2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
35 #include <sys/types.h>
37 #include <sys/param.h>
43 #ifdef HAVE_SYS_FILE_H
66 /** @defgroup internal MDB Internals
69 /** @defgroup compat Windows Compatibility Macros
70 * A bunch of macros to minimize the amount of platform-specific ifdefs
71 * needed throughout the rest of the code. When the features this library
72 * needs are similar enough to POSIX to be hidden in a one-or-two line
73 * replacement, this macro approach is used.
77 #define pthread_t DWORD
78 #define pthread_mutex_t HANDLE
79 #define pthread_key_t DWORD
80 #define pthread_self() GetCurrentThreadId()
81 #define pthread_key_create(x,y) *(x) = TlsAlloc()
82 #define pthread_key_delete(x) TlsFree(x)
83 #define pthread_getspecific(x) TlsGetValue(x)
84 #define pthread_setspecific(x,y) TlsSetValue(x,y)
85 #define pthread_mutex_unlock(x) ReleaseMutex(x)
86 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
87 #define LOCK_MUTEX_R(env) pthread_mutex_lock(env->me_rmutex)
88 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(env->me_rmutex)
89 #define LOCK_MUTEX_W(env) pthread_mutex_lock(env->me_wmutex)
90 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(env->me_wmutex)
91 #define getpid() GetCurrentProcessId()
92 #define fdatasync(fd) !FlushFileBuffers(fd)
93 #define ErrCode() GetLastError()
94 #define GetPageSize(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
95 #define close(fd) CloseHandle(fd)
96 #define munmap(ptr,len) UnmapViewOfFile(ptr)
98 /** Lock the reader mutex.
100 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&env->me_txns->mti_mutex)
101 /** Unlock the reader mutex.
103 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&env->me_txns->mti_mutex)
105 /** Lock the writer mutex.
106 * Only a single write transaction is allowed at a time. Other writers
107 * will block waiting for this mutex.
109 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&env->me_txns->mti_wmutex)
110 /** Unlock the writer mutex.
112 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&env->me_txns->mti_wmutex)
114 /** Get the error code for the last failed system function.
116 #define ErrCode() errno
118 /** An abstraction for a file handle.
119 * On POSIX systems file handles are small integers. On Windows
120 * they're opaque pointers.
124 /** A value for an invalid file handle.
125 * Mainly used to initialize file variables and signify that they are
128 #define INVALID_HANDLE_VALUE -1
130 /** Get the size of a memory page for the system.
131 * This is the basic size that the platform's memory manager uses, and is
132 * fundamental to the use of memory-mapped files.
134 #define GetPageSize(x) (x) = sysconf(_SC_PAGE_SIZE)
140 /** A flag for opening a file and requesting synchronous data writes.
141 * This is only used when writing a meta page. It's not strictly needed;
142 * we could just do a normal write and then immediately perform a flush.
143 * But if this flag is available it saves us an extra system call.
145 * @note If O_DSYNC is undefined but exists in /usr/include,
146 * preferably set some compiler flag to get the definition.
147 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
150 # define MDB_DSYNC O_DSYNC
154 /** A page number in the database.
155 * Note that 64 bit page numbers are overkill, since pages themselves
156 * already represent 12-13 bits of addressable memory, and the OS will
157 * always limit applications to a maximum of 63 bits of address space.
159 * @note In the #MDB_node structure, we only store 48 bits of this value,
160 * which thus limits us to only 60 bits of addressable data.
162 typedef ULONG pgno_t;
164 /** @defgroup debug Debug Macros
168 /** Enable debug output.
169 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
170 * read from and written to the database (used for free space management).
175 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
176 # define DPRINTF (void) /* Vararg macros may be unsupported */
178 /** Print a debug message with printf formatting. */
179 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
180 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
182 # define DPRINTF(fmt, ...) ((void) 0)
184 /** Print a debug string.
185 * The string is printed literally, with no format processing.
187 #define DPUTS(arg) DPRINTF("%s", arg)
190 /** A default memory page size.
191 * The actual size is platform-dependent, but we use this for
192 * boot-strapping. We probably should not be using this any more.
193 * The #GetPageSize() macro is used to get the actual size.
195 * Note that we don't currently support Huge pages. On Linux,
196 * regular data files cannot use Huge pages, and in general
197 * Huge pages aren't actually pageable. We rely on the OS
198 * demand-pager to read our data and page it out when memory
199 * pressure from other processes is high. So until OSs have
200 * actual paging support for Huge pages, they're not viable.
202 #define PAGESIZE 4096
204 /** The minimum number of keys required in a database page.
