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 GET_PAGESIZE(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 GET_PAGESIZE(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 #GET_PAGESIZE() 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 discarded any data pages can then have those 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 {
472 #define mp_pgno mp_p.p_pgno
473 #define mp_next mp_p.p_next
475 pgno_t p_pgno; /**< page number */
476 void * p_next; /**< for in-memory list of freed structs */
478 #define P_BRANCH 0x01 /**< branch page */
479 #define P_LEAF 0x02 /**< leaf page */
480 #define P_OVERFLOW 0x04 /**< overflow page */
481 #define P_META 0x08 /**< meta page */
482 #define P_DIRTY 0x10 /**< dirty page */
483 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
485 #define mp_lower mp_pb.pb.pb_lower
486 #define mp_upper mp_pb.pb.pb_upper
487 #define mp_pages mp_pb.pb_pages
490 indx_t pb_lower; /**< lower bound of free space */
491 indx_t pb_upper; /**< upper bound of free space */
493 uint32_t pb_pages; /**< number of overflow pages */
495 indx_t mp_ptrs[1]; /**< dynamic size */
498 /** Size of the page header, excluding dynamic data at the end */
499 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
501 /** Address of first usable data byte in a page, after the header */
502 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
504 /** Number of nodes on a page */
505 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
507 /** The amount of space remaining in the page */
508 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
510 /** The percentage of space used in the page, in tenths of a percent. */
511 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
512 ((env)->me_psize - PAGEHDRSZ))
513 /** The minimum page fill factor, in tenths of a percent.
514 * Pages emptier than this are candidates for merging.
516 #define FILL_THRESHOLD 250
518 /** Test if a page is a leaf page */
519 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
520 /** Test if a page is a LEAF2 page */
521 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
522 /** Test if a page is a branch page */
523 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
524 /** Test if a page is an overflow page */
525 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
527 /** The number of overflow pages needed to store the given size. */
528 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
530 /** Header for a single key/data pair within a page.
531 * We guarantee 2-byte alignment for nodes.
533 typedef struct MDB_node {
534 /** lo and hi are used for data size on leaf nodes and for
535 * child pgno on branch nodes. On 64 bit platforms, flags
536 * is also used for pgno. (branch nodes ignore flags)
538 unsigned short mn_lo;
539 unsigned short mn_hi; /**< part of dsize or pgno */
540 unsigned short mn_flags; /**< flags for special node types */
541 #define F_BIGDATA 0x01 /**< data put on overflow page */
542 #define F_SUBDATA 0x02 /**< data is a sub-database */
543 #define F_DUPDATA 0x04 /**< data has duplicates */
544 unsigned short mn_ksize; /**< key size */
545 char mn_data[1]; /**< key and data are appended here */
548 /** Size of the node header, excluding dynamic data at the end */
549 #define NODESIZE offsetof(MDB_node, mn_data)
551 /** Size of a node in a branch page with a given key.
552 * This is just the node header plus the key, there is no data.
554 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
556 /** Size of a node in a leaf page with a given key and data.
557 * This is node header plus key plus data size.
559 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
561 /** Address of node \b i in page \b p */
562 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
564 /** Address of the key for the node */
565 #define NODEKEY(node) (void *)((node)->mn_data)
567 /** Address of the data for a node */
568 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
570 /** Get the page number pointed to by a branch node */
571 #if LONG_MAX == 0x7fffffff
572 #define NODEPGNO(node) ((node)->mn_lo | ((node)->mn_hi << 16))
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;} while(0)
577 #define NODEPGNO(node) ((node)->mn_lo | ((node)->mn_hi << 16) | ((unsigned long)(node)->mn_flags << 32))
578 /** Set the page number in a branch node */
579 #define SETPGNO(node,pgno) do { \
580 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
581 (node)->mn_flags = (pgno) >> 32; } while(0)
584 /** Get the size of the data in a leaf node */
585 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
586 /** Set the size of the data for a leaf node */
587 #define SETDSZ(node,size) do { \
588 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
589 /** The size of a key in a node */
590 #define NODEKSZ(node) ((node)->mn_ksize)
592 /** The address of a key in a LEAF2 page.
593 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
594 * There are no node headers, keys are stored contiguously.
596 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
598 /** Set the \b node's key into \b key, if requested. */
599 #define MDB_SET_KEY(node, key) { if ((key) != NULL) { \
600 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
602 /** Information about a single database in the environment. */
603 typedef struct MDB_db {
604 uint32_t md_pad; /**< also ksize for LEAF2 pages */
605 uint16_t md_flags; /**< @ref mdb_open */
606 uint16_t md_depth; /**< depth of this tree */
607 ULONG md_branch_pages; /**< number of internal pages */
608 ULONG md_leaf_pages; /**< number of leaf pages */
609 ULONG md_overflow_pages; /**< number of overflow pages */
610 ULONG md_entries; /**< number of data items */
611 pgno_t md_root; /**< the root page of this tree */
614 /** Handle for the DB used to track free pages. */
616 /** Handle for the default DB. */
619 /** Meta page content. */
620 typedef struct MDB_meta {
621 /** Stamp identifying this as an MDB data file. It must be set
624 /** Version number of this lock file. Must be set to #MDB_VERSION. */
626 void *mm_address; /**< address for fixed mapping */
627 size_t mm_mapsize; /**< size of mmap region */
628 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
629 /** The size of pages used in this DB */
630 #define mm_psize mm_dbs[0].md_pad
631 /** Any persistent environment flags. @ref mdb_env */
632 #define mm_flags mm_dbs[0].md_flags
633 pgno_t mm_last_pg; /**< last used page in file */
634 ULONG mm_txnid; /**< txnid that committed this page */
637 /** Auxiliary DB info.
638 * The information here is mostly static/read-only. There is
639 * only a single copy of this record in the environment.
640 * The \b md_dirty flag is not read-only, but only a write
641 * transaction can ever update it, and only write transactions
642 * need to worry about it.
644 typedef struct MDB_dbx {
645 MDB_val md_name; /**< name of the database */
646 MDB_cmp_func *md_cmp; /**< function for comparing keys */
647 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
648 MDB_rel_func *md_rel; /**< user relocate function */
649 MDB_dbi md_parent; /**< parent DB of a sub-DB */
650 unsigned int md_dirty; /**< TRUE if DB was written in this txn */
653 /** A database transaction.
654 * Every operation requires a transaction handle.
657 pgno_t mt_next_pgno; /**< next unallocated page */
658 /** The ID of this transaction. IDs are integers incrementing from 1.
659 * Only committed write transactions increment the ID. If a transaction
660 * aborts, the ID may be re-used by the next writer.
663 MDB_env *mt_env; /**< the DB environment */
664 /** The list of pages that became unused during this transaction.
669 ID2L dirty_list; /**< modified pages */
670 MDB_reader *reader; /**< this thread's slot in the reader table */
672 /** Array of records for each DB known in the environment. */
674 /** Array of MDB_db records for each known DB */
676 /** Number of DB records in use. This number only ever increments;
677 * we don't decrement it when individual DB handles are closed.
679 unsigned int mt_numdbs;
681 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
682 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
683 unsigned int mt_flags;
684 /** Tracks which of the two meta pages was used at the start
685 * of this transaction.
687 unsigned int mt_toggle;
690 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
691 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
692 * raise this on a 64 bit machine.
694 #define CURSOR_STACK 32
698 /** Cursors are used for all DB operations */
700 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
701 struct MDB_xcursor *mc_xcursor;
702 /** The transaction that owns this cursor */
704 /** The database handle this cursor operates on */
706 unsigned short mc_snum; /**< number of pushed pages */
707 unsigned short mc_top; /**< index of top page, mc_snum-1 */
708 unsigned int mc_flags;
709 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
710 #define C_EOF 0x02 /**< No more data */
711 #define C_XDIRTY 0x04 /**< @deprecated mc_xcursor needs to be flushed */
712 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
713 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
716 /** Context for sorted-dup records.
717 * We could have gone to a fully recursive design, with arbitrarily
718 * deep nesting of sub-databases. But for now we only handle these
719 * levels - main DB, optional sub-DB, sorted-duplicate DB.
721 typedef struct MDB_xcursor {
722 /** A sub-cursor for traversing the Dup DB */
723 MDB_cursor mx_cursor;
724 /** A fake transaction struct for pointing to our own table
728 /** Our private DB information tables. Slots 0 and 1 are always
729 * copies of the corresponding slots in the main transaction. These
730 * hold the FREEDB and MAINDB, respectively. If the main cursor is
731 * on a sub-database, that will be copied to slot 2, and the duplicate
732 * database info will be in slot 3. If the main cursor is on the MAINDB
733 * then the duplicate DB info will be in slot 2 and slot 3 will be unused.
740 /** A set of pages freed by an earlier transaction. */
741 typedef struct MDB_oldpages {
742 /** Usually we only read one record from the FREEDB at a time, but
743 * in case we read more, this will chain them together.
745 struct MDB_oldpages *mo_next;
746 /** The ID of the transaction in which these pages were freed. */
748 /** An #IDL of the pages */
749 pgno_t mo_pages[1]; /* dynamic */
752 /** The database environment. */
754 HANDLE me_fd; /**< The main data file */
755 HANDLE me_lfd; /**< The lock file */
756 HANDLE me_mfd; /**< just for writing the meta pages */
757 /** Failed to update the meta page. Probably an I/O error. */
758 #define MDB_FATAL_ERROR 0x80000000U
760 uint32_t me_extrapad; /**< unused for now */
761 unsigned int me_maxreaders; /**< size of the reader table */
762 unsigned int me_numdbs; /**< number of DBs opened */
763 unsigned int me_maxdbs; /**< size of the DB table */
764 char *me_path; /**< path to the DB files */
765 char *me_map; /**< the memory map of the data file */
766 MDB_txninfo *me_txns; /**< the memory map of the lock file */
767 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
768 MDB_txn *me_txn; /**< current write transaction */
769 size_t me_mapsize; /**< size of the data memory map */
770 off_t me_size; /**< current file size */
771 pgno_t me_maxpg; /**< me_mapsize / me_psize */
772 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
773 unsigned int me_db_toggle; /**< which DB table is current */
774 MDB_dbx *me_dbxs; /**< array of static DB info */
775 MDB_db *me_dbs[2]; /**< two arrays of MDB_db info */
776 MDB_oldpages *me_pghead; /**< list of old page records */
777 pthread_key_t me_txkey; /**< thread-key for readers */
778 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
779 /** IDL of pages that became unused in a write txn */
780 pgno_t me_free_pgs[MDB_IDL_UM_SIZE];
781 /** ID2L of pages that were written during a write txn */
782 ID2 me_dirty_list[MDB_IDL_UM_SIZE];
783 /** rwlock for the DB tables, if #LAZY_LOCKS is false */
784 LAZY_RWLOCK_DEF(me_dblock);
786 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
790 /** max number of pages to commit in one writev() call */
791 #define MDB_COMMIT_PAGES 64
793 static MDB_page *mdb_alloc_page(MDB_cursor *mc, int num);
794 static int mdb_touch(MDB_cursor *mc);
796 static int mdb_search_page_root(MDB_cursor *mc,
797 MDB_val *key, int modify);
798 static int mdb_search_page(MDB_cursor *mc,
799 MDB_val *key, int modify);
801 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
802 static int mdb_env_read_meta(MDB_env *env, int *which);
803 static int mdb_env_write_meta(MDB_txn *txn);
804 static int mdb_get_page(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
806 static MDB_node *mdb_search_node(MDB_cursor *mc, MDB_val *key, int *exactp);
807 static int mdb_add_node(MDB_cursor *mc, indx_t indx,
808 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags);
809 static void mdb_del_node(MDB_page *mp, indx_t indx, int ksize);
810 static int mdb_del0(MDB_cursor *mc, MDB_node *leaf);
811 static int mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
813 static int mdb_rebalance(MDB_cursor *mc);
814 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
815 static int mdb_move_node(MDB_cursor *csrc, MDB_cursor *cdst);
816 static int mdb_merge(MDB_cursor *csrc, MDB_cursor *cdst);
817 static int mdb_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
819 static MDB_page *mdb_new_page(MDB_cursor *mc, uint32_t flags, int num);
821 static void cursor_pop_page(MDB_cursor *mc);
822 static int cursor_push_page(MDB_cursor *mc, MDB_page *mp);
824 static int mdb_sibling(MDB_cursor *mc, int move_right);
825 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
826 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
827 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
829 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
830 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
832 static void mdb_xcursor_init0(MDB_cursor *mc);
833 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
834 static void mdb_xcursor_init2(MDB_cursor *mc);
835 static void mdb_xcursor_fini(MDB_cursor *mc);
837 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
838 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
840 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
843 static MDB_cmp_func memncmp, memnrcmp, intcmp, cintcmp;
847 static SECURITY_DESCRIPTOR mdb_null_sd;
848 static SECURITY_ATTRIBUTES mdb_all_sa;
849 static int mdb_sec_inited;
852 /** Return the library version info. */
854 mdb_version(int *major, int *minor, int *patch)
856 if (major) *major = MDB_VERSION_MAJOR;
857 if (minor) *minor = MDB_VERSION_MINOR;
858 if (patch) *patch = MDB_VERSION_PATCH;
859 return MDB_VERSION_STRING;
862 /** Table of descriptions for MDB @ref errors */
863 static char *const mdb_errstr[] = {
864 "MDB_KEYEXIST: Key/data pair already exists",
865 "MDB_NOTFOUND: No matching key/data pair found",
866 "MDB_PAGE_NOTFOUND: Requested page not found",
867 "MDB_CORRUPTED: Located page was wrong type",
868 "MDB_PANIC: Update of meta page failed",
869 "MDB_VERSION_MISMATCH: Database environment version mismatch"
873 mdb_strerror(int err)
876 return ("Successful return: 0");
878 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
879 return mdb_errstr[err - MDB_KEYEXIST];
881 return strerror(err);
885 /** Display a key in hexadecimal and return the address of the result.
