_numreaders = i+1;
1132 UNLOCK_MUTEX_R(env);
1133 r = &env->me_txns->mti_readers[i];
1134 pthread_setspecific(env->me_txkey, r);
1136 txn->mt_txnid = env->me_txns->mti_txnid;
1137 txn->mt_toggle = env->me_txns->mti_me_toggle;
1138 r->mr_txnid = txn->mt_txnid;
1139 txn->mt_u.reader = r;
1143 txn->mt_txnid = env->me_txns->mti_txnid+1;
1144 txn->mt_toggle = env->me_txns->mti_me_toggle;
1145 txn->mt_u.dirty_list = env->me_dirty_list;
1146 txn->mt_u.dirty_list[0].mid = 0;
1147 txn->mt_free_pgs = env->me_free_pgs;
1148 txn->mt_free_pgs[0] = 0;
1149 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1153 /* Copy the DB arrays */
1154 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
1155 txn->mt_numdbs = env->me_numdbs;
1156 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1157 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1158 if (txn->mt_numdbs > 2)
1159 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
1160 (txn->mt_numdbs - 2) * sizeof(MDB_db));
1161 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1167 mdb_txn_renew(MDB_txn *txn)
1174 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1175 DPUTS("environment had fatal error, must shutdown!");
1179 rc = mdb_txn_renew0(txn);
1180 if (rc == MDB_SUCCESS) {
1181 DPRINTF("renew txn %lu%c %p on mdbenv %p, root page %lu",
1182 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1183 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1189 mdb_txn_begin(MDB_env *env, unsigned int flags, MDB_txn **ret)
1194 if (env->me_flags & MDB_FATAL_ERROR) {
1195 DPUTS("environment had fatal error, must shutdown!");
1198 if ((txn = calloc(1, sizeof(MDB_txn) + env->me_maxdbs * sizeof(MDB_db))) == NULL) {
1199 DPRINTF("calloc: %s", strerror(ErrCode()));
1202 txn->mt_dbs = (MDB_db *)(txn+1);
1203 if (flags & MDB_RDONLY) {
1204 txn->mt_flags |= MDB_TXN_RDONLY;
1208 rc = mdb_txn_renew0(txn);
1213 DPRINTF("begin txn %lu%c %p on mdbenv %p, root page %lu",
1214 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1215 (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1221 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1222 * @param[in] txn the transaction handle to reset
1225 mdb_txn_reset0(MDB_txn *txn)
1227 MDB_env *env = txn->mt_env;
1229 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1230 txn->mt_u.reader->mr_txnid = 0;
1236 /* return all dirty pages to dpage list */
1237 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1238 dp = txn->mt_u.dirty_list[i].mptr;
1239 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1240 dp->mp_next = txn->mt_env->me_dpages;
1241 txn->mt_env->me_dpages = dp;
1243 /* large pages just get freed directly */
1248 while ((mop = txn->mt_env->me_pghead)) {
1249 txn->mt_env->me_pghead = mop->mo_next;
1254 for (i=2; i<env->me_numdbs; i++)
1255 env->me_dbxs[i].md_dirty = 0;
1256 /* The writer mutex was locked in mdb_txn_begin. */
1257 UNLOCK_MUTEX_W(env);
1262 mdb_txn_reset(MDB_txn *txn)
1267 DPRINTF("reset txn %lu%c %p on mdbenv %p, root page %lu",
1268 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1269 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1271 mdb_txn_reset0(txn);
1275 mdb_txn_abort(MDB_txn *txn)
1280 DPRINTF("abort txn %lu%c %p on mdbenv %p, root page %lu",
1281 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1282 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1284 mdb_txn_reset0(txn);
1289 mdb_txn_commit(MDB_txn *txn)
1300 assert(txn != NULL);
1301 assert(txn->mt_env != NULL);
1305 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1310 if (txn != env->me_txn) {
1311 DPUTS("attempt to commit unknown transaction");
1316 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1317 DPUTS("error flag is set, can't commit");
1322 if (!txn->mt_u.dirty_list[0].mid)
1325 DPRINTF("committing txn %lu %p on mdbenv %p, root page %lu",
1326 txn->mt_txnid, txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1329 mc.mc_dbi = FREE_DBI;
1332 /* should only be one record now */
1333 if (env->me_pghead) {
1334 /* make sure first page of freeDB is touched and on freelist */
1335 mdb_search_page(&mc, NULL, 1);
1337 /* save to free list */
1338 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1342 /* make sure last page of freeDB is touched and on freelist */
1343 key.mv_size = MAXKEYSIZE+1;
1345 mdb_search_page(&mc, &key, 1);
1347 mdb_midl_sort(txn->mt_free_pgs);
1351 ULONG *idl = txn->mt_free_pgs;
1352 DPRINTF("IDL write txn %lu root %lu num %lu",
1353 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1354 for (i=0; i<idl[0]; i++) {
1355 DPRINTF("IDL %lu", idl[i+1]);
1359 /* write to last page of freeDB */
1360 key.mv_size = sizeof(pgno_t);
1361 key.mv_data = (char *)&txn->mt_txnid;
1362 data.mv_data = txn->mt_free_pgs;
1363 /* The free list can still grow during this call,
1364 * despite the pre-emptive touches above. So check
1365 * and make sure the entire thing got written.
1368 i = txn->mt_free_pgs[0];
1369 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1370 rc = mdb_cursor_put(&mc, &key, &data, 0);
1375 } while (i != txn->mt_free_pgs[0]);
1377 /* should only be one record now */
1378 if (env->me_pghead) {
1382 mop = env->me_pghead;
1383 key.mv_size = sizeof(pgno_t);
1384 key.mv_data = (char *)&mop->mo_txnid;
1385 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1386 data.mv_data = mop->mo_pages;
1387 mdb_cursor_put(&mc, &key, &data, 0);
1388 free(env->me_pghead);
1389 env->me_pghead = NULL;
1392 /* Update DB root pointers. Their pages have already been
1393 * touched so this is all in-place and cannot fail.
1397 data.mv_size = sizeof(MDB_db);
1399 mc.mc_dbi = MAIN_DBI;
1401 for (i = 2; i < txn->mt_numdbs; i++) {
1402 if (txn->mt_dbxs[i].md_dirty) {
1403 data.mv_data = &txn->mt_dbs[i];
1404 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1409 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
1415 /* Windows actually supports scatter/gather I/O, but only on
1416 * unbuffered file handles. Since we're relying on the OS page
1417 * cache for all our data, that's self-defeating. So we just
1418 * write pages one at a time. We use the ov structure to set
1419 * the write offset, to at least save the overhead of a Seek
1423 memset(&ov, 0, sizeof(ov));
1424 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1426 dp = txn->mt_u.dirty_list[i].mptr;
1427 DPRINTF("committing page %lu", dp->mp_pgno);
1428 size = dp->mp_pgno * env->me_psize;
1429 ov.Offset = size & 0xffffffff;
1430 ov.OffsetHigh = size >> 16;
1431 ov.OffsetHigh >>= 16;
1432 /* clear dirty flag */
1433 dp->mp_flags &= ~P_DIRTY;
1434 wsize = env->me_psize;
1435 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
1436 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
1439 DPRINTF("WriteFile: %d", n);
1446 struct iovec iov[MDB_COMMIT_PAGES];
1450 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1451 dp = txn->mt_u.dirty_list[i].mptr;
1452 if (dp->mp_pgno != next) {
1454 DPRINTF("committing %u dirty pages", n);
1455 rc = writev(env->me_fd, iov, n);
1459 DPUTS("short write, filesystem full?");
1461 DPRINTF("writev: %s", strerror(n));
1468 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
1471 DPRINTF("committing page %lu", dp->mp_pgno);
1472 iov[n].iov_len = env->me_psize;
1473 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
1474 iov[n].iov_base = dp;
1475 size += iov[n].iov_len;
1476 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
1477 /* clear dirty flag */
1478 dp->mp_flags &= ~P_DIRTY;
1479 if (++n >= MDB_COMMIT_PAGES) {
1489 DPRINTF("committing %u dirty pages", n);
1490 rc = writev(env->me_fd, iov, n);
1494 DPUTS("short write, filesystem full?");
1496 DPRINTF("writev: %s", strerror(n));
1503 /* Drop the dirty pages.
1505 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1506 dp = txn->mt_u.dirty_list[i].mptr;
1507 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1508 dp->mp_next = txn->mt_env->me_dpages;
1509 txn->mt_env->me_dpages = dp;
1513 txn->mt_u.dirty_list[i].mid = 0;
1515 txn->mt_u.dirty_list[0].mid = 0;
1517 if ((n = mdb_env_sync(env, 0)) != 0 ||
1518 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
1525 /* update the DB tables */
1527 int toggle = !env->me_db_toggle;
1530 ip = &env->me_dbs[toggle][2];
1531 jp = &txn->mt_dbs[2];
1532 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1533 for (i = 2; i < txn->mt_numdbs; i++) {
1534 if (ip->md_root != jp->md_root)
1539 for (i = 2; i < txn->mt_numdbs; i++) {
1540 if (txn->mt_dbxs[i].md_dirty)
1541 txn->mt_dbxs[i].md_dirty = 0;
1543 env->me_db_toggle = toggle;
1544 env->me_numdbs = txn->mt_numdbs;
1545 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1548 UNLOCK_MUTEX_W(env);
1554 /** Read the environment parameters of a DB environment before
1555 * mapping it into memory.
1556 * @param[in] env the environment handle
1557 * @param[out] meta address of where to store the meta information
1558 * @return 0 on success, non-zero on failure.
1561 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
1563 char page[PAGESIZE];
1568 /* We don't know the page size yet, so use a minimum value.