205 * Setting this to a larger value will place a smaller bound on the
206 * maximum size of a data item. Data items larger than this size will
207 * be pushed into overflow pages instead of being stored directly in
208 * the B-tree node. This value used to default to 4. With a page size
209 * of 4096 bytes that meant that any item larger than 1024 bytes would
210 * go into an overflow page. That also meant that on average 2-3KB of
211 * each overflow page was wasted space. The value cannot be lower than
212 * 2 because then there would no longer be a tree structure. With this
213 * value, items larger than 2KB will go into overflow pages, and on
214 * average only 1KB will be wasted.
216 #define MDB_MINKEYS 2
218 /** A stamp that identifies a file as an MDB file.
219 * There's nothing special about this value other than that it is easily
220 * recognizable, and it will reflect any byte order mismatches.
222 #define MDB_MAGIC 0xBEEFC0DE
224 /** The version number for a database's file format. */
225 #define MDB_VERSION 1
227 /** The maximum size of a key in the database.
228 * While data items have essentially unbounded size, we require that
229 * keys all fit onto a regular page. This limit could be raised a bit
230 * further if needed; to something just under #PAGESIZE / #MDB_MINKEYS.
232 #define MAXKEYSIZE 511
237 * This is used for printing a hex dump of a key's contents.
239 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
240 /** Display a key in hex.
242 * Invoke a function to display a key in hex.
244 #define DKEY(x) mdb_dkey(x, kbuf)
250 /** @defgroup lazylock Lazy Locking
251 * Macros for locks that are't actually needed.
252 * The DB view is always consistent because all writes are wrapped in
253 * the wmutex. Finer-grained locks aren't necessary.
257 /** Use lazy locking. I.e., don't lock these accesses at all. */
261 /** Grab the reader lock */
262 #define LAZY_MUTEX_LOCK(x)
263 /** Release the reader lock */
264 #define LAZY_MUTEX_UNLOCK(x)
265 /** Release the DB table reader/writer lock */
266 #define LAZY_RWLOCK_UNLOCK(x)
267 /** Grab the DB table write lock */
268 #define LAZY_RWLOCK_WRLOCK(x)
269 /** Grab the DB table read lock */
270 #define LAZY_RWLOCK_RDLOCK(x)
271 /** Declare the DB table rwlock */
272 #define LAZY_RWLOCK_DEF(x)
273 /** Initialize the DB table rwlock */
274 #define LAZY_RWLOCK_INIT(x,y)
275 /** Destroy the DB table rwlock */
276 #define LAZY_RWLOCK_DESTROY(x)
278 #define LAZY_MUTEX_LOCK(x) pthread_mutex_lock(x)
279 #define LAZY_MUTEX_UNLOCK(x) pthread_mutex_unlock(x)
280 #define LAZY_RWLOCK_UNLOCK(x) pthread_rwlock_unlock(x)
281 #define LAZY_RWLOCK_WRLOCK(x) pthread_rwlock_wrlock(x)
282 #define LAZY_RWLOCK_RDLOCK(x) pthread_rwlock_rdlock(x)
283 #define LAZY_RWLOCK_DEF(x) pthread_rwlock_t x
284 #define LAZY_RWLOCK_INIT(x,y) pthread_rwlock_init(x,y)
285 #define LAZY_RWLOCK_DESTROY(x) pthread_rwlock_destroy(x)
289 /** An invalid page number.
290 * Mainly used to denote an empty tree.
292 #define P_INVALID (~0UL)
294 /** Test if a flag \b f is set in a flag word \b w. */
295 #define F_ISSET(w, f) (((w) & (f)) == (f))
297 /** Used for offsets within a single page.
298 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
301 typedef uint16_t indx_t;
303 /** Default size of memory map.
304 * This is certainly too small for any actual applications. Apps should always set
305 * the size explicitly using #mdb_env_set_mapsize().
307 #define DEFAULT_MAPSIZE 1048576
309 /** @defgroup readers Reader Lock Table
310 * Readers don't acquire any locks for their data access. Instead, they
311 * simply record their transaction ID in the reader table. The reader
312 * mutex is needed just to find an empty slot in the reader table. The
313 * slot's address is saved in thread-specific data so that subsequent read
314 * transactions started by the same thread need no further locking to proceed.