886 * @param[in] key the key to display
887 * @param[in] buf the buffer to write into. Should always be #DKBUF.
888 * @return The key in hexadecimal form.
891 mdb_dkey(MDB_val *key, char *buf)
894 unsigned char *c = key->mv_data;
896 if (key->mv_size > MAXKEYSIZE)
898 /* may want to make this a dynamic check: if the key is mostly
899 * printable characters, print it as-is instead of converting to hex.
902 for (i=0; i<key->mv_size; i++)
903 ptr += sprintf(ptr, "%02x", *c++);
905 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
912 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
914 return txn->mt_dbxs[dbi].md_cmp(a, b);
917 /** Compare two data items according to a particular database.
918 * This returns a comparison as if the two items were data items of
919 * a sorted duplicates #MDB_DUPSORT database.
920 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
921 * @param[in] dbi A database handle returned by #mdb_open()
922 * @param[in] a The first item to compare
923 * @param[in] b The second item to compare
924 * @return < 0 if a < b, 0 if a == b, > 0 if a > b
927 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
929 if (txn->mt_dbxs[dbi].md_dcmp)
930 return txn->mt_dbxs[dbi].md_dcmp(a, b);
932 return EINVAL; /* too bad you can't distinguish this from a valid result */
935 /** Allocate pages for writing.
936 * If there are free pages available from older transactions, they
937 * will be re-used first. Otherwise a new page will be allocated.
938 * @param[in] mc cursor A cursor handle identifying the transaction and
939 * database for which we are allocating.
940 * @param[in] num the number of pages to allocate.
941 * @return Address of the allocated page(s). Requests for multiple pages
942 * will always be satisfied by a single contiguous chunk of memory.
945 mdb_alloc_page(MDB_cursor *mc, int num)
947 MDB_txn *txn = mc->mc_txn;
949 pgno_t pgno = P_INVALID;
952 if (txn->mt_txnid > 2) {
954 if (!txn->mt_env->me_pghead && mc->mc_dbi != FREE_DBI &&
955 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
956 /* See if there's anything in the free DB */
962 m2.mc_dbi = FREE_DBI;
965 mdb_search_page(&m2, NULL, 0);
966 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
967 kptr = (ULONG *)NODEKEY(leaf);
971 oldest = txn->mt_txnid - 1;
972 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
973 ULONG mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
974 if (mr && mr < oldest)
979 if (oldest > *kptr) {
980 /* It's usable, grab it.
986 mdb_read_data(txn, leaf, &data);
987 idl = (ULONG *)data.mv_data;
988 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
989 mop->mo_next = txn->mt_env->me_pghead;
990 mop->mo_txnid = *kptr;
991 txn->mt_env->me_pghead = mop;
992 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
997 DPRINTF("IDL read txn %lu root %lu num %lu",
998 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
999 for (i=0; i<idl[0]; i++) {
1000 DPRINTF("IDL %lu", idl[i+1]);
1004 /* drop this IDL from the DB */
1005 m2.mc_ki[m2.mc_top] = 0;
1006 m2.mc_flags = C_INITIALIZED;
1007 mdb_cursor_del(&m2, 0);
1010 if (txn->mt_env->me_pghead) {
1011 MDB_oldpages *mop = txn->mt_env->me_pghead;
1013 /* FIXME: For now, always use fresh pages. We
1014 * really ought to search the free list for a
1019 /* peel pages off tail, so we only have to truncate the list */
1020 pgno = MDB_IDL_LAST(mop->mo_pages);
1021 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1023 if (mop->mo_pages[2] > mop->mo_pages[1])
1024 mop->mo_pages[0] = 0;
1028 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1029 txn->mt_env->me_pghead = mop->mo_next;
1036 if (pgno == P_INVALID) {
1037 /* DB size is maxed out */
1038 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg)
1041 if (txn->mt_env->me_dpages && num == 1) {
1042 np = txn->mt_env->me_dpages;
1043 txn->mt_env->me_dpages = np->mp_next;
1045 if ((np = malloc(txn->mt_env->me_psize * num )) == NULL)
1048 if (pgno == P_INVALID) {
1049 np->mp_pgno = txn->mt_next_pgno;
1050 txn->mt_next_pgno += num;
1054 mid.mid = np->mp_pgno;
1056 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1061 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1062 * @param[in] mc cursor pointing to the page to be touched
1063 * @return 0 on success, non-zero on failure.
1066 mdb_touch(MDB_cursor *mc)
1068 MDB_page *mp = mc->mc_pg[mc->mc_top];
1071 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1073 if ((np = mdb_alloc_page(mc, 1)) == NULL)
1075 DPRINTF("touched db %u page %lu -> %lu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1076 assert(mp->mp_pgno != np->mp_pgno);
1077 mdb_midl_append(mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1079 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1082 mp->mp_flags |= P_DIRTY;
1084 mc->mc_pg[mc->mc_top] = mp;
1085 /** If this page has a parent, update the parent to point to
1089 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1095 mdb_env_sync(MDB_env *env, int force)
1098 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1099 if (fdatasync(env->me_fd))
1106 mdb_txn_reset0(MDB_txn *txn);
1108 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1109 * @param[in] txn the transaction handle to initialize
1110 * @return 0 on success, non-zero on failure. This can only
1111 * fail for read-only transactions, and then only if the
1112 * reader table is full.
1115 mdb_txn_renew0(MDB_txn *txn)
1117 MDB_env *env = txn->mt_env;
1119 if (txn->mt_flags & MDB_TXN_RDONLY) {
1120 MDB_reader *r = pthread_getspecific(env->me_txkey);
1123 pid_t pid = getpid();
1124 pthread_t tid = pthread_self();
1127 for (i=0; i<env->me_txns->mti_numreaders; i++)
1128 if (env->me_txns->mti_readers[i].mr_pid == 0)
1130 if (i == env->me_maxreaders) {
1131 UNLOCK_MUTEX_R(env);
1134 env->me_txns->mti_readers[i].mr_pid = pid;
1135 env->me_txns->mti_readers[i].mr_tid = tid;
1136 if (i >= env->me_txns->mti_numreaders)
1137 env->me_txns->mti_numreaders = i+1;
1138 UNLOCK_MUTEX_R(env);
1139 r = &env->me_txns->mti_readers[i];
1140 pthread_setspecific(env->me_txkey, r);
1142 txn->mt_txnid = env->me_txns->mti_txnid;
1143 txn->mt_toggle = env->me_txns->mti_me_toggle;
1144 r->mr_txnid = txn->mt_txnid;
1145 txn->mt_u.reader = r;
1149 txn->mt_txnid = env->me_txns->mti_txnid+1;
1150 txn->mt_toggle = env->me_txns->mti_me_toggle;
1151 txn->mt_u.dirty_list = env->me_dirty_list;
1152 txn->mt_u.dirty_list[0].mid = 0;
1153 txn->mt_free_pgs = env->me_free_pgs;
1154 txn->mt_free_pgs[0] = 0;
1155 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1159 /* Copy the DB arrays */
1160 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
1161 txn->mt_numdbs = env->me_numdbs;
1162 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1163 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1164 if (txn->mt_numdbs > 2)
1165 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
1166 (txn->mt_numdbs - 2) * sizeof(MDB_db));
1167 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1173 mdb_txn_renew(MDB_txn *txn)
1180 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1181 DPUTS("environment had fatal error, must shutdown!");
1185 rc = mdb_txn_renew0(txn);
1186 if (rc == MDB_SUCCESS) {
1187 DPRINTF("renew txn %lu%c %p on mdbenv %p, root page %lu",
1188 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1189 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1195 mdb_txn_begin(MDB_env *env, unsigned int flags, MDB_txn **ret)
1200 if (env->me_flags & MDB_FATAL_ERROR) {
1201 DPUTS("environment had fatal error, must shutdown!");
1204 if ((txn = calloc(1, sizeof(MDB_txn) + env->me_maxdbs * sizeof(MDB_db))) == NULL) {
1205 DPRINTF("calloc: %s", strerror(ErrCode()));
1208 txn->mt_dbs = (MDB_db *)(txn+1);
1209 if (flags & MDB_RDONLY) {
1210 txn->mt_flags |= MDB_TXN_RDONLY;
1214 rc = mdb_txn_renew0(txn);
1219 DPRINTF("begin txn %lu%c %p on mdbenv %p, root page %lu",
1220 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1221 (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1227 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1228 * @param[in] txn the transaction handle to reset
1231 mdb_txn_reset0(MDB_txn *txn)
1233 MDB_env *env = txn->mt_env;
1235 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1236 txn->mt_u.reader->mr_txnid = 0;
1242 /* return all dirty pages to dpage list */
1243 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1244 dp = txn->mt_u.dirty_list[i].mptr;
1245 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1246 dp->mp_next = txn->mt_env->me_dpages;
1247 txn->mt_env->me_dpages = dp;
1249 /* large pages just get freed directly */
1254 while ((mop = txn->mt_env->me_pghead)) {
1255 txn->mt_env->me_pghead = mop->mo_next;
1260 for (i=2; i<env->me_numdbs; i++)
1261 env->me_dbxs[i].md_dirty = 0;
1262 /* The writer mutex was locked in mdb_txn_begin. */
1263 UNLOCK_MUTEX_W(env);
1268 mdb_txn_reset(MDB_txn *txn)
1273 DPRINTF("reset txn %lu%c %p on mdbenv %p, root page %lu",
1274 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1275 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1277 mdb_txn_reset0(txn);
1281 mdb_txn_abort(MDB_txn *txn)
1286 DPRINTF("abort txn %lu%c %p on mdbenv %p, root page %lu",
1287 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1288 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1290 mdb_txn_reset0(txn);
1295 mdb_txn_commit(MDB_txn *txn)
1306 assert(txn != NULL);
1307 assert(txn->mt_env != NULL);
1311 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1316 if (txn != env->me_txn) {
1317 DPUTS("attempt to commit unknown transaction");
1322 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1323 DPUTS("error flag is set, can't commit");
1328 if (!txn->mt_u.dirty_list[0].mid)
1331 DPRINTF("committing txn %lu %p on mdbenv %p, root page %lu",
1332 txn->mt_txnid, txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1335 mc.mc_dbi = FREE_DBI;
1338 /* should only be one record now */
1339 if (env->me_pghead) {
1340 /* make sure first page of freeDB is touched and on freelist */
1341 mdb_search_page(&mc, NULL, 1);
1343 /* save to free list */
1344 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1348 /* make sure last page of freeDB is touched and on freelist */
1349 key.mv_size = MAXKEYSIZE+1;
1351 mdb_search_page(&mc, &key, 1);
1353 mdb_midl_sort(txn->mt_free_pgs);
1357 ULONG *idl = txn->mt_free_pgs;
1358 DPRINTF("IDL write txn %lu root %lu num %lu",
1359 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1360 for (i=0; i<idl[0]; i++) {
1361 DPRINTF("IDL %lu", idl[i+1]);
1365 /* write to last page of freeDB */
1366 key.mv_size = sizeof(pgno_t);
1367 key.mv_data = (char *)&txn->mt_txnid;
1368 data.mv_data = txn->mt_free_pgs;
1369 /* The free list can still grow during this call,
1370 * despite the pre-emptive touches above. So check
1371 * and make sure the entire thing got written.