1572 if (!ReadFile(env->me_fd, page, PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
1574 if ((rc = read(env->me_fd, page, PAGESIZE)) == 0)
1579 else if (rc != PAGESIZE) {
1583 DPRINTF("read: %s", strerror(err));
1587 p = (MDB_page *)page;
1589 if (!F_ISSET(p->mp_flags, P_META)) {
1590 DPRINTF("page %lu not a meta page", p->mp_pgno);
1595 if (m->mm_magic != MDB_MAGIC) {
1596 DPUTS("meta has invalid magic");
1600 if (m->mm_version != MDB_VERSION) {
1601 DPRINTF("database is version %u, expected version %u",
1602 m->mm_version, MDB_VERSION);
1603 return MDB_VERSION_MISMATCH;
1606 memcpy(meta, m, sizeof(*m));
1610 /** Write the environment parameters of a freshly created DB environment.
1611 * @param[in] env the environment handle
1612 * @param[out] meta address of where to store the meta information
1613 * @return 0 on success, non-zero on failure.
1616 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
1623 DPUTS("writing new meta page");
1627 meta->mm_magic = MDB_MAGIC;
1628 meta->mm_version = MDB_VERSION;
1629 meta->mm_psize = psize;
1630 meta->mm_last_pg = 1;
1631 meta->mm_flags = env->me_flags & 0xffff;
1632 meta->mm_flags |= MDB_INTEGERKEY;
1633 meta->mm_dbs[0].md_root = P_INVALID;
1634 meta->mm_dbs[1].md_root = P_INVALID;
1636 p = calloc(2, psize);
1638 p->mp_flags = P_META;
1641 memcpy(m, meta, sizeof(*meta));
1643 q = (MDB_page *)((char *)p + psize);
1646 q->mp_flags = P_META;
1649 memcpy(m, meta, sizeof(*meta));
1654 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
1655 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
1658 rc = write(env->me_fd, p, psize * 2);
1659 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
1665 /** Update the environment info to commit a transaction.
1666 * @param[in] txn the transaction that's being committed
1667 * @return 0 on success, non-zero on failure.
1670 mdb_env_write_meta(MDB_txn *txn)
1673 MDB_meta meta, metab;
1675 int rc, len, toggle;
1681 assert(txn != NULL);
1682 assert(txn->mt_env != NULL);
1684 toggle = !txn->mt_toggle;
1685 DPRINTF("writing meta page %d for root page %lu",
1686 toggle, txn->mt_dbs[MAIN_DBI].md_root);
1690 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
1691 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
1693 ptr = (char *)&meta;
1694 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
1695 len = sizeof(MDB_meta) - off;
1698 meta.mm_dbs[0] = txn->mt_dbs[0];
1699 meta.mm_dbs[1] = txn->mt_dbs[1];
1700 meta.mm_last_pg = txn->mt_next_pgno - 1;
1701 meta.mm_txnid = txn->mt_txnid;
1704 off += env->me_psize;
1707 /* Write to the SYNC fd */
1710 memset(&ov, 0, sizeof(ov));
1712 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
1715 rc = pwrite(env->me_mfd, ptr, len, off);
1720 DPUTS("write failed, disk error?");
1721 /* On a failure, the pagecache still contains the new data.
1722 * Write some old data back, to prevent it from being used.
1723 * Use the non-SYNC fd; we know it will fail anyway.
1725 meta.mm_last_pg = metab.mm_last_pg;
1726 meta.mm_txnid = metab.mm_txnid;
1728 WriteFile(env->me_fd, ptr, len, NULL, &ov);
1730 r2 = pwrite(env->me_fd, ptr, len, off);
1732 env->me_flags |= MDB_FATAL_ERROR;
1735 /* Memory ordering issues are irrelevant; since the entire writer
1736 * is wrapped by wmutex, all of these changes will become visible
1737 * after the wmutex is unlocked. Since the DB is multi-version,
1738 * readers will get consistent data regardless of how fresh or
1739 * how stale their view of these values is.
1741 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
1742 txn->mt_env->me_txns->mti_me_toggle = toggle;
1743 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
1744 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
1749 /** Check both meta pages to see which one is newer.
1750 * @param[in] env the environment handle
1751 * @param[out] which address of where to store the meta toggle ID
1752 * @return 0 on success, non-zero on failure.
1755 mdb_env_read_meta(MDB_env *env, int *which)
1759 assert(env != NULL);
1761 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1764 DPRINTF("Using meta page %d", toggle);
1771 mdb_env_create(MDB_env **env)
1775 e = calloc(1, sizeof(MDB_env));
1776 if (!e) return ENOMEM;
1778 e->me_maxreaders = DEFAULT_READERS;
1780 e->me_fd = INVALID_HANDLE_VALUE;
1781 e->me_lfd = INVALID_HANDLE_VALUE;
1782 e->me_mfd = INVALID_HANDLE_VALUE;
1788 mdb_env_set_mapsize(MDB_env *env, size_t size)
1792 env->me_mapsize = size;
1797 mdb_env_set_maxdbs(MDB_env *env, int dbs)
1801 env->me_maxdbs = dbs;
1806 mdb_env_set_maxreaders(MDB_env *env, int readers)
1810 env->me_maxreaders = readers;
1815 mdb_env_get_maxreaders(MDB_env *env, int *readers)
1817 if (!env || !readers)
1819 *readers = env->me_maxreaders;
1823 /** Further setup required for opening an MDB environment
1826 mdb_env_open2(MDB_env *env, unsigned int flags)
1828 int i, newenv = 0, toggle;
1832 env->me_flags = flags;
1834 memset(&meta, 0, sizeof(meta));
1836 if ((i = mdb_env_read_header(env, &meta)) != 0) {
1839 DPUTS("new mdbenv");
1843 if (!env->me_mapsize) {
1844 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
1850 LONG sizelo, sizehi;
1851 sizelo = env->me_mapsize & 0xffffffff;
1852 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
1854 /* Windows won't create mappings for zero length files.
1855 * Just allocate the maxsize right now.
1858 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
1859 if (!SetEndOfFile(env->me_fd))
1861 SetFilePointer(env->me_fd, 0, NULL, 0);
1863 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
1864 sizehi, sizelo, NULL);
1867 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
1875 if (meta.mm_address && (flags & MDB_FIXEDMAP))
1877 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
1879 if (env->me_map == MAP_FAILED)
1884 meta.mm_mapsize = env->me_mapsize;
1885 if (flags & MDB_FIXEDMAP)
1886 meta.mm_address = env->me_map;
1887 i = mdb_env_init_meta(env, &meta);
1888 if (i != MDB_SUCCESS) {
1889 munmap(env->me_map, env->me_mapsize);
1893 env->me_psize = meta.mm_psize;
1895 env->me_maxpg = env->me_mapsize / env->me_psize;
1897 p = (MDB_page *)env->me_map;
1898 env->me_metas[0] = METADATA(p);
1899 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
1901 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
1904 DPRINTF("opened database version %u, pagesize %u",
1905 env->me_metas[toggle]->mm_version, env->me_psize);
1906 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
1907 DPRINTF("entries: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
1908 DPRINTF("branch pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
1909 DPRINTF("leaf pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
1910 DPRINTF("overflow pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
1911 DPRINTF("root: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
1917 /* Windows doesn't support destructor callbacks for thread-specific storage */
1919 mdb_env_reader_dest(void *ptr)
1921 MDB_reader *reader = ptr;
1923 reader->mr_txnid = 0;
1929 /* downgrade the exclusive lock on the region back to shared */
1931 mdb_env_share_locks(MDB_env *env)
1935 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1937 env->me_txns->mti_me_toggle = toggle;
1938 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
1943 /* First acquire a shared lock. The Unlock will
1944 * then release the existing exclusive lock.
1946 memset(&ov, 0, sizeof(ov));
1947 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
1948 UnlockFile(env->me_lfd, 0, 0, 1, 0);
1952 struct flock lock_info;
1953 /* The shared lock replaces the existing lock */
1954 memset((void *)&lock_info, 0, sizeof(lock_info));
1955 lock_info.l_type = F_RDLCK;
1956 lock_info.l_whence = SEEK_SET;
1957 lock_info.l_start = 0;
1958 lock_info.l_len = 1;
1959 fcntl(env->me_lfd, F_SETLK, &lock_info);
1965 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
1973 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
1974 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
1975 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
1979 /* Try to get exclusive lock. If we succeed, then
1980 * nobody is using the lock region and we should initialize it.
1983 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
1987 memset(&ov, 0, sizeof(ov));
1988 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
1994 size = GetFileSize(env->me_lfd, NULL);
1996 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2000 /* Try to get exclusive lock. If we succeed, then
2001 * nobody is using the lock region and we should initialize it.
2004 struct flock lock_info;
2005 memset((void *)&lock_info, 0, sizeof(lock_info));
2006 lock_info.l_type = F_WRLCK;
2007 lock_info.l_whence = SEEK_SET;
2008 lock_info.l_start = 0;
2009 lock_info.l_len = 1;
2010 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2014 lock_info.l_type = F_RDLCK;
2015 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2022 size = lseek(env->me_lfd, 0, SEEK_END);
2024 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2025 if (size < rsize && *excl) {
2027 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2028 if (!SetEndOfFile(env->me_lfd)) {
2033 if (ftruncate(env->me_lfd, rsize) != 0) {
2040 size = rsize - sizeof(MDB_txninfo);
2041 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2046 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2052 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2054 if (!env->me_txns) {
2060 env->me_txns = mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2062 if (env->me_txns == MAP_FAILED) {
2070 if (!mdb_sec_inited) {
2071 InitializeSecurityDescriptor(&mdb_null_sd,
2072 SECURITY_DESCRIPTOR_REVISION);
2073 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2074 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2075 mdb_all_sa.bInheritHandle = FALSE;
2076 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2079 /* FIXME: only using up to 20 characters of the env path here,
2080 * probably not enough to assure uniqueness...