316 * Since the database uses multi-version concurrency control, readers don't
317 * actually need any locking. This table is used to keep track of which
318 * readers are using data from which old transactions, so that we'll know
319 * when a particular old transaction is no longer in use. Old transactions
320 * that have 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) {(key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node);}
601 /** Information about a single database in the environment. */
602 typedef struct MDB_db {
603 uint32_t md_pad; /**< also ksize for LEAF2 pages */
604 uint16_t md_flags; /**< @ref mdb_open */
605 uint16_t md_depth; /**< depth of this tree */
606 ULONG md_branch_pages; /**< number of internal pages */
607 ULONG md_leaf_pages; /**< number of leaf pages */
608 ULONG md_overflow_pages; /**< number of overflow pages */
609 ULONG md_entries; /**< number of data items */
610 pgno_t md_root; /**< the root page of this tree */
613 /** Handle for the DB used to track free pages. */
615 /** Handle for the default DB. */
618 /** Identify a data item as a valid sub-DB record */
619 #define MDB_SUBDATA 0x8200
621 /** Meta page content. */
622 typedef struct MDB_meta {
623 /** Stamp identifying this as an MDB data file. It must be set
626 /** Version number of this lock file. Must be set to #MDB_VERSION. */
628 void *mm_address; /**< address for fixed mapping */
629 size_t mm_mapsize; /**< size of mmap region */
630 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
631 /** The size of pages used in this DB */
632 #define mm_psize mm_dbs[0].md_pad
633 /** Any persistent environment flags. @ref mdb_env */
634 #define mm_flags mm_dbs[0].md_flags
635 pgno_t mm_last_pg; /**< last used page in file */
636 ULONG mm_txnid; /**< txnid that committed this page */
639 /** Auxiliary DB info.
640 * The information here is mostly static/read-only. There is
641 * only a single copy of this record in the environment.
642 * The \b md_dirty flag is not read-only, but only a write
643 * transaction can ever update it, and only write transactions
644 * need to worry about it.
646 typedef struct MDB_dbx {
647 MDB_val md_name; /**< name of the database */
648 MDB_cmp_func *md_cmp; /**< function for comparing keys */
649 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
650 MDB_rel_func *md_rel; /**< user relocate function */
651 MDB_dbi md_parent; /**< parent DB of a sub-DB */
652 unsigned int md_dirty; /**< TRUE if DB was written in this txn */
655 /** A database transaction.
656 * Every operation requires a transaction handle.
659 pgno_t mt_next_pgno; /**< next unallocated page */
660 /** The ID of this transaction. IDs are integers incrementing from 1.
661 * Only committed write transactions increment the ID. If a transaction
662 * aborts, the ID may be re-used by the next writer.
665 MDB_env *mt_env; /**< the DB environment */
666 /** The list of pages that became unused during this transaction.
671 ID2L dirty_list; /**< modified pages */
672 MDB_reader *reader; /**< this thread's slot in the reader table */
674 /** Array of records for each DB known in the environment. */
676 /** Array of MDB_db records for each known DB */
678 /** Number of DB records in use. This number only ever increments;
679 * we don't decrement it when individual DB handles are closed.
681 unsigned int mt_numdbs;
683 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
684 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
685 unsigned int mt_flags;
686 /** Tracks which of the two meta pages was used at the start
687 * of this transaction.
689 unsigned int mt_toggle;
692 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
693 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
694 * raise this on a 64 bit machine.
696 #define CURSOR_STACK 32
700 /** Cursors are used for all DB operations */
702 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
703 struct MDB_xcursor *mc_xcursor;
704 /** The transaction that owns this cursor */
706 /** The database handle this cursor operates on */
708 /** The database record for this cursor */
710 /** The database auxiliary record for this cursor */
712 unsigned short mc_snum; /**< number of pushed pages */
713 unsigned short mc_top; /**< index of top page, mc_snum-1 */
714 unsigned int mc_flags;
715 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
716 #define C_EOF 0x02 /**< No more data */
717 #define C_XDIRTY 0x04 /**< @deprecated mc_xcursor needs to be flushed */
718 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
719 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
722 /** Context for sorted-dup records.
723 * We could have gone to a fully recursive design, with arbitrarily
724 * deep nesting of sub-databases. But for now we only handle these
725 * levels - main DB, optional sub-DB, sorted-duplicate DB.
727 typedef struct MDB_xcursor {
728 /** A sub-cursor for traversing the Dup DB */
729 MDB_cursor mx_cursor;
730 /** The database record for this Dup DB */
732 /** The auxiliary DB record for this Dup DB */
736 /** A set of pages freed by an earlier transaction. */
737 typedef struct MDB_oldpages {
738 /** Usually we only read one record from the FREEDB at a time, but
739 * in case we read more, this will chain them together.