1374 i = txn->mt_free_pgs[0];
1375 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1376 rc = mdb_cursor_put(&mc, &key, &data, 0);
1381 } while (i != txn->mt_free_pgs[0]);
1383 /* should only be one record now */
1384 if (env->me_pghead) {
1388 mop = env->me_pghead;
1389 key.mv_size = sizeof(pgno_t);
1390 key.mv_data = (char *)&mop->mo_txnid;
1391 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1392 data.mv_data = mop->mo_pages;
1393 mdb_cursor_put(&mc, &key, &data, 0);
1394 free(env->me_pghead);
1395 env->me_pghead = NULL;
1398 /* Update DB root pointers. Their pages have already been
1399 * touched so this is all in-place and cannot fail.
1403 data.mv_size = sizeof(MDB_db);
1405 mc.mc_dbi = MAIN_DBI;
1407 for (i = 2; i < txn->mt_numdbs; i++) {
1408 if (txn->mt_dbxs[i].md_dirty) {
1409 data.mv_data = &txn->mt_dbs[i];
1410 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1415 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
1421 /* Windows actually supports scatter/gather I/O, but only on
1422 * unbuffered file handles. Since we're relying on the OS page
1423 * cache for all our data, that's self-defeating. So we just
1424 * write pages one at a time. We use the ov structure to set
1425 * the write offset, to at least save the overhead of a Seek
1429 memset(&ov, 0, sizeof(ov));
1430 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1432 dp = txn->mt_u.dirty_list[i].mptr;
1433 DPRINTF("committing page %lu", dp->mp_pgno);
1434 size = dp->mp_pgno * env->me_psize;
1435 ov.Offset = size & 0xffffffff;
1436 ov.OffsetHigh = size >> 16;
1437 ov.OffsetHigh >>= 16;
1438 /* clear dirty flag */
1439 dp->mp_flags &= ~P_DIRTY;
1440 wsize = env->me_psize;
1441 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
1442 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
1445 DPRINTF("WriteFile: %d", n);
1452 struct iovec iov[MDB_COMMIT_PAGES];
1456 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1457 dp = txn->mt_u.dirty_list[i].mptr;
1458 if (dp->mp_pgno != next) {
1460 DPRINTF("committing %u dirty pages", n);
1461 rc = writev(env->me_fd, iov, n);
1465 DPUTS("short write, filesystem full?");
1467 DPRINTF("writev: %s", strerror(n));
1474 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
1477 DPRINTF("committing page %lu", dp->mp_pgno);
1478 iov[n].iov_len = env->me_psize;
1479 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
1480 iov[n].iov_base = dp;
1481 size += iov[n].iov_len;
1482 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
1483 /* clear dirty flag */
1484 dp->mp_flags &= ~P_DIRTY;
1485 if (++n >= MDB_COMMIT_PAGES) {
1495 DPRINTF("committing %u dirty pages", n);
1496 rc = writev(env->me_fd, iov, n);
1500 DPUTS("short write, filesystem full?");
1502 DPRINTF("writev: %s", strerror(n));
1509 /* Drop the dirty pages.
1511 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1512 dp = txn->mt_u.dirty_list[i].mptr;
1513 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1514 dp->mp_next = txn->mt_env->me_dpages;
1515 txn->mt_env->me_dpages = dp;
1519 txn->mt_u.dirty_list[i].mid = 0;
1521 txn->mt_u.dirty_list[0].mid = 0;
1523 if ((n = mdb_env_sync(env, 0)) != 0 ||
1524 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
1531 /* update the DB tables */
1533 int toggle = !env->me_db_toggle;
1536 ip = &env->me_dbs[toggle][2];
1537 jp = &txn->mt_dbs[2];
1538 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1539 for (i = 2; i < txn->mt_numdbs; i++) {
1540 if (ip->md_root != jp->md_root)
1545 for (i = 2; i < txn->mt_numdbs; i++) {
1546 if (txn->mt_dbxs[i].md_dirty)
1547 txn->mt_dbxs[i].md_dirty = 0;
1549 env->me_db_toggle = toggle;
1550 env->me_numdbs = txn->mt_numdbs;
1551 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1554 UNLOCK_MUTEX_W(env);
1560 /** Read the environment parameters of a DB environment before
1561 * mapping it into memory.
1562 * @param[in] env the environment handle
1563 * @param[out] meta address of where to store the meta information
1564 * @return 0 on success, non-zero on failure.
1567 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
1569 char page[PAGESIZE];
1574 /* We don't know the page size yet, so use a minimum value.
1578 if (!ReadFile(env->me_fd, page, PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
1580 if ((rc = read(env->me_fd, page, PAGESIZE)) == 0)
1585 else if (rc != PAGESIZE) {
1589 DPRINTF("read: %s", strerror(err));
1593 p = (MDB_page *)page;
1595 if (!F_ISSET(p->mp_flags, P_META)) {
1596 DPRINTF("page %lu not a meta page", p->mp_pgno);
1601 if (m->mm_magic != MDB_MAGIC) {
1602 DPUTS("meta has invalid magic");
1606 if (m->mm_version != MDB_VERSION) {
1607 DPRINTF("database is version %u, expected version %u",
1608 m->mm_version, MDB_VERSION);
1609 return MDB_VERSION_MISMATCH;
1612 memcpy(meta, m, sizeof(*m));
1616 /** Write the environment parameters of a freshly created DB environment.
1617 * @param[in] env the environment handle
1618 * @param[out] meta address of where to store the meta information
1619 * @return 0 on success, non-zero on failure.
1622 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
1629 DPUTS("writing new meta page");
1631 GET_PAGESIZE(psize);
1633 meta->mm_magic = MDB_MAGIC;
1634 meta->mm_version = MDB_VERSION;
1635 meta->mm_psize = psize;
1636 meta->mm_last_pg = 1;
1637 meta->mm_flags = env->me_flags & 0xffff;
1638 meta->mm_flags |= MDB_INTEGERKEY;
1639 meta->mm_dbs[0].md_root = P_INVALID;
1640 meta->mm_dbs[1].md_root = P_INVALID;
1642 p = calloc(2, psize);
1644 p->mp_flags = P_META;
1647 memcpy(m, meta, sizeof(*meta));
1649 q = (MDB_page *)((char *)p + psize);
1652 q->mp_flags = P_META;
1655 memcpy(m, meta, sizeof(*meta));
1660 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
1661 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
1664 rc = write(env->me_fd, p, psize * 2);
1665 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
1671 /** Update the environment info to commit a transaction.
1672 * @param[in] txn the transaction that's being committed
1673 * @return 0 on success, non-zero on failure.
1676 mdb_env_write_meta(MDB_txn *txn)
1679 MDB_meta meta, metab;
1681 int rc, len, toggle;
1687 assert(txn != NULL);
1688 assert(txn->mt_env != NULL);
1690 toggle = !txn->mt_toggle;
1691 DPRINTF("writing meta page %d for root page %lu",
1692 toggle, txn->mt_dbs[MAIN_DBI].md_root);
1696 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
1697 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
1699 ptr = (char *)&meta;
1700 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
1701 len = sizeof(MDB_meta) - off;
1704 meta.mm_dbs[0] = txn->mt_dbs[0];
1705 meta.mm_dbs[1] = txn->mt_dbs[1];
1706 meta.mm_last_pg = txn->mt_next_pgno - 1;
1707 meta.mm_txnid = txn->mt_txnid;
1710 off += env->me_psize;
1713 /* Write to the SYNC fd */
1716 memset(&ov, 0, sizeof(ov));
1718 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
1721 rc = pwrite(env->me_mfd, ptr, len, off);
1726 DPUTS("write failed, disk error?");
1727 /* On a failure, the pagecache still contains the new data.
1728 * Write some old data back, to prevent it from being used.
1729 * Use the non-SYNC fd; we know it will fail anyway.
1731 meta.mm_last_pg = metab.mm_last_pg;
1732 meta.mm_txnid = metab.mm_txnid;
1734 WriteFile(env->me_fd, ptr, len, NULL, &ov);
1736 r2 = pwrite(env->me_fd, ptr, len, off);
1738 env->me_flags |= MDB_FATAL_ERROR;
1741 /* Memory ordering issues are irrelevant; since the entire writer
1742 * is wrapped by wmutex, all of these changes will become visible
1743 * after the wmutex is unlocked. Since the DB is multi-version,
1744 * readers will get consistent data regardless of how fresh or
1745 * how stale their view of these values is.
1747 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
1748 txn->mt_env->me_txns->mti_me_toggle = toggle;
1749 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
1750 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
1755 /** Check both meta pages to see which one is newer.
1756 * @param[in] env the environment handle
1757 * @param[out] which address of where to store the meta toggle ID
1758 * @return 0 on success, non-zero on failure.
1761 mdb_env_read_meta(MDB_env *env, int *which)
1765 assert(env != NULL);
1767 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1770 DPRINTF("Using meta page %d", toggle);
1777 mdb_env_create(MDB_env **env)
1781 e = calloc(1, sizeof(MDB_env));
1782 if (!e) return ENOMEM;
1784 e->me_maxreaders = DEFAULT_READERS;
1786 e->me_fd = INVALID_HANDLE_VALUE;
1787 e->me_lfd = INVALID_HANDLE_VALUE;
1788 e->me_mfd = INVALID_HANDLE_VALUE;
1794 mdb_env_set_mapsize(MDB_env *env, size_t size)
1798 env->me_mapsize = size;
1803 mdb_env_set_maxdbs(MDB_env *env, int dbs)
1807 env->me_maxdbs = dbs;
1812 mdb_env_set_maxreaders(MDB_env *env, int readers)
1816 env->me_maxreaders = readers;
1821 mdb_env_get_maxreaders(MDB_env *env, int *readers)
1823 if (!env || !readers)
1825 *readers = env->me_maxreaders;
1829 /** Further setup required for opening an MDB environment
1832 mdb_env_open2(MDB_env *env, unsigned int flags)
1834 int i, newenv = 0, toggle;
1838 env->me_flags = flags;
1840 memset(&meta, 0, sizeof(meta));
1842 if ((i = mdb_env_read_header(env, &meta)) != 0) {
1845 DPUTS("new mdbenv");
1849 if (!env->me_mapsize) {
1850 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
1856 LONG sizelo, sizehi;
1857 sizelo = env->me_mapsize & 0xffffffff;
1858 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
1860 /* Windows won't create mappings for zero length files.
1861 * Just allocate the maxsize right now.
1864 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
1865 if (!SetEndOfFile(env->me_fd))
1867 SetFilePointer(env->me_fd, 0, NULL, 0);
1869 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
1870 sizehi, sizelo, NULL);
1873 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
1881 if (meta.mm_address && (flags & MDB_FIXEDMAP))
1883 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
1885 if (env->me_map == MAP_FAILED)
1890 meta.mm_mapsize = env->me_mapsize;
1891 if (flags & MDB_FIXEDMAP)
1892 meta.mm_address = env->me_map;
1893 i = mdb_env_init_meta(env, &meta);
1894 if (i != MDB_SUCCESS) {
1895 munmap(env->me_map, env->me_mapsize);
1899 env->me_psize = meta.mm_psize;
1901 env->me_maxpg = env->me_mapsize / env->me_psize;
1903 p = (MDB_page *)env->me_map;
1904 env->me_metas[0] = METADATA(p);
1905 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
1907 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
1910 DPRINTF("opened database version %u, pagesize %u",
1911 env->me_metas[toggle]->mm_version, env->me_psize);
1912 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
1913 DPRINTF("entries: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
1914 DPRINTF("branch pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
1915 DPRINTF("leaf pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
1916 DPRINTF("overflow pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
1917 DPRINTF("root: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
1923 /* Windows doesn't support destructor callbacks for thread-specific storage */
1925 mdb_env_reader_dest(void *ptr)
1927 MDB_reader *reader = ptr;
1929 reader->mr_txnid = 0;
1935 /* downgrade the exclusive lock on the region back to shared */
1937 mdb_env_share_locks(MDB_env *env)
1941 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1943 env->me_txns->mti_me_toggle = toggle;
1944 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
1949 /* First acquire a shared lock. The Unlock will
1950 * then release the existing exclusive lock.