2082 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%.20s", lpath);
2083 ptr = env->me_txns->mti_rmname + sizeof("Global\\MDBr");
2084 while ((ptr = strchr(ptr, '\\')))
2086 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2087 if (!env->me_rmutex) {
2091 sprintf(env->me_txns->mti_rmname, "Global\\MDBw%.20s", lpath);
2092 ptr = env->me_txns->mti_rmname + sizeof("Global\\MDBw");
2093 while ((ptr = strchr(ptr, '\\')))
2095 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2096 if (!env->me_wmutex) {
2101 pthread_mutexattr_t mattr;
2103 pthread_mutexattr_init(&mattr);
2104 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2108 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2109 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2111 env->me_txns->mti_version = MDB_VERSION;
2112 env->me_txns->mti_magic = MDB_MAGIC;
2113 env->me_txns->mti_txnid = 0;
2114 env->me_txns->mti_numreaders = 0;
2115 env->me_txns->mti_me_toggle = 0;
2118 if (env->me_txns->mti_magic != MDB_MAGIC) {
2119 DPUTS("lock region has invalid magic");
2123 if (env->me_txns->mti_version != MDB_VERSION) {
2124 DPRINTF("lock region is version %u, expected version %u",
2125 env->me_txns->mti_version, MDB_VERSION);
2126 rc = MDB_VERSION_MISMATCH;
2130 if (rc != EACCES && rc != EAGAIN) {
2134 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2135 if (!env->me_rmutex) {
2139 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2140 if (!env->me_wmutex) {
2150 env->me_lfd = INVALID_HANDLE_VALUE;
2155 /** The name of the lock file in the DB environment */
2156 #define LOCKNAME "/lock.mdb"
2157 /** The name of the data file in the DB environment */
2158 #define DATANAME "/data.mdb"
2160 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2162 int oflags, rc, len, excl;
2163 char *lpath, *dpath;
2166 lpath = malloc(len + sizeof(LOCKNAME) + len + sizeof(DATANAME));
2169 dpath = lpath + len + sizeof(LOCKNAME);
2170 sprintf(lpath, "%s" LOCKNAME, path);
2171 sprintf(dpath, "%s" DATANAME, path);
2173 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
2178 if (F_ISSET(flags, MDB_RDONLY)) {
2179 oflags = GENERIC_READ;
2180 len = OPEN_EXISTING;
2182 oflags = GENERIC_READ|GENERIC_WRITE;
2185 mode = FILE_ATTRIBUTE_NORMAL;
2186 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2187 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2192 if (F_ISSET(flags, MDB_RDONLY))
2195 oflags = O_RDWR | O_CREAT;
2197 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
2203 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
2204 /* synchronous fd for meta writes */
2206 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2207 mode |= FILE_FLAG_WRITE_THROUGH;
2208 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2209 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2214 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2215 oflags |= MDB_DSYNC;
2216 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
2221 env->me_path = strdup(path);
2222 DPRINTF("opened dbenv %p", (void *) env);
2223 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
2224 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
2226 mdb_env_share_locks(env);
2227 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
2228 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
2229 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
2235 if (env->me_fd != INVALID_HANDLE_VALUE) {
2237 env->me_fd = INVALID_HANDLE_VALUE;
2239 if (env->me_lfd != INVALID_HANDLE_VALUE) {
2241 env->me_lfd = INVALID_HANDLE_VALUE;
2249 mdb_env_close(MDB_env *env)
2256 while (env->me_dpages) {
2257 dp = env->me_dpages;
2258 env->me_dpages = dp->mp_next;
2262 free(env->me_dbs[1]);
2263 free(env->me_dbs[0]);
2267 LAZY_RWLOCK_DESTROY(&env->me_dblock);
2268 pthread_key_delete(env->me_txkey);
2271 munmap(env->me_map, env->me_mapsize);
2276 pid_t pid = getpid();
2278 for (i=0; i<env->me_txns->mti_numreaders; i++)
2279 if (env->me_txns->mti_readers[i].mr_pid == pid)
2280 env->me_txns->mti_readers[i].mr_pid = 0;
2281 munmap(env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
2287 /* only for aligned ints */
2289 intcmp(const MDB_val *a, const MDB_val *b)
2291 if (a->mv_size == sizeof(long))
2293 unsigned long *la, *lb;
2298 unsigned int *ia, *ib;
2305 /* ints must always be the same size */
2307 cintcmp(const MDB_val *a, const MDB_val *b)
2309 #if __BYTE_ORDER == __LITTLE_ENDIAN
2310 unsigned short *u, *c;
2313 u = a->mv_data + a->mv_size;
2314 c = b->mv_data + a->mv_size;
2317 } while(!x && u > (unsigned short *)a->mv_data);
2320 return memcmp(a->mv_data, b->mv_data, a->mv_size);
2325 memncmp(const MDB_val *a, const MDB_val *b)
2331 len_diff = a->mv_size - b->mv_size;
2334 diff = memcmp(a->mv_data, b->mv_data, len);
2335 return diff ? diff : len_diff;
2339 memnrcmp(const MDB_val *a, const MDB_val *b)
2341 const unsigned char *p1, *p2, *p1_lim;
2344 if (b->mv_size == 0)
2345 return a->mv_size != 0;
2346 if (a->mv_size == 0)
2349 p1 = (const unsigned char *)a->mv_data + a->mv_size - 1;
2350 p2 = (const unsigned char *)b->mv_data + b->mv_size - 1;
2352 len_diff = a->mv_size - b->mv_size;
2354 p1_lim = p1 - a->mv_size;
2356 p1_lim = p1 - b->mv_size;
2358 while (p1 > p1_lim) {
2368 /* Search for key within a leaf page, using binary search.
2369 * Returns the smallest entry larger or equal to the key.
2370 * If exactp is non-null, stores whether the found entry was an exact match
2371 * in *exactp (1 or 0).
2372 * If kip is non-null, stores the index of the found entry in *kip.
2373 * If no entry larger or equal to the key is found, returns NULL.
2376 mdb_search_node(MDB_cursor *mc, MDB_val *key, int *exactp)
2378 unsigned int i = 0, nkeys;
2381 MDB_page *mp = mc->mc_pg[mc->mc_top];
2382 MDB_node *node = NULL;
2387 nkeys = NUMKEYS(mp);
2389 DPRINTF("searching %u keys in %s page %lu",
2390 nkeys, IS_LEAF(mp) ? "leaf" : "branch",
2395 low = IS_LEAF(mp) ? 0 : 1;
2397 cmp = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_cmp;
2399 nodekey.mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2400 node = NODEPTR(mp, 0); /* fake */
2402 while (low <= high) {
2403 i = (low + high) >> 1;
2406 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
2408 node = NODEPTR(mp, i);
2410 nodekey.mv_size = node->mn_ksize;
2411 nodekey.mv_data = NODEKEY(node);
2414 rc = cmp(key, &nodekey);
2418 DPRINTF("found leaf index %u [%s], rc = %i",
2419 i, DKEY(&nodekey), rc);
2421 DPRINTF("found branch index %u [%s -> %lu], rc = %i",
2422 i, DKEY(&nodekey), NODEPGNO(node), rc);
2433 if (rc > 0) { /* Found entry is less than the key. */
2434 i++; /* Skip to get the smallest entry larger than key. */
2436 node = NODEPTR(mp, i);
2439 *exactp = (rc == 0);
2440 /* store the key index */
2441 mc->mc_ki[mc->mc_top] = i;
2443 /* There is no entry larger or equal to the key. */
2446 /* nodeptr is fake for LEAF2 */
2451 cursor_pop_page(MDB_cursor *mc)
2456 top = mc->mc_pg[mc->mc_top];
2461 DPRINTF("popped page %lu off db %u cursor %p", top->mp_pgno,
2462 mc->mc_dbi, (void *) mc);
2467 cursor_push_page(MDB_cursor *mc, MDB_page *mp)
2469 DPRINTF("pushing page %lu on db %u cursor %p", mp->mp_pgno,
2470 mc->mc_dbi, (void *) mc);
2472 if (mc->mc_snum >= CURSOR_STACK)
2475 mc->mc_top = mc->mc_snum++;
2476 mc->mc_pg[mc->mc_top] = mp;
2477 mc->mc_ki[mc->mc_top] = 0;
2483 mdb_get_page(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
2487 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
2489 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
2490 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
2491 p = txn->mt_u.dirty_list[x].mptr;
2495 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
2496 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
2500 DPRINTF("page %lu not found", pgno);
2503 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
2507 mdb_search_page_root(MDB_cursor *mc, MDB_val *key, int modify)
2509 MDB_page *mp = mc->mc_pg[mc->mc_top];
2514 while (IS_BRANCH(mp)) {
2517 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
2518 assert(NUMKEYS(mp) > 1);
2519 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
2521 if (key == NULL) /* Initialize cursor to first page. */
2522 mc->mc_ki[mc->mc_top] = 0;
2523 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
2524 /* cursor to last page */
2525 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
2528 node = mdb_search_node(mc, key, &exact);
2530 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
2532 assert(mc->mc_ki[mc->mc_top] > 0);
2533 mc->mc_ki[mc->mc_top]--;
2538 DPRINTF("following index %u for key [%s]",
2539 mc->mc_ki[mc->mc_top], DKEY(key));
2540 assert(mc->mc_ki[mc->mc_top] < NUMKEYS(mp));
2541 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2543 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mp)))
2546 if ((rc = cursor_push_page(mc, mp)))
2550 if ((rc = mdb_touch(mc)) != 0)
2552 mp = mc->mc_pg[mc->mc_top];
2557 DPRINTF("internal error, index points to a %02X page!?",
2559 return MDB_CORRUPTED;
2562 DPRINTF("found leaf page %lu for key [%s]", mp->mp_pgno,
2563 key ? DKEY(key) : NULL);
2568 /* Search for the page a given key should be in.
2569 * Pushes parent pages on the cursor stack.
2570 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
2571 * If modify is true, visited pages are updated with new page numbers.
2574 mdb_search_page(MDB_cursor *mc, MDB_val *key, int modify)
2579 /* Make sure the txn is still viable, then find the root from
2580 * the txn's db table.