741 struct MDB_oldpages *mo_next;
742 /** The ID of the transaction in which these pages were freed. */
744 /** An #IDL of the pages */
745 pgno_t mo_pages[1]; /* dynamic */
748 /** The database environment. */
750 HANDLE me_fd; /**< The main data file */
751 HANDLE me_lfd; /**< The lock file */
752 HANDLE me_mfd; /**< just for writing the meta pages */
753 /** Failed to update the meta page. Probably an I/O error. */
754 #define MDB_FATAL_ERROR 0x80000000U
756 uint32_t me_extrapad; /**< unused for now */
757 unsigned int me_maxreaders; /**< size of the reader table */
758 unsigned int me_numdbs; /**< number of DBs opened */
759 unsigned int me_maxdbs; /**< size of the DB table */
760 char *me_path; /**< path to the DB files */
761 char *me_map; /**< the memory map of the data file */
762 MDB_txninfo *me_txns; /**< the memory map of the lock file */
763 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
764 MDB_txn *me_txn; /**< current write transaction */
765 size_t me_mapsize; /**< size of the data memory map */
766 off_t me_size; /**< current file size */
767 pgno_t me_maxpg; /**< me_mapsize / me_psize */
768 unsigned int me_psize; /**< size of a page, from #GetPageSize */
769 unsigned int me_db_toggle; /**< which DB table is current */
770 MDB_dbx *me_dbxs; /**< array of static DB info */
771 MDB_db *me_dbs[2]; /**< two arrays of MDB_db info */
772 MDB_oldpages *me_pghead; /**< list of old page records */
773 pthread_key_t me_txkey; /**< thread-key for readers */
774 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
775 /** IDL of pages that became unused in a write txn */
776 pgno_t me_free_pgs[MDB_IDL_UM_SIZE];
777 /** ID2L of pages that were written during a write txn */
778 ID2 me_dirty_list[MDB_IDL_UM_SIZE];
779 /** rwlock for the DB tables, if #LAZY_LOCKS is false */
780 LAZY_RWLOCK_DEF(me_dblock);
782 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
786 /** max number of pages to commit in one writev() call */
787 #define MDB_COMMIT_PAGES 64
789 static MDB_page *mdb_alloc_page(MDB_cursor *mc, int num);
790 static int mdb_touch(MDB_cursor *mc);
792 static int mdb_search_page_root(MDB_cursor *mc,
793 MDB_val *key, int modify);
794 static int mdb_search_page(MDB_cursor *mc,
795 MDB_val *key, int modify);
797 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
798 static int mdb_env_read_meta(MDB_env *env, int *which);
799 static int mdb_env_write_meta(MDB_txn *txn);
800 static int mdb_get_page(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
802 static MDB_node *mdb_search_node(MDB_cursor *mc, MDB_val *key, int *exactp);
803 static int mdb_add_node(MDB_cursor *mc, indx_t indx,
804 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags);
805 static void mdb_del_node(MDB_page *mp, indx_t indx, int ksize);
806 static int mdb_del0(MDB_cursor *mc, MDB_node *leaf);
807 static int mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
809 static int mdb_rebalance(MDB_cursor *mc);
810 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
811 static int mdb_move_node(MDB_cursor *csrc, MDB_cursor *cdst);
812 static int mdb_merge(MDB_cursor *csrc, MDB_cursor *cdst);
813 static int mdb_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
815 static MDB_page *mdb_new_page(MDB_cursor *mc, uint32_t flags, int num);
817 static void cursor_pop_page(MDB_cursor *mc);
818 static int cursor_push_page(MDB_cursor *mc, MDB_page *mp);
820 static int mdb_sibling(MDB_cursor *mc, int move_right);
821 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
822 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
823 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
825 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
826 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
828 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi);
829 static void mdb_xcursor_init0(MDB_cursor *mc);
830 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
832 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
833 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
835 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
838 static MDB_cmp_func memncmp, memnrcmp, intcmp, cintcmp;
842 static SECURITY_DESCRIPTOR mdb_null_sd;
843 static SECURITY_ATTRIBUTES mdb_all_sa;
844 static int mdb_sec_inited;
847 /** Return the library version info. */
849 mdb_version(int *major, int *minor, int *patch)
851 if (major) *major = MDB_VERSION_MAJOR;
852 if (minor) *minor = MDB_VERSION_MINOR;
853 if (patch) *patch = MDB_VERSION_PATCH;
854 return MDB_VERSION_STRING;
857 /** Table of descriptions for MDB @ref errors */
858 static char *const mdb_errstr[] = {
859 "MDB_KEYEXIST: Key/data pair already exists",
860 "MDB_NOTFOUND: No matching key/data pair found",
861 "MDB_PAGE_NOTFOUND: Requested page not found",
862 "MDB_CORRUPTED: Located page was wrong type",
863 "MDB_PANIC: Update of meta page failed",
864 "MDB_VERSION_MISMATCH: Database environment version mismatch"
868 mdb_strerror(int err)
871 return ("Successful return: 0");
873 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
874 return mdb_errstr[err - MDB_KEYEXIST];
876 return strerror(err);
880 /** Display a key in hexadecimal and return the address of the result.