1952 memset(&ov, 0, sizeof(ov));
1953 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
1954 UnlockFile(env->me_lfd, 0, 0, 1, 0);
1958 struct flock lock_info;
1959 /* The shared lock replaces the existing lock */
1960 memset((void *)&lock_info, 0, sizeof(lock_info));
1961 lock_info.l_type = F_RDLCK;
1962 lock_info.l_whence = SEEK_SET;
1963 lock_info.l_start = 0;
1964 lock_info.l_len = 1;
1965 fcntl(env->me_lfd, F_SETLK, &lock_info);
1971 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
1979 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
1980 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
1981 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
1985 /* Try to get exclusive lock. If we succeed, then
1986 * nobody is using the lock region and we should initialize it.
1989 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
1993 memset(&ov, 0, sizeof(ov));
1994 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2000 size = GetFileSize(env->me_lfd, NULL);
2002 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2006 /* Try to get exclusive lock. If we succeed, then
2007 * nobody is using the lock region and we should initialize it.
2010 struct flock lock_info;
2011 memset((void *)&lock_info, 0, sizeof(lock_info));
2012 lock_info.l_type = F_WRLCK;
2013 lock_info.l_whence = SEEK_SET;
2014 lock_info.l_start = 0;
2015 lock_info.l_len = 1;
2016 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2020 lock_info.l_type = F_RDLCK;
2021 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2028 size = lseek(env->me_lfd, 0, SEEK_END);
2030 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2031 if (size < rsize && *excl) {
2033 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2034 if (!SetEndOfFile(env->me_lfd)) {
2039 if (ftruncate(env->me_lfd, rsize) != 0) {
2046 size = rsize - sizeof(MDB_txninfo);
2047 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2052 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2058 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2060 if (!env->me_txns) {
2066 env->me_txns = mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2068 if (env->me_txns == MAP_FAILED) {
2076 if (!mdb_sec_inited) {
2077 InitializeSecurityDescriptor(&mdb_null_sd,
2078 SECURITY_DESCRIPTOR_REVISION);
2079 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2080 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2081 mdb_all_sa.bInheritHandle = FALSE;
2082 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2085 /* FIXME: only using up to 20 characters of the env path here,
2086 * probably not enough to assure uniqueness...
2088 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%.20s", lpath);
2089 ptr = env->me_txns->mti_rmname + sizeof("Global\\MDBr");
2090 while ((ptr = strchr(ptr, '\\')))
2092 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2093 if (!env->me_rmutex) {
2097 sprintf(env->me_txns->mti_rmname, "Global\\MDBw%.20s", lpath);
2098 ptr = env->me_txns->mti_rmname + sizeof("Global\\MDBw");
2099 while ((ptr = strchr(ptr, '\\')))
2101 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2102 if (!env->me_wmutex) {
2107 pthread_mutexattr_t mattr;
2109 pthread_mutexattr_init(&mattr);
2110 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2114 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2115 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2117 env->me_txns->mti_version = MDB_VERSION;
2118 env->me_txns->mti_magic = MDB_MAGIC;
2119 env->me_txns->mti_txnid = 0;
2120 env->me_txns->mti_numreaders = 0;
2121 env->me_txns->mti_me_toggle = 0;
2124 if (env->me_txns->mti_magic != MDB_MAGIC) {
2125 DPUTS("lock region has invalid magic");
2129 if (env->me_txns->mti_version != MDB_VERSION) {
2130 DPRINTF("lock region is version %u, expected version %u",
2131 env->me_txns->mti_version, MDB_VERSION);
2132 rc = MDB_VERSION_MISMATCH;
2136 if (rc != EACCES && rc != EAGAIN) {
2140 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2141 if (!env->me_rmutex) {
2145 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2146 if (!env->me_wmutex) {
2156 env->me_lfd = INVALID_HANDLE_VALUE;
2161 /** The name of the lock file in the DB environment */
2162 #define LOCKNAME "/lock.mdb"
2163 /** The name of the data file in the DB environment */
2164 #define DATANAME "/data.mdb"
2166 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2168 int oflags, rc, len, excl;
2169 char *lpath, *dpath;
2172 lpath = malloc(len + sizeof(LOCKNAME) + len + sizeof(DATANAME));
2175 dpath = lpath + len + sizeof(LOCKNAME);
2176 sprintf(lpath, "%s" LOCKNAME, path);
2177 sprintf(dpath, "%s" DATANAME, path);
2179 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
2184 if (F_ISSET(flags, MDB_RDONLY)) {
2185 oflags = GENERIC_READ;
2186 len = OPEN_EXISTING;
2188 oflags = GENERIC_READ|GENERIC_WRITE;
2191 mode = FILE_ATTRIBUTE_NORMAL;
2192 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2193 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2198 if (F_ISSET(flags, MDB_RDONLY))
2201 oflags = O_RDWR | O_CREAT;
2203 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
2209 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
2210 /* synchronous fd for meta writes */
2212 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2213 mode |= FILE_FLAG_WRITE_THROUGH;
2214 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2215 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2220 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2221 oflags |= MDB_DSYNC;
2222 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
2227 env->me_path = strdup(path);
2228 DPRINTF("opened dbenv %p", (void *) env);
2229 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
2230 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
2232 mdb_env_share_locks(env);
2233 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
2234 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
2235 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
2241 if (env->me_fd != INVALID_HANDLE_VALUE) {
2243 env->me_fd = INVALID_HANDLE_VALUE;
2245 if (env->me_lfd != INVALID_HANDLE_VALUE) {
2247 env->me_lfd = INVALID_HANDLE_VALUE;
2255 mdb_env_close(MDB_env *env)
2262 while (env->me_dpages) {
2263 dp = env->me_dpages;
2264 env->me_dpages = dp->mp_next;
2268 free(env->me_dbs[1]);
2269 free(env->me_dbs[0]);
2273 LAZY_RWLOCK_DESTROY(&env->me_dblock);
2274 pthread_key_delete(env->me_txkey);
2277 munmap(env->me_map, env->me_mapsize);
2282 pid_t pid = getpid();
2284 for (i=0; i<env->me_txns->mti_numreaders; i++)
2285 if (env->me_txns->mti_readers[i].mr_pid == pid)
2286 env->me_txns->mti_readers[i].mr_pid = 0;
2287 munmap(env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
2293 /* only for aligned ints */
2295 intcmp(const MDB_val *a, const MDB_val *b)
2297 if (a->mv_size == sizeof(long))
2299 unsigned long *la, *lb;
2304 unsigned int *ia, *ib;
2311 /* ints must always be the same size */
2313 cintcmp(const MDB_val *a, const MDB_val *b)
2315 #if __BYTE_ORDER == __LITTLE_ENDIAN
2316 unsigned short *u, *c;
2319 u = a->mv_data + a->mv_size;
2320 c = b->mv_data + a->mv_size;
2323 } while(!x && u > (unsigned short *)a->mv_data);
2326 return memcmp(a->mv_data, b->mv_data, a->mv_size);
2331 memncmp(const MDB_val *a, const MDB_val *b)
2337 len_diff = a->mv_size - b->mv_size;
2340 diff = memcmp(a->mv_data, b->mv_data, len);
2341 return diff ? diff : len_diff;
2345 memnrcmp(const MDB_val *a, const MDB_val *b)
2347 const unsigned char *p1, *p2, *p1_lim;
2350 if (b->mv_size == 0)
2351 return a->mv_size != 0;
2352 if (a->mv_size == 0)
2355 p1 = (const unsigned char *)a->mv_data + a->mv_size - 1;
2356 p2 = (const unsigned char *)b->mv_data + b->mv_size - 1;
2358 len_diff = a->mv_size - b->mv_size;
2360 p1_lim = p1 - a->mv_size;
2362 p1_lim = p1 - b->mv_size;
2364 while (p1 > p1_lim) {
2374 /* Search for key within a leaf page, using binary search.
2375 * Returns the smallest entry larger or equal to the key.
2376 * If exactp is non-null, stores whether the found entry was an exact match
2377 * in *exactp (1 or 0).
2378 * If kip is non-null, stores the index of the found entry in *kip.
2379 * If no entry larger or equal to the key is found, returns NULL.
2382 mdb_search_node(MDB_cursor *mc, MDB_val *key, int *exactp)
2384 unsigned int i = 0, nkeys;
2387 MDB_page *mp = mc->mc_pg[mc->mc_top];
2388 MDB_node *node = NULL;
2393 nkeys = NUMKEYS(mp);
2395 DPRINTF("searching %u keys in %s page %lu",
2396 nkeys, IS_LEAF(mp) ? "leaf" : "branch",
2401 low = IS_LEAF(mp) ? 0 : 1;
2403 cmp = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_cmp;
2405 nodekey.mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2406 node = NODEPTR(mp, 0); /* fake */
2408 while (low <= high) {
2409 i = (low + high) >> 1;
2412 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
2414 node = NODEPTR(mp, i);
2416 nodekey.mv_size = node->mn_ksize;
2417 nodekey.mv_data = NODEKEY(node);
2420 rc = cmp(key, &nodekey);
2424 DPRINTF("found leaf index %u [%s], rc = %i",
2425 i, DKEY(&nodekey), rc);
2427 DPRINTF("found branch index %u [%s -> %lu], rc = %i",
2428 i, DKEY(&nodekey), NODEPGNO(node), rc);
2439 if (rc > 0) { /* Found entry is less than the key. */
2440 i++; /* Skip to get the smallest entry larger than key. */
2442 node = NODEPTR(mp, i);
2445 *exactp = (rc == 0);
2446 /* store the key index */
2447 mc->mc_ki[mc->mc_top] = i;
2449 /* There is no entry larger or equal to the key. */
2452 /* nodeptr is fake for LEAF2 */
2457 cursor_pop_page(MDB_cursor *mc)
2462 top = mc->mc_pg[mc->mc_top];
2467 DPRINTF("popped page %lu off db %u cursor %p", top->mp_pgno,
2468 mc->mc_dbi, (void *) mc);
2473 cursor_push_page(MDB_cursor *mc, MDB_page *mp)
2475 DPRINTF("pushing page %lu on db %u cursor %p", mp->mp_pgno,
2476 mc->mc_dbi, (void *) mc);
2478 if (mc->mc_snum >= CURSOR_STACK)
2481 mc->mc_top = mc->mc_snum++;
2482 mc->mc_pg[mc->mc_top] = mp;
2483 mc->mc_ki[mc->mc_top] = 0;
2489 mdb_get_page(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
2493 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
2495 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
2496 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
2497 p = txn->mt_u.dirty_list[x].mptr;
2501 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
2502 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
2506 DPRINTF("page %lu not found", pgno);
2509 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
2513 mdb_search_page_root(MDB_cursor *mc, MDB_val *key, int modify)
2515 MDB_page *mp = mc->mc_pg[mc->mc_top];
2520 while (IS_BRANCH(mp)) {
2523 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
2524 assert(NUMKEYS(mp) > 1);
2525 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
2527 if (key == NULL) /* Initialize cursor to first page. */
2528 mc->mc_ki[mc->mc_top] = 0;
2529 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
2530 /* cursor to last page */
2531 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
2534 node = mdb_search_node(mc, key, &exact);
2536 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
2538 assert(mc->mc_ki[mc->mc_top] > 0);
2539 mc->mc_ki[mc->mc_top]--;
2544 DPRINTF("following index %u for key [%s]",
2545 mc->mc_ki[mc->mc_top], DKEY(key));
2546 assert(mc->mc_ki[mc->mc_top] < NUMKEYS(mp));
2547 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2549 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mp)))
2552 if ((rc = cursor_push_page(mc, mp)))
2556 if ((rc = mdb_touch(mc)) != 0)
2558 mp = mc->mc_pg[mc->mc_top];
2563 DPRINTF("internal error, index points to a %02X page!?",
2565 return MDB_CORRUPTED;
2568 DPRINTF("found leaf page %lu for key [%s]", mp->mp_pgno,
2569 key ? DKEY(key) : NULL);
2574 /* Search for the page a given key should be in.
2575 * Pushes parent pages on the cursor stack.
2576 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
2577 * If modify is true, visited pages are updated with new page numbers.
2580 mdb_search_page(MDB_cursor *mc, MDB_val *key, int modify)
2585 /* Make sure the txn is still viable, then find the root from
2586 * the txn's db table.