2582 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
2583 DPUTS("transaction has failed, must abort");
2586 root = mc->mc_txn->mt_dbs[mc->mc_dbi].md_root;
2588 if (root == P_INVALID) { /* Tree is empty. */
2589 DPUTS("tree is empty");
2590 return MDB_NOTFOUND;
2593 if ((rc = mdb_get_page(mc->mc_txn, root, &mc->mc_pg[0])))
2599 DPRINTF("db %u root page %lu has flags 0x%X",
2600 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
2603 /* For sub-databases, update main root first */
2604 if (mc->mc_dbi > MAIN_DBI && !mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty) {
2606 mc2.mc_txn = mc->mc_txn;
2607 mc2.mc_dbi = MAIN_DBI;
2608 rc = mdb_search_page(&mc2, &mc->mc_txn->mt_dbxs[mc->mc_dbi].md_name, 1);
2611 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
2613 if (!F_ISSET(mc->mc_pg[0]->mp_flags, P_DIRTY)) {
2614 if ((rc = mdb_touch(mc)))
2616 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = mc->mc_pg[0]->mp_pgno;
2620 return mdb_search_page_root(mc, key, modify);
2624 mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
2626 MDB_page *omp; /* overflow mpage */
2630 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
2631 data->mv_size = NODEDSZ(leaf);
2632 data->mv_data = NODEDATA(leaf);
2636 /* Read overflow data.
2638 data->mv_size = NODEDSZ(leaf);
2639 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
2640 if ((rc = mdb_get_page(txn, pgno, &omp))) {
2641 DPRINTF("read overflow page %lu failed", pgno);
2644 data->mv_data = METADATA(omp);
2650 mdb_get(MDB_txn *txn, MDB_dbi dbi,
2651 MDB_val *key, MDB_val *data)
2660 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
2662 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
2665 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2672 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
2673 mc.mc_xcursor = &mx;
2674 mdb_xcursor_init0(&mc);
2676 mc.mc_xcursor = NULL;
2678 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
2682 mdb_sibling(MDB_cursor *mc, int move_right)
2689 if (mc->mc_snum < 2) {
2690 return MDB_NOTFOUND; /* root has no siblings */
2692 ptop = mc->mc_top-1;
2694 DPRINTF("parent page is page %lu, index %u",
2695 mc->mc_pg[ptop]->mp_pgno, mc->mc_ki[ptop]);
2697 cursor_pop_page(mc);
2698 if (move_right ? (mc->mc_ki[ptop] + 1u >= NUMKEYS(mc->mc_pg[ptop]))
2699 : (mc->mc_ki[ptop] == 0)) {
2700 DPRINTF("no more keys left, moving to %s sibling",
2701 move_right ? "right" : "left");
2702 if ((rc = mdb_sibling(mc, move_right)) != MDB_SUCCESS)
2709 DPRINTF("just moving to %s index key %u",
2710 move_right ? "right" : "left", mc->mc_ki[ptop]);
2712 assert(IS_BRANCH(mc->mc_pg[ptop]));
2714 indx = NODEPTR(mc->mc_pg[ptop], mc->mc_ki[ptop]);
2715 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(indx), &mp)))
2718 cursor_push_page(mc, mp);
2724 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
2730 if (mc->mc_flags & C_EOF) {
2731 return MDB_NOTFOUND;
2734 assert(mc->mc_flags & C_INITIALIZED);
2736 mp = mc->mc_pg[mc->mc_top];
2738 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) {
2739 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2740 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2741 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
2742 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
2743 if (op != MDB_NEXT || rc == MDB_SUCCESS)
2747 mc->mc_xcursor->mx_cursor.mc_flags = 0;
2748 if (op == MDB_NEXT_DUP)
2749 return MDB_NOTFOUND;
2753 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) mc);
2755 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
2756 DPUTS("=====> move to next sibling page");
2757 if (mdb_sibling(mc, 1) != MDB_SUCCESS) {
2758 mc->mc_flags |= C_EOF;
2759 return MDB_NOTFOUND;
2761 mp = mc->mc_pg[mc->mc_top];
2762 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
2764 mc->mc_ki[mc->mc_top]++;
2766 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
2767 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
2770 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2771 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
2775 assert(IS_LEAF(mp));
2776 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2778 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2779 mdb_xcursor_init1(mc, leaf);
2782 if ((rc = mdb_read_data(mc->mc_txn, leaf, data) != MDB_SUCCESS))
2785 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2786 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
2787 if (rc != MDB_SUCCESS)
2792 MDB_SET_KEY(leaf, key);
2797 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
2803 assert(mc->mc_flags & C_INITIALIZED);
2805 mp = mc->mc_pg[mc->mc_top];
2807 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) {
2808 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2809 if (op == MDB_PREV || op == MDB_PREV_DUP) {
2810 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2811 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
2812 if (op != MDB_PREV || rc == MDB_SUCCESS)
2815 mc->mc_xcursor->mx_cursor.mc_flags = 0;
2816 if (op == MDB_PREV_DUP)
2817 return MDB_NOTFOUND;
2822 DPRINTF("cursor_prev: top page is %lu in cursor %p", mp->mp_pgno, (void *) mc);
2824 if (mc->mc_ki[mc->mc_top] == 0) {
2825 DPUTS("=====> move to prev sibling page");
2826 if (mdb_sibling(mc, 0) != MDB_SUCCESS) {
2827 mc->mc_flags &= ~C_INITIALIZED;
2828 return MDB_NOTFOUND;
2830 mp = mc->mc_pg[mc->mc_top];
2831 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
2832 DPRINTF("prev page is %lu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
2834 mc->mc_ki[mc->mc_top]--;
2836 mc->mc_flags &= ~C_EOF;
2838 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
2839 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
2842 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2843 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
2847 assert(IS_LEAF(mp));
2848 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2850 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2851 mdb_xcursor_init1(mc, leaf);
2854 if ((rc = mdb_read_data(mc->mc_txn, leaf, data) != MDB_SUCCESS))
2857 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2858 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
2859 if (rc != MDB_SUCCESS)
2864 MDB_SET_KEY(leaf, key);
2869 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
2870 MDB_cursor_op op, int *exactp)
2878 assert(key->mv_size > 0);
2880 /* See if we're already on the right page */
2881 if (mc->mc_flags & C_INITIALIZED) {
2884 if (mc->mc_pg[mc->mc_top]->mp_flags & P_LEAF2) {
2885 nodekey.mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2886 nodekey.mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, nodekey.mv_size);
2888 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
2889 MDB_SET_KEY(leaf, &nodekey);
2891 rc = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_cmp(key, &nodekey);
2893 /* Probably happens rarely, but first node on the page
2894 * was the one we wanted.
2896 mc->mc_ki[mc->mc_top] = 0;
2901 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
2906 if (NUMKEYS(mc->mc_pg[mc->mc_top]) > 1) {
2907 if (mc->mc_pg[mc->mc_top]->mp_flags & P_LEAF2) {
2908 nodekey.mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top],
2909 NUMKEYS(mc->mc_pg[mc->mc_top])-1, nodekey.mv_size);
2911 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
2912 MDB_SET_KEY(leaf, &nodekey);
2914 rc = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_cmp(key, &nodekey);
2916 /* last node was the one we wanted */
2917 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top])-1;
2921 /* This is definitely the right page, skip search_page */
2926 /* If any parents have right-sibs, search.
2927 * Otherwise, there's nothing further.
2929 for (i=0; i<mc->mc_top; i++)
2931 NUMKEYS(mc->mc_pg[i])-1)
2933 if (i == mc->mc_top) {
2934 /* There are no other pages */
2935 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
2936 return MDB_NOTFOUND;
2941 rc = mdb_search_page(mc, key, 0);
2942 if (rc != MDB_SUCCESS)
2945 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
2948 leaf = mdb_search_node(mc, key, exactp);
2949 if (exactp != NULL && !*exactp) {
2950 /* MDB_SET specified and not an exact match. */
2951 return MDB_NOTFOUND;
2955 DPUTS("===> inexact leaf not found, goto sibling");
2956 if ((rc = mdb_sibling(mc, 1)) != MDB_SUCCESS)
2957 return rc; /* no entries matched */
2958 mc->mc_ki[mc->mc_top] = 0;
2959 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
2960 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
2964 mc->mc_flags |= C_INITIALIZED;
2965 mc->mc_flags &= ~C_EOF;
2967 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
2968 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2969 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
2973 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2974 mdb_xcursor_init1(mc, leaf);
2977 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2978 if (op == MDB_SET || op == MDB_SET_RANGE) {
2979 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
2982 if (op == MDB_GET_BOTH) {
2988 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
2989 if (rc != MDB_SUCCESS)
2992 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
2994 if ((rc = mdb_read_data(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
2996 rc = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dcmp(data, &d2);
2998 if (op == MDB_GET_BOTH || rc > 0)
2999 return MDB_NOTFOUND;
3003 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3008 /* The key already matches in all other cases */
3009 if (op == MDB_SET_RANGE)
3010 MDB_SET_KEY(leaf, key);
3011 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3017 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3022 rc = mdb_search_page(mc, NULL, 0);
3023 if (rc != MDB_SUCCESS)
3025 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3027 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3028 mc->mc_flags |= C_INITIALIZED;
3029 mc->mc_flags &= ~C_EOF;
3031 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3032 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
3033 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
3038 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3039 mdb_xcursor_init1(mc, leaf);
3040 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3045 mc->mc_xcursor->mx_cursor.mc_flags = 0;
3046 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3050 MDB_SET_KEY(leaf, key);
3055 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3061 lkey.mv_size = MAXKEYSIZE+1;
3062 lkey.mv_data = NULL;
3064 rc = mdb_search_page(mc, &lkey, 0);
3065 if (rc != MDB_SUCCESS)
3067 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3069 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
3070 mc->mc_flags |= C_INITIALIZED;
3071 mc->mc_flags &= ~C_EOF;
3073 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
3075 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3076 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
3077 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
3082 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3083 mdb_xcursor_init1(mc, leaf);
3084 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3088 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3093 MDB_SET_KEY(leaf, key);
3098 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3108 case MDB_GET_BOTH_RANGE:
3109 if (data == NULL || mc->mc_xcursor == NULL) {
3116 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3118 } else if (op == MDB_SET_RANGE)
3119 rc = mdb_cursor_set(mc, key, data, op, NULL);
3121 rc = mdb_cursor_set(mc, key, data, op, &exact);
3123 case MDB_GET_MULTIPLE:
3125 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPFIXED) ||
3126 !(mc->mc_flags & C_INITIALIZED)) {
3131 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
3132 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
3135 case MDB_NEXT_MULTIPLE:
3137 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPFIXED)) {
3141 if (!(mc->mc_flags & C_INITIALIZED))
3142 rc = mdb_cursor_first(mc, key, data);
3144 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
3145 if (rc == MDB_SUCCESS) {
3146 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
3149 mx = &mc->mc_xcursor->mx_cursor;
3150 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
3151 mx->mc_txn->mt_dbs[mx->mc_dbi].md_pad;
3152 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
3153 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
3161 case MDB_NEXT_NODUP:
3162 if (!(mc->mc_flags & C_INITIALIZED))
3163 rc = mdb_cursor_first(mc, key, data);
3165 rc = mdb_cursor_next(mc, key, data, op);
3169 case MDB_PREV_NODUP:
3170 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
3171 rc = mdb_cursor_last(mc, key, data);
3173 rc = mdb_cursor_prev(mc, key, data, op);
3176 rc = mdb_cursor_first(mc, key, data);
3180 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) ||
3181 !(mc->mc_flags & C_INITIALIZED) ||
3182 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3186 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3189 rc = mdb_cursor_last(mc, key, data);
3193 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) ||
3194 !(mc->mc_flags & C_INITIALIZED) ||
3195 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3199 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3202 DPRINTF("unhandled/unimplemented cursor operation %u", op);
3211 mdb_cursor_touch(MDB_cursor *mc)
3215 if (mc->mc_dbi > MAIN_DBI && !mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty) {
3217 mc2.mc_txn = mc->mc_txn;
3218 mc2.mc_dbi = MAIN_DBI;
3219 rc = mdb_search_page(&mc2, &mc->mc_txn->mt_dbxs[mc->mc_dbi].md_name, 1);
3221 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
3223 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
3224 if (!F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) {
3228 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root =
3229 mc->mc_pg[mc->mc_top]->mp_pgno;
3233 mc->mc_top = mc->mc_snum-1;
3238 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3242 MDB_val xdata, *rdata, dkey;
3244 char dbuf[PAGESIZE];
3250 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
3253 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
3254 mc->mc_dbi, DKEY(key), key->mv_size, data->mv_size);
3258 if (flags == MDB_CURRENT) {
3259 if (!(mc->mc_flags & C_INITIALIZED))
3262 } else if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_root == P_INVALID) {
3264 /* new database, write a root leaf page */
3265 DPUTS("allocating new root leaf page");
3266 if ((np = mdb_new_page(mc, P_LEAF, 1)) == NULL) {
3270 cursor_push_page(mc, np);
3271 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = np->mp_pgno;
3272 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth++;
3273 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
3274 if ((mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
3276 np->mp_flags |= P_LEAF2;
3277 mc->mc_flags |= C_INITIALIZED;
3282 rc = mdb_cursor_set(mc, key, NULL, MDB_SET, &exact);
3283 if (flags == MDB_NOOVERWRITE && rc == 0) {
3284 DPRINTF("duplicate key [%s]", DKEY(key));
3285 return MDB_KEYEXIST;
3287 if (rc && rc != MDB_NOTFOUND)
3291 /* Cursor is positioned, now make sure all pages are writable */
3292 rc2 = mdb_cursor_touch(mc);
3293 if (rc2) return rc2;
3296 /* The key already exists */
3297 if (rc == MDB_SUCCESS) {
3298 /* there's only a key anyway, so this is a no-op */
3299 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3300 unsigned int ksize = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
3301 if (key->mv_size != ksize)
3303 if (flags == MDB_CURRENT) {
3304 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
3305 memcpy(ptr, key->mv_data, ksize);
3310 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3313 if (F_ISSET(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags, MDB_DUPSORT)) {
3314 /* Was a single item before, must convert now */
3315 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3316 dkey.mv_size = NODEDSZ(leaf);
3317 dkey.mv_data = dbuf;
3318 memcpy(dbuf, NODEDATA(leaf), dkey.mv_size);
3319 /* data matches, ignore it */
3320 if (!mdb_dcmp(mc->mc_txn, mc->mc_dbi, data, &dkey))
3321 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
3322 memset(&dummy, 0, sizeof(dummy));
3323 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPFIXED) {
3324 dummy.md_pad = data->mv_size;
3325 dummy.md_flags = MDB_DUPFIXED;
3326 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_INTEGERDUP)
3327 dummy.md_flags |= MDB_INTEGERKEY;
3329 dummy.md_root = P_INVALID;
3330 if (dkey.mv_size == sizeof(MDB_db)) {
3331 memcpy(NODEDATA(leaf), &dummy, sizeof(dummy));
3334 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
3337 xdata.mv_size = sizeof(MDB_db);
3338 xdata.mv_data = &dummy;
3343 /* same size, just replace it */
3344 if (!F_ISSET(leaf->mn_flags, F_BIGDATA) &&
3345 NODEDSZ(leaf) == data->mv_size) {
3346 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
3349 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
3351 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
3357 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
3358 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
3359 rc = mdb_split(mc, key, rdata, P_INVALID);
3361 /* There is room already in this leaf page. */
3362 rc = mdb_add_node(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, 0);
3365 if (rc != MDB_SUCCESS)
3366 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
3368 /* Remember if we just added a subdatabase */
3369 if (flags & F_SUBDATA) {
3370 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3371 leaf->mn_flags |= F_SUBDATA;
3374 /* Now store the actual data in the child DB. Note that we're
3375 * storing the user data in the keys field, so there are strict
3376 * size limits on dupdata. The actual data fields of the child
3377 * DB are all zero size.
3380 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3382 if (flags == MDB_CURRENT)
3383 mdb_xcursor_init2(mc);
3385 mdb_xcursor_init1(mc, leaf);
3388 if (flags == MDB_NODUPDATA)
3389 flags = MDB_NOOVERWRITE;
3390 /* converted, write the original data first */
3392 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, flags);
3394 leaf->mn_flags |= F_DUPDATA;
3396 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, flags);
3397 mdb_xcursor_fini(mc);
3398 memcpy(NODEDATA(leaf),
3399 &mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi],
3402 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries++;
3409 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
3414 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
3417 if (!mc->mc_flags & C_INITIALIZED)
3420 rc = mdb_cursor_touch(mc);
3423 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3425 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3426 if (flags != MDB_NODUPDATA) {
3427 mdb_xcursor_init2(mc);
3428 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
3429 mdb_xcursor_fini(mc);