2588 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
2589 DPUTS("transaction has failed, must abort");
2592 root = mc->mc_txn->mt_dbs[mc->mc_dbi].md_root;
2594 if (root == P_INVALID) { /* Tree is empty. */
2595 DPUTS("tree is empty");
2596 return MDB_NOTFOUND;
2599 if ((rc = mdb_get_page(mc->mc_txn, root, &mc->mc_pg[0])))
2605 DPRINTF("db %u root page %lu has flags 0x%X",
2606 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
2609 /* For sub-databases, update main root first */
2610 if (mc->mc_dbi > MAIN_DBI && !mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty) {
2612 mc2.mc_txn = mc->mc_txn;
2613 mc2.mc_dbi = MAIN_DBI;
2614 rc = mdb_search_page(&mc2, &mc->mc_txn->mt_dbxs[mc->mc_dbi].md_name, 1);
2617 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
2619 if (!F_ISSET(mc->mc_pg[0]->mp_flags, P_DIRTY)) {
2620 if ((rc = mdb_touch(mc)))
2622 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = mc->mc_pg[0]->mp_pgno;
2626 return mdb_search_page_root(mc, key, modify);
2630 mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
2632 MDB_page *omp; /* overflow mpage */
2636 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
2637 data->mv_size = NODEDSZ(leaf);
2638 data->mv_data = NODEDATA(leaf);
2642 /* Read overflow data.
2644 data->mv_size = NODEDSZ(leaf);
2645 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
2646 if ((rc = mdb_get_page(txn, pgno, &omp))) {
2647 DPRINTF("read overflow page %lu failed", pgno);
2650 data->mv_data = METADATA(omp);
2656 mdb_get(MDB_txn *txn, MDB_dbi dbi,
2657 MDB_val *key, MDB_val *data)
2666 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
2668 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
2671 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2678 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
2679 mc.mc_xcursor = &mx;
2680 mdb_xcursor_init0(&mc);
2682 mc.mc_xcursor = NULL;
2684 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
2688 mdb_sibling(MDB_cursor *mc, int move_right)
2695 if (mc->mc_snum < 2) {
2696 return MDB_NOTFOUND; /* root has no siblings */
2698 ptop = mc->mc_top-1;
2700 DPRINTF("parent page is page %lu, index %u",
2701 mc->mc_pg[ptop]->mp_pgno, mc->mc_ki[ptop]);
2703 cursor_pop_page(mc);
2704 if (move_right ? (mc->mc_ki[ptop] + 1u >= NUMKEYS(mc->mc_pg[ptop]))
2705 : (mc->mc_ki[ptop] == 0)) {
2706 DPRINTF("no more keys left, moving to %s sibling",
2707 move_right ? "right" : "left");
2708 if ((rc = mdb_sibling(mc, move_right)) != MDB_SUCCESS)
2715 DPRINTF("just moving to %s index key %u",
2716 move_right ? "right" : "left", mc->mc_ki[ptop]);
2718 assert(IS_BRANCH(mc->mc_pg[ptop]));
2720 indx = NODEPTR(mc->mc_pg[ptop], mc->mc_ki[ptop]);
2721 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(indx), &mp)))
2724 cursor_push_page(mc, mp);
2730 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
2736 if (mc->mc_flags & C_EOF) {
2737 return MDB_NOTFOUND;
2740 assert(mc->mc_flags & C_INITIALIZED);
2742 mp = mc->mc_pg[mc->mc_top];
2744 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) {
2745 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2746 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2747 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
2748 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
2749 if (op != MDB_NEXT || rc == MDB_SUCCESS)
2753 mc->mc_xcursor->mx_cursor.mc_flags = 0;
2754 if (op == MDB_NEXT_DUP)
2755 return MDB_NOTFOUND;
2759 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) mc);
2761 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
2762 DPUTS("=====> move to next sibling page");
2763 if (mdb_sibling(mc, 1) != MDB_SUCCESS) {
2764 mc->mc_flags |= C_EOF;
2765 return MDB_NOTFOUND;
2767 mp = mc->mc_pg[mc->mc_top];
2768 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
2770 mc->mc_ki[mc->mc_top]++;
2772 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
2773 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
2776 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2777 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
2781 assert(IS_LEAF(mp));
2782 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2784 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2785 mdb_xcursor_init1(mc, leaf);
2788 if ((rc = mdb_read_data(mc->mc_txn, leaf, data) != MDB_SUCCESS))
2791 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2792 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
2793 if (rc != MDB_SUCCESS)
2798 MDB_SET_KEY(leaf, key);
2803 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
2809 assert(mc->mc_flags & C_INITIALIZED);
2811 mp = mc->mc_pg[mc->mc_top];
2813 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) {
2814 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2815 if (op == MDB_PREV || op == MDB_PREV_DUP) {
2816 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2817 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
2818 if (op != MDB_PREV || rc == MDB_SUCCESS)
2821 mc->mc_xcursor->mx_cursor.mc_flags = 0;
2822 if (op == MDB_PREV_DUP)
2823 return MDB_NOTFOUND;
2828 DPRINTF("cursor_prev: top page is %lu in cursor %p", mp->mp_pgno, (void *) mc);
2830 if (mc->mc_ki[mc->mc_top] == 0) {
2831 DPUTS("=====> move to prev sibling page");
2832 if (mdb_sibling(mc, 0) != MDB_SUCCESS) {
2833 mc->mc_flags &= ~C_INITIALIZED;
2834 return MDB_NOTFOUND;
2836 mp = mc->mc_pg[mc->mc_top];
2837 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
2838 DPRINTF("prev page is %lu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
2840 mc->mc_ki[mc->mc_top]--;
2842 mc->mc_flags &= ~C_EOF;
2844 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
2845 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
2848 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2849 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
2853 assert(IS_LEAF(mp));
2854 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2856 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2857 mdb_xcursor_init1(mc, leaf);
2860 if ((rc = mdb_read_data(mc->mc_txn, leaf, data) != MDB_SUCCESS))
2863 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2864 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
2865 if (rc != MDB_SUCCESS)
2870 MDB_SET_KEY(leaf, key);
2875 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
2876 MDB_cursor_op op, int *exactp)
2884 assert(key->mv_size > 0);
2886 /* See if we're already on the right page */
2887 if (mc->mc_flags & C_INITIALIZED) {
2890 if (mc->mc_pg[mc->mc_top]->mp_flags & P_LEAF2) {
2891 nodekey.mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2892 nodekey.mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, nodekey.mv_size);
2894 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
2895 MDB_SET_KEY(leaf, &nodekey);
2897 rc = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_cmp(key, &nodekey);
2899 /* Probably happens rarely, but first node on the page
2900 * was the one we wanted.
2902 mc->mc_ki[mc->mc_top] = 0;
2906 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
2911 if (NUMKEYS(mc->mc_pg[mc->mc_top]) > 1) {
2912 if (mc->mc_pg[mc->mc_top]->mp_flags & P_LEAF2) {
2913 nodekey.mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top],
2914 NUMKEYS(mc->mc_pg[mc->mc_top])-1, nodekey.mv_size);
2916 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
2917 MDB_SET_KEY(leaf, &nodekey);
2919 rc = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_cmp(key, &nodekey);
2921 /* last node was the one we wanted */
2922 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top])-1;
2926 /* This is definitely the right page, skip search_page */
2931 /* If any parents have right-sibs, search.
2932 * Otherwise, there's nothing further.
2934 for (i=0; i<mc->mc_top; i++)
2936 NUMKEYS(mc->mc_pg[i])-1)
2938 if (i == mc->mc_top) {
2939 /* There are no other pages */
2940 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
2941 return MDB_NOTFOUND;
2946 rc = mdb_search_page(mc, key, 0);
2947 if (rc != MDB_SUCCESS)
2950 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
2953 leaf = mdb_search_node(mc, key, exactp);
2954 if (exactp != NULL && !*exactp) {
2955 /* MDB_SET specified and not an exact match. */
2956 return MDB_NOTFOUND;
2960 DPUTS("===> inexact leaf not found, goto sibling");
2961 if ((rc = mdb_sibling(mc, 1)) != MDB_SUCCESS)
2962 return rc; /* no entries matched */
2963 mc->mc_ki[mc->mc_top] = 0;
2964 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
2965 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
2969 mc->mc_flags |= C_INITIALIZED;
2970 mc->mc_flags &= ~C_EOF;
2972 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
2973 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2974 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
2978 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2979 mdb_xcursor_init1(mc, leaf);
2982 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2983 if (op == MDB_SET || op == MDB_SET_RANGE) {
2984 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
2987 if (op == MDB_GET_BOTH) {
2993 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
2994 if (rc != MDB_SUCCESS)
2997 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
2999 if ((rc = mdb_read_data(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
3001 rc = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dcmp(data, &d2);
3003 if (op == MDB_GET_BOTH || rc > 0)
3004 return MDB_NOTFOUND;
3008 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3013 /* The key already matches in all other cases */
3014 if (op == MDB_SET_RANGE)
3015 MDB_SET_KEY(leaf, key);
3016 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3022 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3027 rc = mdb_search_page(mc, NULL, 0);
3028 if (rc != MDB_SUCCESS)
3030 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3032 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3033 mc->mc_flags |= C_INITIALIZED;
3034 mc->mc_flags &= ~C_EOF;
3036 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3037 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
3038 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
3043 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3044 mdb_xcursor_init1(mc, leaf);
3045 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3050 mc->mc_xcursor->mx_cursor.mc_flags = 0;
3051 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3055 MDB_SET_KEY(leaf, key);
3060 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3066 lkey.mv_size = MAXKEYSIZE+1;
3067 lkey.mv_data = NULL;
3069 rc = mdb_search_page(mc, &lkey, 0);
3070 if (rc != MDB_SUCCESS)
3072 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3074 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
3075 mc->mc_flags |= C_INITIALIZED;
3076 mc->mc_flags &= ~C_EOF;
3078 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
3080 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3081 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
3082 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
3087 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3088 mdb_xcursor_init1(mc, leaf);
3089 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3093 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3098 MDB_SET_KEY(leaf, key);
3103 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3113 case MDB_GET_BOTH_RANGE:
3114 if (data == NULL || mc->mc_xcursor == NULL) {
3121 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3123 } else if (op == MDB_SET_RANGE)
3124 rc = mdb_cursor_set(mc, key, data, op, NULL);
3126 rc = mdb_cursor_set(mc, key, data, op, &exact);
3128 case MDB_GET_MULTIPLE:
3130 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPFIXED) ||
3131 !(mc->mc_flags & C_INITIALIZED)) {
3136 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
3137 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
3140 case MDB_NEXT_MULTIPLE:
3142 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPFIXED)) {
3146 if (!(mc->mc_flags & C_INITIALIZED))
3147 rc = mdb_cursor_first(mc, key, data);
3149 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
3150 if (rc == MDB_SUCCESS) {
3151 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
3154 mx = &mc->mc_xcursor->mx_cursor;
3155 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
3156 mx->mc_txn->mt_dbs[mx->mc_dbi].md_pad;
3157 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
3158 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
3166 case MDB_NEXT_NODUP:
3167 if (!(mc->mc_flags & C_INITIALIZED))
3168 rc = mdb_cursor_first(mc, key, data);
3170 rc = mdb_cursor_next(mc, key, data, op);
3174 case MDB_PREV_NODUP:
3175 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
3176 rc = mdb_cursor_last(mc, key, data);
3178 rc = mdb_cursor_prev(mc, key, data, op);
3181 rc = mdb_cursor_first(mc, key, data);
3185 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) ||
3186 !(mc->mc_flags & C_INITIALIZED) ||
3187 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3191 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3194 rc = mdb_cursor_last(mc, key, data);
3198 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) ||
3199 !(mc->mc_flags & C_INITIALIZED) ||
3200 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3204 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3207 DPRINTF("unhandled/unimplemented cursor operation %u", op);
3216 mdb_cursor_touch(MDB_cursor *mc)
3220 if (mc->mc_dbi > MAIN_DBI && !mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty) {
3222 mc2.mc_txn = mc->mc_txn;
3223 mc2.mc_dbi = MAIN_DBI;
3224 rc = mdb_search_page(&mc2, &mc->mc_txn->mt_dbxs[mc->mc_dbi].md_name, 1);
3226 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
3228 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
3229 if (!F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) {
3233 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root =
3234 mc->mc_pg[mc->mc_top]->mp_pgno;
3238 mc->mc_top = mc->mc_snum-1;
3243 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3247 MDB_val xdata, *rdata, dkey;
3249 char dbuf[PAGESIZE];
3255 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
3258 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
3259 mc->mc_dbi, DKEY(key), key->mv_size, data->mv_size);
3263 if (flags == MDB_CURRENT) {
3264 if (!(mc->mc_flags & C_INITIALIZED))
3267 } else if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_root == P_INVALID) {
3269 /* new database, write a root leaf page */
3270 DPUTS("allocating new root leaf page");
3271 if ((np = mdb_new_page(mc, P_LEAF, 1)) == NULL) {
3275 cursor_push_page(mc, np);
3276 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = np->mp_pgno;
3277 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth++;
3278 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
3279 if ((mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
3281 np->mp_flags |= P_LEAF2;
3282 mc->mc_flags |= C_INITIALIZED;
3288 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
3289 if (flags == MDB_NOOVERWRITE && rc == 0) {
3290 DPRINTF("duplicate key [%s]", DKEY(key));
3292 return MDB_KEYEXIST;
3294 if (rc && rc != MDB_NOTFOUND)
3298 /* Cursor is positioned, now make sure all pages are writable */
3299 rc2 = mdb_cursor_touch(mc);
3300 if (rc2) return rc2;
3303 /* The key already exists */
3304 if (rc == MDB_SUCCESS) {
3305 /* there's only a key anyway, so this is a no-op */
3306 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3307 unsigned int ksize = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
3308 if (key->mv_size != ksize)
3310 if (flags == MDB_CURRENT) {
3311 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
3312 memcpy(ptr, key->mv_data, ksize);
3317 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3320 if (F_ISSET(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags, MDB_DUPSORT)) {
3321 /* Was a single item before, must convert now */
3322 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3323 dkey.mv_size = NODEDSZ(leaf);
3324 dkey.mv_data = dbuf;
3325 memcpy(dbuf, NODEDATA(leaf), dkey.mv_size);
3326 /* data matches, ignore it */
3327 if (!mdb_dcmp(mc->mc_txn, mc->mc_dbi, data, &dkey))
3328 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
3329 memset(&dummy, 0, sizeof(dummy));
3330 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPFIXED) {
3331 dummy.md_pad = data->mv_size;
3332 dummy.md_flags = MDB_DUPFIXED;
3333 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_INTEGERDUP)
3334 dummy.md_flags |= MDB_INTEGERKEY;
3336 dummy.md_root = P_INVALID;
3337 if (dkey.mv_size == sizeof(MDB_db)) {
3338 memcpy(NODEDATA(leaf), &dummy, sizeof(dummy));
3341 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
3344 xdata.mv_size = sizeof(MDB_db);
3345 xdata.mv_data = &dummy;
3346 /* new sub-DB, must fully init xcursor */
3347 if (flags == MDB_CURRENT)
3353 /* same size, just replace it */
3354 if (!F_ISSET(leaf->mn_flags, F_BIGDATA) &&
3355 NODEDSZ(leaf) == data->mv_size) {
3356 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
3359 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
3361 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
3367 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
3368 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
3369 rc = mdb_split(mc, key, rdata, P_INVALID);
3371 /* There is room already in this leaf page. */
3372 rc = mdb_add_node(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, 0);
3375 if (rc != MDB_SUCCESS)
3376 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
3378 /* Remember if we just added a subdatabase */
3379 if (flags & F_SUBDATA) {
3380 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3381 leaf->mn_flags |= F_SUBDATA;
3384 /* Now store the actual data in the child DB. Note that we're
3385 * storing the user data in the keys field, so there are strict
3386 * size limits on dupdata. The actual data fields of the child
3387 * DB are all zero size.