3430 /* If sub-DB still has entries, we're done */
3431 if (mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_root
3433 memcpy(NODEDATA(leaf),
3434 &mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi],
3436 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries--;
3439 /* otherwise fall thru and delete the sub-DB */
3442 /* add all the child DB's pages to the free list */
3443 rc = mdb_search_page(&mc->mc_xcursor->mx_cursor, NULL, 0);
3444 if (rc == MDB_SUCCESS) {
3449 mx = &mc->mc_xcursor->mx_cursor;
3450 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries -=
3451 mx->mc_txn->mt_dbs[mx->mc_dbi].md_entries;
3453 cursor_pop_page(mx);
3455 while (mx->mc_snum > 1) {
3456 for (i=0; i<NUMKEYS(mx->mc_pg[mx->mc_top]); i++) {
3458 ni = NODEPTR(mx->mc_pg[mx->mc_top], i);
3461 mdb_midl_append(mc->mc_txn->mt_free_pgs, pg);
3463 rc = mdb_sibling(mx, 1);
3468 mdb_midl_append(mc->mc_txn->mt_free_pgs,
3469 mx->mc_txn->mt_dbs[mx->mc_dbi].md_root);
3473 return mdb_del0(mc, leaf);
3476 /* Allocate a page and initialize it
3479 mdb_new_page(MDB_cursor *mc, uint32_t flags, int num)
3483 if ((np = mdb_alloc_page(mc, num)) == NULL)
3485 DPRINTF("allocated new mpage %lu, page size %u",
3486 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
3487 np->mp_flags = flags | P_DIRTY;
3488 np->mp_lower = PAGEHDRSZ;
3489 np->mp_upper = mc->mc_txn->mt_env->me_psize;
3492 mc->mc_txn->mt_dbs[mc->mc_dbi].md_branch_pages++;
3493 else if (IS_LEAF(np))
3494 mc->mc_txn->mt_dbs[mc->mc_dbi].md_leaf_pages++;
3495 else if (IS_OVERFLOW(np)) {
3496 mc->mc_txn->mt_dbs[mc->mc_dbi].md_overflow_pages += num;
3504 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
3508 sz = LEAFSIZE(key, data);
3509 if (data->mv_size >= env->me_psize / MDB_MINKEYS) {
3510 /* put on overflow page */
3511 sz -= data->mv_size - sizeof(pgno_t);
3515 return sz + sizeof(indx_t);
3519 mdb_branch_size(MDB_env *env, MDB_val *key)
3524 if (sz >= env->me_psize / MDB_MINKEYS) {
3525 /* put on overflow page */
3526 /* not implemented */
3527 /* sz -= key->size - sizeof(pgno_t); */
3530 return sz + sizeof(indx_t);
3534 mdb_add_node(MDB_cursor *mc, indx_t indx,
3535 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
3538 size_t node_size = NODESIZE;
3541 MDB_page *mp = mc->mc_pg[mc->mc_top];
3542 MDB_page *ofp = NULL; /* overflow page */
3545 assert(mp->mp_upper >= mp->mp_lower);
3547 DPRINTF("add to %s page %lu index %i, data size %zu key size %zu [%s]",
3548 IS_LEAF(mp) ? "leaf" : "branch",
3549 mp->mp_pgno, indx, data ? data->mv_size : 0,
3550 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
3553 /* Move higher keys up one slot. */
3554 int ksize = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad, dif;
3555 char *ptr = LEAF2KEY(mp, indx, ksize);
3556 dif = NUMKEYS(mp) - indx;
3558 memmove(ptr+ksize, ptr, dif*ksize);
3559 /* insert new key */
3560 memcpy(ptr, key->mv_data, ksize);
3562 /* Just using these for counting */
3563 mp->mp_lower += sizeof(indx_t);
3564 mp->mp_upper -= ksize - sizeof(indx_t);
3569 node_size += key->mv_size;
3573 if (F_ISSET(flags, F_BIGDATA)) {
3574 /* Data already on overflow page. */
3575 node_size += sizeof(pgno_t);
3576 } else if (data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
3577 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
3578 /* Put data on overflow page. */
3579 DPRINTF("data size is %zu, put on overflow page",
3581 node_size += sizeof(pgno_t);
3582 if ((ofp = mdb_new_page(mc, P_OVERFLOW, ovpages)) == NULL)
3584 DPRINTF("allocated overflow page %lu", ofp->mp_pgno);
3587 node_size += data->mv_size;
3590 node_size += node_size & 1;
3592 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
3593 DPRINTF("not enough room in page %lu, got %u ptrs",
3594 mp->mp_pgno, NUMKEYS(mp));
3595 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
3596 mp->mp_upper - mp->mp_lower);
3597 DPRINTF("node size = %zu", node_size);
3601 /* Move higher pointers up one slot. */
3602 for (i = NUMKEYS(mp); i > indx; i--)
3603 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
3605 /* Adjust free space offsets. */
3606 ofs = mp->mp_upper - node_size;
3607 assert(ofs >= mp->mp_lower + sizeof(indx_t));
3608 mp->mp_ptrs[indx] = ofs;
3610 mp->mp_lower += sizeof(indx_t);
3612 /* Write the node data. */
3613 node = NODEPTR(mp, indx);
3614 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
3615 node->mn_flags = flags;
3617 SETDSZ(node,data->mv_size);
3622 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
3627 if (F_ISSET(flags, F_BIGDATA))
3628 memcpy(node->mn_data + key->mv_size, data->mv_data,
3631 memcpy(node->mn_data + key->mv_size, data->mv_data,
3634 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
3636 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
3644 mdb_del_node(MDB_page *mp, indx_t indx, int ksize)
3647 indx_t i, j, numkeys, ptr;
3651 DPRINTF("delete node %u on %s page %lu", indx,
3652 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
3653 assert(indx < NUMKEYS(mp));
3656 int x = NUMKEYS(mp) - 1 - indx;
3657 base = LEAF2KEY(mp, indx, ksize);
3659 memmove(base, base + ksize, x * ksize);
3660 mp->mp_lower -= sizeof(indx_t);
3661 mp->mp_upper += ksize - sizeof(indx_t);
3665 node = NODEPTR(mp, indx);
3666 sz = NODESIZE + node->mn_ksize;
3668 if (F_ISSET(node->mn_flags, F_BIGDATA))
3669 sz += sizeof(pgno_t);
3671 sz += NODEDSZ(node);
3675 ptr = mp->mp_ptrs[indx];
3676 numkeys = NUMKEYS(mp);
3677 for (i = j = 0; i < numkeys; i++) {
3679 mp->mp_ptrs[j] = mp->mp_ptrs[i];
3680 if (mp->mp_ptrs[i] < ptr)
3681 mp->mp_ptrs[j] += sz;
3686 base = (char *)mp + mp->mp_upper;
3687 memmove(base + sz, base, ptr - mp->mp_upper);
3689 mp->mp_lower -= sizeof(indx_t);
3694 mdb_xcursor_init0(MDB_cursor *mc)
3696 MDB_xcursor *mx = mc->mc_xcursor;
3699 mx->mx_txn = *mc->mc_txn;
3700 mx->mx_txn.mt_dbxs = mx->mx_dbxs;
3701 mx->mx_txn.mt_dbs = mx->mx_dbs;
3702 mx->mx_dbxs[0] = mc->mc_txn->mt_dbxs[0];
3703 mx->mx_dbxs[1] = mc->mc_txn->mt_dbxs[1];
3704 if (mc->mc_dbi > 1) {
3705 mx->mx_dbxs[2] = mc->mc_txn->mt_dbxs[mc->mc_dbi];
3710 mx->mx_dbxs[dbn+1].md_parent = dbn;
3711 mx->mx_dbxs[dbn+1].md_cmp = mx->mx_dbxs[dbn].md_dcmp;
3712 mx->mx_dbxs[dbn+1].md_rel = mx->mx_dbxs[dbn].md_rel;
3713 mx->mx_dbxs[dbn+1].md_dirty = 0;
3714 mx->mx_txn.mt_numdbs = dbn+2;
3715 mx->mx_txn.mt_u = mc->mc_txn->mt_u;
3717 mx->mx_cursor.mc_xcursor = NULL;
3718 mx->mx_cursor.mc_txn = &mx->mx_txn;
3719 mx->mx_cursor.mc_dbi = dbn+1;
3723 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
3725 MDB_db *db = NODEDATA(node);
3726 MDB_xcursor *mx = mc->mc_xcursor;
3728 mx->mx_dbs[0] = mc->mc_txn->mt_dbs[0];
3729 mx->mx_dbs[1] = mc->mc_txn->mt_dbs[1];
3730 if (mc->mc_dbi > 1) {
3731 mx->mx_dbs[2] = mc->mc_txn->mt_dbs[mc->mc_dbi];
3732 mx->mx_dbxs[2].md_dirty = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty;
3737 DPRINTF("Sub-db %u for db %u root page %lu", dbn, mc->mc_dbi, db->md_root);
3738 mx->mx_dbs[dbn] = *db;
3739 if (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY))
3740 mx->mx_dbxs[dbn].md_dirty = 1;
3741 mx->mx_dbxs[dbn].md_name.mv_data = NODEKEY(node);
3742 mx->mx_dbxs[dbn].md_name.mv_size = node->mn_ksize;
3743 mx->mx_txn.mt_next_pgno = mc->mc_txn->mt_next_pgno;
3744 mx->mx_cursor.mc_snum = 0;
3745 mx->mx_cursor.mc_flags = 0;
3749 mdb_xcursor_init2(MDB_cursor *mc)
3751 MDB_xcursor *mx = mc->mc_xcursor;
3753 mx->mx_dbs[0] = mc->mc_txn->mt_dbs[0];
3754 mx->mx_dbs[1] = mc->mc_txn->mt_dbs[1];
3755 if (mc->mc_dbi > 1) {
3756 mx->mx_dbs[2] = mc->mc_txn->mt_dbs[mc->mc_dbi];
3757 mx->mx_dbxs[2].md_dirty = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty;
3762 DPRINTF("Sub-db %u for db %u root page %lu", dbn, mc->mc_dbi,
3763 mx->mx_dbs[dbn].md_root);
3764 mx->mx_txn.mt_next_pgno = mc->mc_txn->mt_next_pgno;
3768 mdb_xcursor_fini(MDB_cursor *mc)
3770 MDB_xcursor *mx = mc->mc_xcursor;
3771 mc->mc_txn->mt_next_pgno = mx->mx_txn.mt_next_pgno;
3772 mc->mc_txn->mt_dbs[0] = mx->mx_dbs[0];
3773 mc->mc_txn->mt_dbs[1] = mx->mx_dbs[1];
3774 if (mc->mc_dbi > 1) {
3775 mc->mc_txn->mt_dbs[mc->mc_dbi] = mx->mx_dbs[2];
3776 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = mx->mx_dbxs[2].md_dirty;
3781 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
3784 size_t size = sizeof(MDB_cursor);
3786 if (txn == NULL || ret == NULL || !dbi || dbi >= txn->mt_numdbs)
3789 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
3790 size += sizeof(MDB_xcursor);
3792 if ((mc = calloc(1, size)) != NULL) {
3795 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
3796 MDB_xcursor *mx = (MDB_xcursor *)(mc + 1);
3797 mc->mc_xcursor = mx;
3798 mdb_xcursor_init0(mc);
3809 /* Return the count of duplicate data items for the current key */
3811 mdb_cursor_count(MDB_cursor *mc, unsigned long *countp)
3815 if (mc == NULL || countp == NULL)
3818 if (!(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT))
3821 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3822 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3825 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
3828 *countp = mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_entries;
3834 mdb_cursor_close(MDB_cursor *mc)
3842 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
3844 indx_t ptr, i, numkeys;
3851 node = NODEPTR(mp, indx);
3852 ptr = mp->mp_ptrs[indx];
3853 DPRINTF("update key %u (ofs %u) [%.*s] to [%s] on page %lu",
3855 (int)node->mn_ksize, (char *)NODEKEY(node),
3859 delta = key->mv_size - node->mn_ksize;
3861 if (delta > 0 && SIZELEFT(mp) < delta) {
3862 DPRINTF("OUCH! Not enough room, delta = %d", delta);
3866 numkeys = NUMKEYS(mp);
3867 for (i = 0; i < numkeys; i++) {
3868 if (mp->mp_ptrs[i] <= ptr)
3869 mp->mp_ptrs[i] -= delta;
3872 base = (char *)mp + mp->mp_upper;
3873 len = ptr - mp->mp_upper + NODESIZE;
3874 memmove(base - delta, base, len);
3875 mp->mp_upper -= delta;
3877 node = NODEPTR(mp, indx);
3878 node->mn_ksize = key->mv_size;
3881 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
3886 /* Move a node from csrc to cdst.
3889 mdb_move_node(MDB_cursor *csrc, MDB_cursor *cdst)
3896 /* Mark src and dst as dirty. */
3897 if ((rc = mdb_touch(csrc)) ||
3898 (rc = mdb_touch(cdst)))
3901 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3902 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
3903 key.mv_size = csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_pad;
3904 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3906 data.mv_data = NULL;
3908 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
3909 unsigned int snum = csrc->mc_snum;
3910 /* must find the lowest key below src */
3911 mdb_search_page_root(csrc, NULL, 0);
3912 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
3913 csrc->mc_snum = snum--;
3914 csrc->mc_top = snum;
3916 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
3918 key.mv_size = NODEKSZ(srcnode);
3919 key.mv_data = NODEKEY(srcnode);
3920 data.mv_size = NODEDSZ(srcnode);
3921 data.mv_data = NODEDATA(srcnode);
3923 DPRINTF("moving %s node %u [%s] on page %lu to node %u on page %lu",
3924 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
3925 csrc->mc_ki[csrc->mc_top],
3927 csrc->mc_pg[csrc->mc_top]->mp_pgno,
3928 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
3930 /* Add the node to the destination page.
3932 rc = mdb_add_node(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, NODEPGNO(srcnode),
3934 if (rc != MDB_SUCCESS)
3937 /* Delete the node from the source page.
3939 mdb_del_node(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3941 /* Update the parent separators.
3943 if (csrc->mc_ki[csrc->mc_top] == 0) {
3944 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
3945 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3946 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3948 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
3949 key.mv_size = NODEKSZ(srcnode);
3950 key.mv_data = NODEKEY(srcnode);
3952 DPRINTF("update separator for source page %lu to [%s]",
3953 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
3954 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
3955 &key)) != MDB_SUCCESS)
3958 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
3960 nullkey.mv_size = 0;
3961 assert(mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey) == MDB_SUCCESS);
3965 if (cdst->mc_ki[cdst->mc_top] == 0) {
3966 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
3967 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3968 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
3970 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
3971 key.mv_size = NODEKSZ(srcnode);
3972 key.mv_data = NODEKEY(srcnode);
3974 DPRINTF("update separator for destination page %lu to [%s]",
3975 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
3976 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
3977 &key)) != MDB_SUCCESS)
3980 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
3982 nullkey.mv_size = 0;
3983 assert(mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey) == MDB_SUCCESS);
3991 mdb_merge(MDB_cursor *csrc, MDB_cursor *cdst)
3998 DPRINTF("merging page %lu into %lu", csrc->mc_pg[csrc->mc_top]->mp_pgno, cdst->mc_pg[cdst->mc_top]->mp_pgno);
4000 assert(csrc->mc_snum > 1); /* can't merge root page */
4001 assert(cdst->mc_snum > 1);
4003 /* Mark dst as dirty. */
4004 if ((rc = mdb_touch(cdst)))
4007 /* Move all nodes from src to dst.
4009 j = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
4010 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
4011 key.mv_size = csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_pad;
4012 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
4013 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
4014 rc = mdb_add_node(cdst, j, &key, NULL, 0, 0);
4015 if (rc != MDB_SUCCESS)
4017 key.mv_data = (char *)key.mv_data + key.mv_size;
4020 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
4021 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
4023 key.mv_size = srcnode->mn_ksize;
4024 key.mv_data = NODEKEY(srcnode);
4025 data.mv_size = NODEDSZ(srcnode);
4026 data.mv_data = NODEDATA(srcnode);
4027 rc = mdb_add_node(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
4028 if (rc != MDB_SUCCESS)
4033 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
4034 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);
4036 /* Unlink the src page from parent and add to free list.
4038 mdb_del_node(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
4039 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
4041 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
4045 mdb_midl_append(csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
4046 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
4047 csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_leaf_pages--;
4049 csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_branch_pages--;
4050 cursor_pop_page(csrc);
4052 return mdb_rebalance(csrc);
4056 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
4060 cdst->mc_txn = csrc->mc_txn;
4061 cdst->mc_dbi = csrc->mc_dbi;
4062 cdst->mc_snum = csrc->mc_snum;
4063 cdst->mc_top = csrc->mc_top;
4064 cdst->mc_flags = csrc->mc_flags;
4066 for (i=0; i<csrc->mc_snum; i++) {
4067 cdst->mc_pg[i] = csrc->mc_pg[i];
4068 cdst->mc_ki[i] = csrc->mc_ki[i];
4073 mdb_rebalance(MDB_cursor *mc)
4081 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
4082 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
4083 mc->mc_pg[mc->mc_top]->mp_pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
4085 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
4086 DPRINTF("no need to rebalance page %lu, above fill threshold",
4087 mc->mc_pg[mc->mc_top]->mp_pgno);
4091 if (mc->mc_snum < 2) {
4092 if (NUMKEYS(mc->mc_pg[mc->mc_top]) == 0) {
4093 DPUTS("tree is completely empty");
4094 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = P_INVALID;
4095 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth = 0;
4096 mc->mc_txn->mt_dbs[mc->mc_dbi].md_leaf_pages = 0;
4097 mdb_midl_append(mc->mc_txn->mt_free_pgs, mc->mc_pg[mc->mc_top]->mp_pgno);
4098 } else if (IS_BRANCH(mc->mc_pg[mc->mc_top]) && NUMKEYS(mc->mc_pg[mc->mc_top]) == 1) {
4099 DPUTS("collapsing root page!");
4100 mdb_midl_append(mc->mc_txn->mt_free_pgs, mc->mc_pg[mc->mc_top]->mp_pgno);
4101 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = NODEPGNO(NODEPTR(mc->mc_pg[mc->mc_top], 0));
4102 if ((rc = mdb_get_page(mc->mc_txn, mc->mc_txn->mt_dbs[mc->mc_dbi].md_root, &root)))
4104 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth--;
4105 mc->mc_txn->mt_dbs[mc->mc_dbi].md_branch_pages--;
4107 DPUTS("root page doesn't need rebalancing");
4111 /* The parent (branch page) must have at least 2 pointers,
4112 * otherwise the tree is invalid.
4114 ptop = mc->mc_top-1;
4115 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
4117 /* Leaf page fill factor is below the threshold.
4118 * Try to move keys from left or right neighbor, or
4119 * merge with a neighbor page.
4124 mdb_cursor_copy(mc, &mn);
4125 mn.mc_xcursor = NULL;
4127 if (mc->mc_ki[ptop] == 0) {
4128 /* We're the leftmost leaf in our parent.
4130 DPUTS("reading right neighbor");
4132 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
4133 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
4135 mn.mc_ki[mn.mc_top] = 0;
4136 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
4138 /* There is at least one neighbor to the left.
4140 DPUTS("reading left neighbor");
4142 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
4143 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
4145 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
4146 mc->mc_ki[mc->mc_top] = 0;
4149 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
4150 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);
4152 /* If the neighbor page is above threshold and has at least two
4153 * keys, move one key from it.
4155 * Otherwise we should try to merge them.
4157 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
4158 return mdb_move_node(&mn, mc);
4159 else { /* FIXME: if (has_enough_room()) */
4160 if (mc->mc_ki[ptop] == 0)
4161 return mdb_merge(&mn, mc);
4163 return mdb_merge(mc, &mn);
4168 mdb_del0(MDB_cursor *mc, MDB_node *leaf)
4172 /* add overflow pages to free list */
4173 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4177 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4178 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4179 for (i=0; i<ovpages; i++) {
4180 DPRINTF("freed ov page %lu", pg);
4181 mdb_midl_append(mc->mc_txn->mt_free_pgs, pg);
4185 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad);
4186 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries--;
4187 rc = mdb_rebalance(mc);
4188 if (rc != MDB_SUCCESS)
4189 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4195 mdb_del(MDB_txn *txn, MDB_dbi dbi,
4196 MDB_val *key, MDB_val *data)
4201 MDB_val rdata, *xdata;
4205 assert(key != NULL);
4207 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
4209 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4212 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
4216 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4223 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4224 mc.mc_xcursor = &mx;
4225 mdb_xcursor_init0(&mc);
4227 mc.mc_xcursor = NULL;
4239 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
4241 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
4245 /* Split page <mc->top>, and insert <key,(data|newpgno)> in either left or
4246 * right sibling, at index <mc->ki> (as if unsplit). Updates mc->top and
4247 * mc->ki with the actual values after split, ie if mc->top and mc->ki
4248 * refer to a node in the new right sibling page.
4251 mdb_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
4254 int rc = MDB_SUCCESS, ins_new = 0;
4257 unsigned int i, j, split_indx, nkeys, pmax;
4259 MDB_val sepkey, rkey, rdata;
4261 MDB_page *mp, *rp, *pp;
4266 mp = mc->mc_pg[mc->mc_top];
4267 newindx = mc->mc_ki[mc->mc_top];
4269 DPRINTF("-----> splitting %s page %lu and adding [%s] at index %i",
4270 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
4271 DKEY(newkey), mc->mc_ki[mc->mc_top]);
4273 if (mc->mc_snum < 2) {
4274 if ((pp = mdb_new_page(mc, P_BRANCH, 1)) == NULL)
4276 /* shift current top to make room for new parent */
4277 mc->mc_pg[1] = mc->mc_pg[0];
4278 mc->mc_ki[1] = mc->mc_ki[0];
4281 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = pp->mp_pgno;
4282 DPRINTF("root split! new root = %lu", pp->mp_pgno);
4283 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth++;
4285 /* Add left (implicit) pointer. */
4286 if ((rc = mdb_add_node(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
4287 /* undo the pre-push */
4288 mc->mc_pg[0] = mc->mc_pg[1];
4289 mc->mc_ki[0] = mc->mc_ki[1];
4290 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = mp->mp_pgno;
4291 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth--;
4298 ptop = mc->mc_top-1;
4299 DPRINTF("parent branch page is %lu", mc->mc_pg[ptop]->mp_pgno);
4302 /* Create a right sibling. */
4303 if ((rp = mdb_new_page(mc, mp->mp_flags, 1)) == NULL)
4305 mdb_cursor_copy(mc, &mn);
4306 mn.mc_pg[mn.mc_top] = rp;
4307 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
4308 DPRINTF("new right sibling: page %lu", rp->mp_pgno);
4310 nkeys = NUMKEYS(mp);
4311 split_indx = nkeys / 2 + 1;
4316 unsigned int lsize, rsize, ksize;
4317 /* Move half of the keys to the right sibling */
4319 x = mc->mc_ki[mc->mc_top] - split_indx;
4320 ksize = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
4321 split = LEAF2KEY(mp, split_indx, ksize);
4322 rsize = (nkeys - split_indx) * ksize;
4323 lsize = (nkeys - split_indx) * sizeof(indx_t);
4324 mp->mp_lower -= lsize;
4325 rp->mp_lower += lsize;
4326 mp->mp_upper += rsize - lsize;
4327 rp->mp_upper -= rsize - lsize;
4328 sepkey.mv_size = ksize;
4329 if (newindx == split_indx) {
4330 sepkey.mv_data = newkey->mv_data;
4332 sepkey.mv_data = split;
4335 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
4336 memcpy(rp->mp_ptrs, split, rsize);
4337 sepkey.mv_data = rp->mp_ptrs;
4338 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
4339 memcpy(ins, newkey->mv_data, ksize);
4340 mp->mp_lower += sizeof(indx_t);
4341 mp->mp_upper -= ksize - sizeof(indx_t);
4344 memcpy(rp->mp_ptrs, split, x * ksize);
4345 ins = LEAF2KEY(rp, x, ksize);
4346 memcpy(ins, newkey->mv_data, ksize);
4347 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
4348 rp->mp_lower += sizeof(indx_t);
4349 rp->mp_upper -= ksize - sizeof(indx_t);
4350 mc->mc_ki[mc->mc_top] = x;
4351 mc->mc_pg[mc->mc_top] = rp;
4356 /* For leaf pages, check the split point based on what
4357 * fits where, since otherwise add_node can fail.