3390 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3392 if (flags == MDB_CURRENT)
3393 mdb_xcursor_init2(mc);
3395 mdb_xcursor_init1(mc, leaf);
3398 if (flags == MDB_NODUPDATA)
3399 flags = MDB_NOOVERWRITE;
3400 /* converted, write the original data first */
3402 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, flags);
3404 leaf->mn_flags |= F_DUPDATA;
3406 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, flags);
3407 mdb_xcursor_fini(mc);
3408 memcpy(NODEDATA(leaf),
3409 &mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi],
3412 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries++;
3419 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
3424 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
3427 if (!mc->mc_flags & C_INITIALIZED)
3430 rc = mdb_cursor_touch(mc);
3433 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3435 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3436 if (flags != MDB_NODUPDATA) {
3437 mdb_xcursor_init2(mc);
3438 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
3439 mdb_xcursor_fini(mc);
3440 /* If sub-DB still has entries, we're done */
3441 if (mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_root
3443 memcpy(NODEDATA(leaf),
3444 &mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi],
3446 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries--;
3449 /* otherwise fall thru and delete the sub-DB */
3452 /* add all the child DB's pages to the free list */
3453 rc = mdb_search_page(&mc->mc_xcursor->mx_cursor, NULL, 0);
3454 if (rc == MDB_SUCCESS) {
3459 mx = &mc->mc_xcursor->mx_cursor;
3460 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries -=
3461 mx->mc_txn->mt_dbs[mx->mc_dbi].md_entries;
3463 cursor_pop_page(mx);
3465 while (mx->mc_snum > 1) {
3466 for (i=0; i<NUMKEYS(mx->mc_pg[mx->mc_top]); i++) {
3468 ni = NODEPTR(mx->mc_pg[mx->mc_top], i);
3471 mdb_midl_append(mc->mc_txn->mt_free_pgs, pg);
3473 rc = mdb_sibling(mx, 1);
3478 mdb_midl_append(mc->mc_txn->mt_free_pgs,
3479 mx->mc_txn->mt_dbs[mx->mc_dbi].md_root);
3483 return mdb_del0(mc, leaf);
3486 /* Allocate a page and initialize it
3489 mdb_new_page(MDB_cursor *mc, uint32_t flags, int num)
3493 if ((np = mdb_alloc_page(mc, num)) == NULL)
3495 DPRINTF("allocated new mpage %lu, page size %u",
3496 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
3497 np->mp_flags = flags | P_DIRTY;
3498 np->mp_lower = PAGEHDRSZ;
3499 np->mp_upper = mc->mc_txn->mt_env->me_psize;
3502 mc->mc_txn->mt_dbs[mc->mc_dbi].md_branch_pages++;
3503 else if (IS_LEAF(np))
3504 mc->mc_txn->mt_dbs[mc->mc_dbi].md_leaf_pages++;
3505 else if (IS_OVERFLOW(np)) {
3506 mc->mc_txn->mt_dbs[mc->mc_dbi].md_overflow_pages += num;
3514 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
3518 sz = LEAFSIZE(key, data);
3519 if (data->mv_size >= env->me_psize / MDB_MINKEYS) {
3520 /* put on overflow page */
3521 sz -= data->mv_size - sizeof(pgno_t);
3525 return sz + sizeof(indx_t);
3529 mdb_branch_size(MDB_env *env, MDB_val *key)
3534 if (sz >= env->me_psize / MDB_MINKEYS) {
3535 /* put on overflow page */
3536 /* not implemented */
3537 /* sz -= key->size - sizeof(pgno_t); */
3540 return sz + sizeof(indx_t);
3544 mdb_add_node(MDB_cursor *mc, indx_t indx,
3545 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
3548 size_t node_size = NODESIZE;
3551 MDB_page *mp = mc->mc_pg[mc->mc_top];
3552 MDB_page *ofp = NULL; /* overflow page */
3555 assert(mp->mp_upper >= mp->mp_lower);
3557 DPRINTF("add to %s page %lu index %i, data size %zu key size %zu [%s]",
3558 IS_LEAF(mp) ? "leaf" : "branch",
3559 mp->mp_pgno, indx, data ? data->mv_size : 0,
3560 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
3563 /* Move higher keys up one slot. */
3564 int ksize = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad, dif;
3565 char *ptr = LEAF2KEY(mp, indx, ksize);
3566 dif = NUMKEYS(mp) - indx;
3568 memmove(ptr+ksize, ptr, dif*ksize);
3569 /* insert new key */
3570 memcpy(ptr, key->mv_data, ksize);
3572 /* Just using these for counting */
3573 mp->mp_lower += sizeof(indx_t);
3574 mp->mp_upper -= ksize - sizeof(indx_t);
3579 node_size += key->mv_size;
3583 if (F_ISSET(flags, F_BIGDATA)) {
3584 /* Data already on overflow page. */
3585 node_size += sizeof(pgno_t);
3586 } else if (data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
3587 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
3588 /* Put data on overflow page. */
3589 DPRINTF("data size is %zu, put on overflow page",
3591 node_size += sizeof(pgno_t);
3592 if ((ofp = mdb_new_page(mc, P_OVERFLOW, ovpages)) == NULL)
3594 DPRINTF("allocated overflow page %lu", ofp->mp_pgno);
3597 node_size += data->mv_size;
3600 node_size += node_size & 1;
3602 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
3603 DPRINTF("not enough room in page %lu, got %u ptrs",
3604 mp->mp_pgno, NUMKEYS(mp));
3605 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
3606 mp->mp_upper - mp->mp_lower);
3607 DPRINTF("node size = %zu", node_size);
3611 /* Move higher pointers up one slot. */
3612 for (i = NUMKEYS(mp); i > indx; i--)
3613 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
3615 /* Adjust free space offsets. */
3616 ofs = mp->mp_upper - node_size;
3617 assert(ofs >= mp->mp_lower + sizeof(indx_t));
3618 mp->mp_ptrs[indx] = ofs;
3620 mp->mp_lower += sizeof(indx_t);
3622 /* Write the node data. */
3623 node = NODEPTR(mp, indx);
3624 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
3625 node->mn_flags = flags;
3627 SETDSZ(node,data->mv_size);
3632 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
3637 if (F_ISSET(flags, F_BIGDATA))
3638 memcpy(node->mn_data + key->mv_size, data->mv_data,
3641 memcpy(node->mn_data + key->mv_size, data->mv_data,
3644 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
3646 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
3654 mdb_del_node(MDB_page *mp, indx_t indx, int ksize)
3657 indx_t i, j, numkeys, ptr;
3661 DPRINTF("delete node %u on %s page %lu", indx,
3662 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
3663 assert(indx < NUMKEYS(mp));
3666 int x = NUMKEYS(mp) - 1 - indx;
3667 base = LEAF2KEY(mp, indx, ksize);
3669 memmove(base, base + ksize, x * ksize);
3670 mp->mp_lower -= sizeof(indx_t);
3671 mp->mp_upper += ksize - sizeof(indx_t);
3675 node = NODEPTR(mp, indx);
3676 sz = NODESIZE + node->mn_ksize;
3678 if (F_ISSET(node->mn_flags, F_BIGDATA))
3679 sz += sizeof(pgno_t);
3681 sz += NODEDSZ(node);
3685 ptr = mp->mp_ptrs[indx];
3686 numkeys = NUMKEYS(mp);
3687 for (i = j = 0; i < numkeys; i++) {
3689 mp->mp_ptrs[j] = mp->mp_ptrs[i];
3690 if (mp->mp_ptrs[i] < ptr)
3691 mp->mp_ptrs[j] += sz;
3696 base = (char *)mp + mp->mp_upper;
3697 memmove(base + sz, base, ptr - mp->mp_upper);
3699 mp->mp_lower -= sizeof(indx_t);
3704 mdb_xcursor_init0(MDB_cursor *mc)
3706 MDB_xcursor *mx = mc->mc_xcursor;
3709 mx->mx_txn = *mc->mc_txn;
3710 mx->mx_txn.mt_dbxs = mx->mx_dbxs;
3711 mx->mx_txn.mt_dbs = mx->mx_dbs;
3712 mx->mx_dbxs[0] = mc->mc_txn->mt_dbxs[0];
3713 mx->mx_dbxs[1] = mc->mc_txn->mt_dbxs[1];
3714 if (mc->mc_dbi > 1) {
3715 mx->mx_dbxs[2] = mc->mc_txn->mt_dbxs[mc->mc_dbi];
3720 mx->mx_dbxs[dbn+1].md_parent = dbn;
3721 mx->mx_dbxs[dbn+1].md_cmp = mx->mx_dbxs[dbn].md_dcmp;
3722 mx->mx_dbxs[dbn+1].md_rel = mx->mx_dbxs[dbn].md_rel;
3723 mx->mx_dbxs[dbn+1].md_dirty = 0;
3724 mx->mx_txn.mt_numdbs = dbn+2;
3725 mx->mx_txn.mt_u = mc->mc_txn->mt_u;
3727 mx->mx_cursor.mc_xcursor = NULL;
3728 mx->mx_cursor.mc_txn = &mx->mx_txn;
3729 mx->mx_cursor.mc_dbi = dbn+1;
3733 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
3735 MDB_db *db = NODEDATA(node);
3736 MDB_xcursor *mx = mc->mc_xcursor;
3738 mx->mx_dbs[0] = mc->mc_txn->mt_dbs[0];
3739 mx->mx_dbs[1] = mc->mc_txn->mt_dbs[1];
3740 if (mc->mc_dbi > 1) {
3741 mx->mx_dbs[2] = mc->mc_txn->mt_dbs[mc->mc_dbi];
3742 mx->mx_dbxs[2].md_dirty = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty;
3747 DPRINTF("Sub-db %u for db %u root page %lu", dbn, mc->mc_dbi, db->md_root);
3748 mx->mx_dbs[dbn] = *db;
3749 if (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY))
3750 mx->mx_dbxs[dbn].md_dirty = 1;
3751 mx->mx_dbxs[dbn].md_name.mv_data = NODEKEY(node);
3752 mx->mx_dbxs[dbn].md_name.mv_size = node->mn_ksize;
3753 mx->mx_txn.mt_next_pgno = mc->mc_txn->mt_next_pgno;
3754 mx->mx_cursor.mc_snum = 0;
3755 mx->mx_cursor.mc_flags = 0;
3759 mdb_xcursor_init2(MDB_cursor *mc)
3761 MDB_xcursor *mx = mc->mc_xcursor;
3763 mx->mx_dbs[0] = mc->mc_txn->mt_dbs[0];
3764 mx->mx_dbs[1] = mc->mc_txn->mt_dbs[1];
3765 if (mc->mc_dbi > 1) {
3766 mx->mx_dbs[2] = mc->mc_txn->mt_dbs[mc->mc_dbi];
3767 mx->mx_dbxs[2].md_dirty = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty;
3772 DPRINTF("Sub-db %u for db %u root page %lu", dbn, mc->mc_dbi,
3773 mx->mx_dbs[dbn].md_root);
3774 mx->mx_txn.mt_next_pgno = mc->mc_txn->mt_next_pgno;
3778 mdb_xcursor_fini(MDB_cursor *mc)
3780 MDB_xcursor *mx = mc->mc_xcursor;
3781 mc->mc_txn->mt_next_pgno = mx->mx_txn.mt_next_pgno;
3782 mc->mc_txn->mt_dbs[0] = mx->mx_dbs[0];
3783 mc->mc_txn->mt_dbs[1] = mx->mx_dbs[1];
3784 if (mc->mc_dbi > 1) {
3785 mc->mc_txn->mt_dbs[mc->mc_dbi] = mx->mx_dbs[2];
3786 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = mx->mx_dbxs[2].md_dirty;
3791 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
3794 size_t size = sizeof(MDB_cursor);
3796 if (txn == NULL || ret == NULL || !dbi || dbi >= txn->mt_numdbs)
3799 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
3800 size += sizeof(MDB_xcursor);
3802 if ((mc = calloc(1, size)) != NULL) {
3805 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
3806 MDB_xcursor *mx = (MDB_xcursor *)(mc + 1);
3807 mc->mc_xcursor = mx;
3808 mdb_xcursor_init0(mc);
3819 /* Return the count of duplicate data items for the current key */
3821 mdb_cursor_count(MDB_cursor *mc, unsigned long *countp)
3825 if (mc == NULL || countp == NULL)
3828 if (!(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT))
3831 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3832 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3835 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
3838 *countp = mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_entries;
3844 mdb_cursor_close(MDB_cursor *mc)
3852 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
3854 indx_t ptr, i, numkeys;
3861 node = NODEPTR(mp, indx);
3862 ptr = mp->mp_ptrs[indx];
3863 DPRINTF("update key %u (ofs %u) [%.*s] to [%s] on page %lu",
3865 (int)node->mn_ksize, (char *)NODEKEY(node),
3869 delta = key->mv_size - node->mn_ksize;
3871 if (delta > 0 && SIZELEFT(mp) < delta) {
3872 DPRINTF("OUCH! Not enough room, delta = %d", delta);
3876 numkeys = NUMKEYS(mp);
3877 for (i = 0; i < numkeys; i++) {
3878 if (mp->mp_ptrs[i] <= ptr)
3879 mp->mp_ptrs[i] -= delta;
3882 base = (char *)mp + mp->mp_upper;
3883 len = ptr - mp->mp_upper + NODESIZE;
3884 memmove(base - delta, base, len);
3885 mp->mp_upper -= delta;
3887 node = NODEPTR(mp, indx);
3888 node->mn_ksize = key->mv_size;
3891 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
3896 /* Move a node from csrc to cdst.
3899 mdb_move_node(MDB_cursor *csrc, MDB_cursor *cdst)
3906 /* Mark src and dst as dirty. */
3907 if ((rc = mdb_touch(csrc)) ||
3908 (rc = mdb_touch(cdst)))
3911 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3912 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
3913 key.mv_size = csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_pad;
3914 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3916 data.mv_data = NULL;
3918 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
3919 unsigned int snum = csrc->mc_snum;
3920 /* must find the lowest key below src */
3921 mdb_search_page_root(csrc, NULL, 0);
3922 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
3923 csrc->mc_snum = snum--;
3924 csrc->mc_top = snum;
3926 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
3928 key.mv_size = NODEKSZ(srcnode);
3929 key.mv_data = NODEKEY(srcnode);
3930 data.mv_size = NODEDSZ(srcnode);
3931 data.mv_data = NODEDATA(srcnode);
3933 DPRINTF("moving %s node %u [%s] on page %lu to node %u on page %lu",
3934 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
3935 csrc->mc_ki[csrc->mc_top],
3937 csrc->mc_pg[csrc->mc_top]->mp_pgno,
3938 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
3940 /* Add the node to the destination page.
3942 rc = mdb_add_node(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, NODEPGNO(srcnode),
3944 if (rc != MDB_SUCCESS)
3947 /* Delete the node from the source page.
3949 mdb_del_node(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3951 /* Update the parent separators.
3953 if (csrc->mc_ki[csrc->mc_top] == 0) {
3954 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
3955 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3956 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3958 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
3959 key.mv_size = NODEKSZ(srcnode);
3960 key.mv_data = NODEKEY(srcnode);
3962 DPRINTF("update separator for source page %lu to [%s]",
3963 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
3964 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
3965 &key)) != MDB_SUCCESS)
3968 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
3970 nullkey.mv_size = 0;
3971 assert(mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey) == MDB_SUCCESS);
3975 if (cdst->mc_ki[cdst->mc_top] == 0) {
3976 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
3977 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3978 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
3980 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
3981 key.mv_size = NODEKSZ(srcnode);
3982 key.mv_data = NODEKEY(srcnode);
3984 DPRINTF("update separator for destination page %lu to [%s]",
3985 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
3986 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
3987 &key)) != MDB_SUCCESS)
3990 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
3992 nullkey.mv_size = 0;
3993 assert(mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey) == MDB_SUCCESS);
4001 mdb_merge(MDB_cursor *csrc, MDB_cursor *cdst)
4008 DPRINTF("merging page %lu into %lu", csrc->mc_pg[csrc->mc_top]->mp_pgno, cdst->mc_pg[cdst->mc_top]->mp_pgno);
4010 assert(csrc->mc_snum > 1); /* can't merge root page */
4011 assert(cdst->mc_snum > 1);
4013 /* Mark dst as dirty. */
4014 if ((rc = mdb_touch(cdst)))
4017 /* Move all nodes from src to dst.
4019 j = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
4020 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
4021 key.mv_size = csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_pad;
4022 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
4023 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
4024 rc = mdb_add_node(cdst, j, &key, NULL, 0, 0);
4025 if (rc != MDB_SUCCESS)
4027 key.mv_data = (char *)key.mv_data + key.mv_size;
4030 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
4031 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
4033 key.mv_size = srcnode->mn_ksize;
4034 key.mv_data = NODEKEY(srcnode);
4035 data.mv_size = NODEDSZ(srcnode);
4036 data.mv_data = NODEDATA(srcnode);
4037 rc = mdb_add_node(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
4038 if (rc != MDB_SUCCESS)
4043 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
4044 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);
4046 /* Unlink the src page from parent and add to free list.
4048 mdb_del_node(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
4049 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
4051 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
4055 mdb_midl_append(csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
4056 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
4057 csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_leaf_pages--;
4059 csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_branch_pages--;
4060 cursor_pop_page(csrc);
4062 return mdb_rebalance(csrc);
4066 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
4070 cdst->mc_txn = csrc->mc_txn;
4071 cdst->mc_dbi = csrc->mc_dbi;
4072 cdst->mc_snum = csrc->mc_snum;
4073 cdst->mc_top = csrc->mc_top;
4074 cdst->mc_flags = csrc->mc_flags;
4076 for (i=0; i<csrc->mc_snum; i++) {
4077 cdst->mc_pg[i] = csrc->mc_pg[i];
4078 cdst->mc_ki[i] = csrc->mc_ki[i];
4083 mdb_rebalance(MDB_cursor *mc)
4091 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
4092 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
4093 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);
4095 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
4096 DPRINTF("no need to rebalance page %lu, above fill threshold",
4097 mc->mc_pg[mc->mc_top]->mp_pgno);
4101 if (mc->mc_snum < 2) {
4102 if (NUMKEYS(mc->mc_pg[mc->mc_top]) == 0) {
4103 DPUTS("tree is completely empty");
4104 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = P_INVALID;
4105 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth = 0;
4106 mc->mc_txn->mt_dbs[mc->mc_dbi].md_leaf_pages = 0;
4107 mdb_midl_append(mc->mc_txn->mt_free_pgs, mc->mc_pg[mc->mc_top]->mp_pgno);
4108 } else if (IS_BRANCH(mc->mc_pg[mc->mc_top]) && NUMKEYS(mc->mc_pg[mc->mc_top]) == 1) {
4109 DPUTS("collapsing root page!");
4110 mdb_midl_append(mc->mc_txn->mt_free_pgs, mc->mc_pg[mc->mc_top]->mp_pgno);
4111 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = NODEPGNO(NODEPTR(mc->mc_pg[mc->mc_top], 0));
4112 if ((rc = mdb_get_page(mc->mc_txn, mc->mc_txn->mt_dbs[mc->mc_dbi].md_root, &root)))
4114 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth--;
4115 mc->mc_txn->mt_dbs[mc->mc_dbi].md_branch_pages--;
4117 DPUTS("root page doesn't need rebalancing");
4121 /* The parent (branch page) must have at least 2 pointers,
4122 * otherwise the tree is invalid.
4124 ptop = mc->mc_top-1;
4125 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
4127 /* Leaf page fill factor is below the threshold.
4128 * Try to move keys from left or right neighbor, or
4129 * merge with a neighbor page.
4134 mdb_cursor_copy(mc, &mn);
4135 mn.mc_xcursor = NULL;
4137 if (mc->mc_ki[ptop] == 0) {
4138 /* We're the leftmost leaf in our parent.
4140 DPUTS("reading right 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] = 0;
4146 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
4148 /* There is at least one neighbor to the left.
4150 DPUTS("reading left neighbor");
4152 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
4153 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
4155 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
4156 mc->mc_ki[mc->mc_top] = 0;
4159 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
4160 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);
4162 /* If the neighbor page is above threshold and has at least two
4163 * keys, move one key from it.
4165 * Otherwise we should try to merge them.
4167 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
4168 return mdb_move_node(&mn, mc);
4169 else { /* FIXME: if (has_enough_room()) */
4170 if (mc->mc_ki[ptop] == 0)
4171 return mdb_merge(&mn, mc);
4173 return mdb_merge(mc, &mn);
4178 mdb_del0(MDB_cursor *mc, MDB_node *leaf)
4182 /* add overflow pages to free list */
4183 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4187 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4188 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4189 for (i=0; i<ovpages; i++) {
4190 DPRINTF("freed ov page %lu", pg);
4191 mdb_midl_append(mc->mc_txn->mt_free_pgs, pg);
4195 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);
4196 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries--;
4197 rc = mdb_rebalance(mc);
4198 if (rc != MDB_SUCCESS)
4199 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4205 mdb_del(MDB_txn *txn, MDB_dbi dbi,
4206 MDB_val *key, MDB_val *data)
4211 MDB_val rdata, *xdata;
4215 assert(key != NULL);
4217 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
4219 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4222 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
4226 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4233 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4234 mc.mc_xcursor = &mx;
4235 mdb_xcursor_init0(&mc);
4237 mc.mc_xcursor = NULL;
4249 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
4251 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
4255 /* Split page <mc->top>, and insert <key,(data|newpgno)> in either left or
4256 * right sibling, at index <mc->ki> (as if unsplit). Updates mc->top and
4257 * mc->ki with the actual values after split, ie if mc->top and mc->ki
4258 * refer to a node in the new right sibling page.