4360 unsigned int psize, nsize;
4361 /* Maximum free space in an empty page */
4362 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
4363 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
4364 if (newindx < split_indx) {
4366 for (i=0; i<split_indx; i++) {
4367 node = NODEPTR(mp, i);
4368 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
4369 if (F_ISSET(node->mn_flags, F_BIGDATA))
4370 psize += sizeof(pgno_t);
4372 psize += NODEDSZ(node);
4381 for (i=nkeys-1; i>=split_indx; i--) {
4382 node = NODEPTR(mp, i);
4383 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
4384 if (F_ISSET(node->mn_flags, F_BIGDATA))
4385 psize += sizeof(pgno_t);
4387 psize += NODEDSZ(node);
4397 /* First find the separating key between the split pages.
4399 if (newindx == split_indx) {
4400 sepkey.mv_size = newkey->mv_size;
4401 sepkey.mv_data = newkey->mv_data;
4403 node = NODEPTR(mp, split_indx);
4404 sepkey.mv_size = node->mn_ksize;
4405 sepkey.mv_data = NODEKEY(node);
4409 DPRINTF("separator is [%s]", DKEY(&sepkey));
4411 /* Copy separator key to the parent.
4413 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
4416 rc = mdb_split(&mn, &sepkey, NULL, rp->mp_pgno);
4418 /* Right page might now have changed parent.
4419 * Check if left page also changed parent.
4421 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
4422 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
4423 mc->mc_pg[ptop] = mn.mc_pg[ptop];
4424 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
4428 rc = mdb_add_node(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
4434 if (rc != MDB_SUCCESS) {
4438 /* Move half of the keys to the right sibling. */
4440 /* grab a page to hold a temporary copy */
4441 if (mc->mc_txn->mt_env->me_dpages) {
4442 copy = mc->mc_txn->mt_env->me_dpages;
4443 mc->mc_txn->mt_env->me_dpages = copy->mp_next;
4445 if ((copy = malloc(mc->mc_txn->mt_env->me_psize)) == NULL)
4449 copy->mp_pgno = mp->mp_pgno;
4450 copy->mp_flags = mp->mp_flags;
4451 copy->mp_lower = PAGEHDRSZ;
4452 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
4453 mc->mc_pg[mc->mc_top] = copy;
4454 for (i = j = 0; i <= nkeys; j++) {
4455 if (i == split_indx) {
4456 /* Insert in right sibling. */
4457 /* Reset insert index for right sibling. */
4458 j = (i == newindx && ins_new);
4459 mc->mc_pg[mc->mc_top] = rp;
4462 if (i == newindx && !ins_new) {
4463 /* Insert the original entry that caused the split. */
4464 rkey.mv_data = newkey->mv_data;
4465 rkey.mv_size = newkey->mv_size;
4467 rdata.mv_data = newdata->mv_data;
4468 rdata.mv_size = newdata->mv_size;
4475 /* Update page and index for the new key. */
4476 mc->mc_ki[mc->mc_top] = j;
4477 } else if (i == nkeys) {
4480 node = NODEPTR(mp, i);
4481 rkey.mv_data = NODEKEY(node);
4482 rkey.mv_size = node->mn_ksize;
4484 rdata.mv_data = NODEDATA(node);
4485 rdata.mv_size = NODEDSZ(node);
4487 pgno = NODEPGNO(node);
4488 flags = node->mn_flags;
4493 if (!IS_LEAF(mp) && j == 0) {
4494 /* First branch index doesn't need key data. */
4498 rc = mdb_add_node(mc, j, &rkey, &rdata, pgno, flags);
4501 /* reset back to original page */
4502 if (newindx < split_indx)
4503 mc->mc_pg[mc->mc_top] = mp;
4505 nkeys = NUMKEYS(copy);
4506 for (i=0; i<nkeys; i++)
4507 mp->mp_ptrs[i] = copy->mp_ptrs[i];
4508 mp->mp_lower = copy->mp_lower;
4509 mp->mp_upper = copy->mp_upper;
4510 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
4511 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
4513 /* return tmp page to freelist */
4514 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
4515 mc->mc_txn->mt_env->me_dpages = copy;
4520 mdb_put(MDB_txn *txn, MDB_dbi dbi,
4521 MDB_val *key, MDB_val *data, unsigned int flags)
4526 assert(key != NULL);
4527 assert(data != NULL);
4529 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4532 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
4536 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4540 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA)) != flags)
4547 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4548 mc.mc_xcursor = &mx;
4549 mdb_xcursor_init0(&mc);
4551 mc.mc_xcursor = NULL;
4553 return mdb_cursor_put(&mc, key, data, flags);
4557 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
4559 /** Only a subset of the @ref mdb_env flags can be changed
4560 * at runtime. Changing other flags requires closing the environment
4561 * and re-opening it with the new flags.
4563 #define CHANGEABLE (MDB_NOSYNC)
4564 if ((flag & CHANGEABLE) != flag)
4567 env->me_flags |= flag;
4569 env->me_flags &= ~flag;
4574 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
4579 *arg = env->me_flags;
4584 mdb_env_get_path(MDB_env *env, const char **arg)
4589 *arg = env->me_path;
4594 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
4596 arg->ms_psize = env->me_psize;
4597 arg->ms_depth = db->md_depth;
4598 arg->ms_branch_pages = db->md_branch_pages;
4599 arg->ms_leaf_pages = db->md_leaf_pages;
4600 arg->ms_overflow_pages = db->md_overflow_pages;
4601 arg->ms_entries = db->md_entries;
4606 mdb_env_stat(MDB_env *env, MDB_stat *arg)
4610 if (env == NULL || arg == NULL)
4613 mdb_env_read_meta(env, &toggle);
4615 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
4619 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
4621 if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEKEY)
4622 txn->mt_dbxs[dbi].md_cmp = memnrcmp;
4623 else if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERKEY)
4624 txn->mt_dbxs[dbi].md_cmp = cintcmp;
4626 txn->mt_dbxs[dbi].md_cmp = memncmp;
4628 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4629 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP) {
4630 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED)
4631 txn->mt_dbxs[dbi].md_dcmp = intcmp;
4633 txn->mt_dbxs[dbi].md_dcmp = cintcmp;
4634 } else if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEDUP) {
4635 txn->mt_dbxs[dbi].md_dcmp = memnrcmp;
4637 txn->mt_dbxs[dbi].md_dcmp = memncmp;
4640 txn->mt_dbxs[dbi].md_dcmp = NULL;
4644 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
4651 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
4652 mdb_default_cmp(txn, FREE_DBI);
4658 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
4659 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
4660 mdb_default_cmp(txn, MAIN_DBI);
4664 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
4665 mdb_default_cmp(txn, MAIN_DBI);
4668 /* Is the DB already open? */
4670 for (i=2; i<txn->mt_numdbs; i++) {
4671 if (len == txn->mt_dbxs[i].md_name.mv_size &&
4672 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
4678 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
4681 /* Find the DB info */
4683 key.mv_data = (void *)name;
4684 rc = mdb_get(txn, MAIN_DBI, &key, &data);
4686 /* Create if requested */
4687 if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
4690 data.mv_size = sizeof(MDB_db);
4691 data.mv_data = &dummy;
4692 memset(&dummy, 0, sizeof(dummy));
4693 dummy.md_root = P_INVALID;
4694 dummy.md_flags = flags & 0xffff;
4696 mc.mc_dbi = MAIN_DBI;
4698 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
4702 /* OK, got info, add to table */
4703 if (rc == MDB_SUCCESS) {
4704 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
4705 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
4706 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
4707 txn->mt_dbxs[txn->mt_numdbs].md_parent = MAIN_DBI;
4708 txn->mt_dbxs[txn->mt_numdbs].md_dirty = dirty;
4709 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
4710 *dbi = txn->mt_numdbs;
4711 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
4712 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
4713 mdb_default_cmp(txn, txn->mt_numdbs);
4720 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
4722 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
4725 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
4728 void mdb_close(MDB_txn *txn, MDB_dbi dbi)
4731 if (dbi <= MAIN_DBI || dbi >= txn->mt_numdbs)
4733 ptr = txn->mt_dbxs[dbi].md_name.mv_data;
4734 txn->mt_dbxs[dbi].md_name.mv_data = NULL;
4735 txn->mt_dbxs[dbi].md_name.mv_size = 0;
4739 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
4741 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4744 txn->mt_dbxs[dbi].md_cmp = cmp;
4748 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
4750 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4753 txn->mt_dbxs[dbi].md_dcmp = cmp;
4757 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
4759 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4762 txn->mt_dbxs[dbi].md_rel = rel;