4261 mdb_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
4264 int rc = MDB_SUCCESS, ins_new = 0;
4267 unsigned int i, j, split_indx, nkeys, pmax;
4269 MDB_val sepkey, rkey, rdata;
4271 MDB_page *mp, *rp, *pp;
4276 mp = mc->mc_pg[mc->mc_top];
4277 newindx = mc->mc_ki[mc->mc_top];
4279 DPRINTF("-----> splitting %s page %lu and adding [%s] at index %i",
4280 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
4281 DKEY(newkey), mc->mc_ki[mc->mc_top]);
4283 if (mc->mc_snum < 2) {
4284 if ((pp = mdb_new_page(mc, P_BRANCH, 1)) == NULL)
4286 /* shift current top to make room for new parent */
4287 mc->mc_pg[1] = mc->mc_pg[0];
4288 mc->mc_ki[1] = mc->mc_ki[0];
4291 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = pp->mp_pgno;
4292 DPRINTF("root split! new root = %lu", pp->mp_pgno);
4293 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth++;
4295 /* Add left (implicit) pointer. */
4296 if ((rc = mdb_add_node(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
4297 /* undo the pre-push */
4298 mc->mc_pg[0] = mc->mc_pg[1];
4299 mc->mc_ki[0] = mc->mc_ki[1];
4300 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = mp->mp_pgno;
4301 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth--;
4308 ptop = mc->mc_top-1;
4309 DPRINTF("parent branch page is %lu", mc->mc_pg[ptop]->mp_pgno);
4312 /* Create a right sibling. */
4313 if ((rp = mdb_new_page(mc, mp->mp_flags, 1)) == NULL)
4315 mdb_cursor_copy(mc, &mn);
4316 mn.mc_pg[mn.mc_top] = rp;
4317 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
4318 DPRINTF("new right sibling: page %lu", rp->mp_pgno);
4320 nkeys = NUMKEYS(mp);
4321 split_indx = nkeys / 2 + 1;
4326 unsigned int lsize, rsize, ksize;
4327 /* Move half of the keys to the right sibling */
4329 x = mc->mc_ki[mc->mc_top] - split_indx;
4330 ksize = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
4331 split = LEAF2KEY(mp, split_indx, ksize);
4332 rsize = (nkeys - split_indx) * ksize;
4333 lsize = (nkeys - split_indx) * sizeof(indx_t);
4334 mp->mp_lower -= lsize;
4335 rp->mp_lower += lsize;
4336 mp->mp_upper += rsize - lsize;
4337 rp->mp_upper -= rsize - lsize;
4338 sepkey.mv_size = ksize;
4339 if (newindx == split_indx) {
4340 sepkey.mv_data = newkey->mv_data;
4342 sepkey.mv_data = split;
4345 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
4346 memcpy(rp->mp_ptrs, split, rsize);
4347 sepkey.mv_data = rp->mp_ptrs;
4348 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
4349 memcpy(ins, newkey->mv_data, ksize);
4350 mp->mp_lower += sizeof(indx_t);
4351 mp->mp_upper -= ksize - sizeof(indx_t);
4354 memcpy(rp->mp_ptrs, split, x * ksize);
4355 ins = LEAF2KEY(rp, x, ksize);
4356 memcpy(ins, newkey->mv_data, ksize);
4357 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
4358 rp->mp_lower += sizeof(indx_t);
4359 rp->mp_upper -= ksize - sizeof(indx_t);
4360 mc->mc_ki[mc->mc_top] = x;
4361 mc->mc_pg[mc->mc_top] = rp;
4366 /* For leaf pages, check the split point based on what
4367 * fits where, since otherwise add_node can fail.
4370 unsigned int psize, nsize;
4371 /* Maximum free space in an empty page */
4372 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
4373 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
4374 if (newindx < split_indx) {
4376 for (i=0; i<split_indx; i++) {
4377 node = NODEPTR(mp, i);
4378 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
4379 if (F_ISSET(node->mn_flags, F_BIGDATA))
4380 psize += sizeof(pgno_t);
4382 psize += NODEDSZ(node);
4391 for (i=nkeys-1; i>=split_indx; i--) {
4392 node = NODEPTR(mp, i);
4393 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
4394 if (F_ISSET(node->mn_flags, F_BIGDATA))
4395 psize += sizeof(pgno_t);
4397 psize += NODEDSZ(node);
4407 /* First find the separating key between the split pages.
4409 if (newindx == split_indx) {
4410 sepkey.mv_size = newkey->mv_size;
4411 sepkey.mv_data = newkey->mv_data;
4413 node = NODEPTR(mp, split_indx);
4414 sepkey.mv_size = node->mn_ksize;
4415 sepkey.mv_data = NODEKEY(node);
4419 DPRINTF("separator is [%s]", DKEY(&sepkey));
4421 /* Copy separator key to the parent.
4423 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
4426 rc = mdb_split(&mn, &sepkey, NULL, rp->mp_pgno);
4428 /* Right page might now have changed parent.
4429 * Check if left page also changed parent.
4431 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
4432 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
4433 mc->mc_pg[ptop] = mn.mc_pg[ptop];
4434 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
4438 rc = mdb_add_node(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
4444 if (rc != MDB_SUCCESS) {
4448 /* Move half of the keys to the right sibling. */
4450 /* grab a page to hold a temporary copy */
4451 if (mc->mc_txn->mt_env->me_dpages) {
4452 copy = mc->mc_txn->mt_env->me_dpages;
4453 mc->mc_txn->mt_env->me_dpages = copy->mp_next;
4455 if ((copy = malloc(mc->mc_txn->mt_env->me_psize)) == NULL)
4459 copy->mp_pgno = mp->mp_pgno;
4460 copy->mp_flags = mp->mp_flags;
4461 copy->mp_lower = PAGEHDRSZ;
4462 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
4463 mc->mc_pg[mc->mc_top] = copy;
4464 for (i = j = 0; i <= nkeys; j++) {
4465 if (i == split_indx) {
4466 /* Insert in right sibling. */
4467 /* Reset insert index for right sibling. */
4468 j = (i == newindx && ins_new);
4469 mc->mc_pg[mc->mc_top] = rp;
4472 if (i == newindx && !ins_new) {
4473 /* Insert the original entry that caused the split. */
4474 rkey.mv_data = newkey->mv_data;
4475 rkey.mv_size = newkey->mv_size;
4477 rdata.mv_data = newdata->mv_data;
4478 rdata.mv_size = newdata->mv_size;
4485 /* Update page and index for the new key. */
4486 mc->mc_ki[mc->mc_top] = j;
4487 } else if (i == nkeys) {
4490 node = NODEPTR(mp, i);
4491 rkey.mv_data = NODEKEY(node);
4492 rkey.mv_size = node->mn_ksize;
4494 rdata.mv_data = NODEDATA(node);
4495 rdata.mv_size = NODEDSZ(node);
4497 pgno = NODEPGNO(node);
4498 flags = node->mn_flags;
4503 if (!IS_LEAF(mp) && j == 0) {
4504 /* First branch index doesn't need key data. */
4508 rc = mdb_add_node(mc, j, &rkey, &rdata, pgno, flags);
4511 /* reset back to original page */
4512 if (newindx < split_indx)
4513 mc->mc_pg[mc->mc_top] = mp;
4515 nkeys = NUMKEYS(copy);
4516 for (i=0; i<nkeys; i++)
4517 mp->mp_ptrs[i] = copy->mp_ptrs[i];
4518 mp->mp_lower = copy->mp_lower;
4519 mp->mp_upper = copy->mp_upper;
4520 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
4521 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
4523 /* return tmp page to freelist */
4524 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
4525 mc->mc_txn->mt_env->me_dpages = copy;
4530 mdb_put(MDB_txn *txn, MDB_dbi dbi,
4531 MDB_val *key, MDB_val *data, unsigned int flags)
4536 assert(key != NULL);
4537 assert(data != NULL);
4539 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4542 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
4546 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4550 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA)) != flags)
4557 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4558 mc.mc_xcursor = &mx;
4559 mdb_xcursor_init0(&mc);
4561 mc.mc_xcursor = NULL;
4563 return mdb_cursor_put(&mc, key, data, flags);
4567 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
4569 /** Only a subset of the @ref mdb_env flags can be changed
4570 * at runtime. Changing other flags requires closing the environment
4571 * and re-opening it with the new flags.
4573 #define CHANGEABLE (MDB_NOSYNC)
4574 if ((flag & CHANGEABLE) != flag)
4577 env->me_flags |= flag;
4579 env->me_flags &= ~flag;
4584 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
4589 *arg = env->me_flags;
4594 mdb_env_get_path(MDB_env *env, const char **arg)
4599 *arg = env->me_path;
4604 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
4606 arg->ms_psize = env->me_psize;
4607 arg->ms_depth = db->md_depth;
4608 arg->ms_branch_pages = db->md_branch_pages;
4609 arg->ms_leaf_pages = db->md_leaf_pages;
4610 arg->ms_overflow_pages = db->md_overflow_pages;
4611 arg->ms_entries = db->md_entries;
4616 mdb_env_stat(MDB_env *env, MDB_stat *arg)
4620 if (env == NULL || arg == NULL)
4623 mdb_env_read_meta(env, &toggle);
4625 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
4629 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
4631 if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEKEY)
4632 txn->mt_dbxs[dbi].md_cmp = memnrcmp;
4633 else if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERKEY)
4634 txn->mt_dbxs[dbi].md_cmp = cintcmp;
4636 txn->mt_dbxs[dbi].md_cmp = memncmp;
4638 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4639 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP) {
4640 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED)
4641 txn->mt_dbxs[dbi].md_dcmp = intcmp;
4643 txn->mt_dbxs[dbi].md_dcmp = cintcmp;
4644 } else if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEDUP) {
4645 txn->mt_dbxs[dbi].md_dcmp = memnrcmp;
4647 txn->mt_dbxs[dbi].md_dcmp = memncmp;
4650 txn->mt_dbxs[dbi].md_dcmp = NULL;
4654 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
4661 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
4662 mdb_default_cmp(txn, FREE_DBI);
4668 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
4669 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
4670 mdb_default_cmp(txn, MAIN_DBI);
4674 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
4675 mdb_default_cmp(txn, MAIN_DBI);
4678 /* Is the DB already open? */
4680 for (i=2; i<txn->mt_numdbs; i++) {
4681 if (len == txn->mt_dbxs[i].md_name.mv_size &&
4682 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
4688 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
4691 /* Find the DB info */
4693 key.mv_data = (void *)name;
4694 rc = mdb_get(txn, MAIN_DBI, &key, &data);
4696 /* Create if requested */
4697 if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
4700 data.mv_size = sizeof(MDB_db);
4701 data.mv_data = &dummy;
4702 memset(&dummy, 0, sizeof(dummy));
4703 dummy.md_root = P_INVALID;
4704 dummy.md_flags = flags & 0xffff;
4706 mc.mc_dbi = MAIN_DBI;
4708 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
4712 /* OK, got info, add to table */
4713 if (rc == MDB_SUCCESS) {
4714 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
4715 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
4716 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
4717 txn->mt_dbxs[txn->mt_numdbs].md_parent = MAIN_DBI;
4718 txn->mt_dbxs[txn->mt_numdbs].md_dirty = dirty;
4719 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
4720 *dbi = txn->mt_numdbs;
4721 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
4722 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
4723 mdb_default_cmp(txn, txn->mt_numdbs);
4730 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
4732 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
4735 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
4738 void mdb_close(MDB_txn *txn, MDB_dbi dbi)
4741 if (dbi <= MAIN_DBI || dbi >= txn->mt_numdbs)
4743 ptr = txn->mt_dbxs[dbi].md_name.mv_data;
4744 txn->mt_dbxs[dbi].md_name.mv_data = NULL;
4745 txn->mt_dbxs[dbi].md_name.mv_size = 0;
4749 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
4751 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4754 txn->mt_dbxs[dbi].md_cmp = cmp;
4758 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
4760 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4763 txn->mt_dbxs[dbi].md_dcmp = cmp;
4767 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
4769 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4772 txn->mt_dbxs[dbi].md_rel = rel;