881 * @param[in] key the key to display
882 * @param[in] buf the buffer to write into. Should always be #DKBUF.
883 * @return The key in hexadecimal form.
886 mdb_dkey(MDB_val *key, char *buf)
889 unsigned char *c = key->mv_data;
891 if (key->mv_size > MAXKEYSIZE)
893 /* may want to make this a dynamic check: if the key is mostly
894 * printable characters, print it as-is instead of converting to hex.
897 for (i=0; i<key->mv_size; i++)
898 ptr += sprintf(ptr, "%02x", *c++);
900 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
907 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
909 return txn->mt_dbxs[dbi].md_cmp(a, b);
912 /** Compare two data items according to a particular database.
913 * This returns a comparison as if the two items were data items of
914 * a sorted duplicates #MDB_DUPSORT database.
915 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
916 * @param[in] dbi A database handle returned by #mdb_open()
917 * @param[in] a The first item to compare
918 * @param[in] b The second item to compare
919 * @return < 0 if a < b, 0 if a == b, > 0 if a > b
922 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
924 if (txn->mt_dbxs[dbi].md_dcmp)
925 return txn->mt_dbxs[dbi].md_dcmp(a, b);
927 return EINVAL; /* too bad you can't distinguish this from a valid result */
930 /** Allocate pages for writing.
931 * If there are free pages available from older transactions, they
932 * will be re-used first. Otherwise a new page will be allocated.
933 * @param[in] mc cursor A cursor handle identifying the transaction and
934 * database for which we are allocating.
935 * @param[in] num the number of pages to allocate.
936 * @return Address of the allocated page(s). Requests for multiple pages
937 * will always be satisfied by a single contiguous chunk of memory.
940 mdb_alloc_page(MDB_cursor *mc, int num)
942 MDB_txn *txn = mc->mc_txn;
944 pgno_t pgno = P_INVALID;
947 if (txn->mt_txnid > 2) {
949 if (!txn->mt_env->me_pghead && mc->mc_dbi != FREE_DBI &&
950 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
951 /* See if there's anything in the free DB */
956 mdb_cursor_init(&m2, txn, FREE_DBI);
957 mdb_search_page(&m2, NULL, 0);
958 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
959 kptr = (ULONG *)NODEKEY(leaf);
963 oldest = txn->mt_txnid - 1;
964 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
965 ULONG mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
966 if (mr && mr < oldest)
971 if (oldest > *kptr) {
972 /* It's usable, grab it.
978 mdb_read_data(txn, leaf, &data);
979 idl = (ULONG *)data.mv_data;
980 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
981 mop->mo_next = txn->mt_env->me_pghead;
982 mop->mo_txnid = *kptr;
983 txn->mt_env->me_pghead = mop;
984 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
989 DPRINTF("IDL read txn %lu root %lu num %lu",
990 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
991 for (i=0; i<idl[0]; i++) {
992 DPRINTF("IDL %lu", idl[i+1]);
996 /* drop this IDL from the DB */
997 m2.mc_ki[m2.mc_top] = 0;
998 m2.mc_flags = C_INITIALIZED;
999 mdb_cursor_del(&m2, 0);
1002 if (txn->mt_env->me_pghead) {
1003 MDB_oldpages *mop = txn->mt_env->me_pghead;
1005 /* FIXME: For now, always use fresh pages. We
1006 * really ought to search the free list for a
1011 /* peel pages off tail, so we only have to truncate the list */
1012 pgno = MDB_IDL_LAST(mop->mo_pages);
1013 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1015 if (mop->mo_pages[2] > mop->mo_pages[1])
1016 mop->mo_pages[0] = 0;
1020 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1021 txn->mt_env->me_pghead = mop->mo_next;
1028 if (pgno == P_INVALID) {
1029 /* DB size is maxed out */
1030 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1031 assert(txn->mt_next_pgno + num < txn->mt_env->me_maxpg);
1035 if (txn->mt_env->me_dpages && num == 1) {
1036 np = txn->mt_env->me_dpages;
1037 txn->mt_env->me_dpages = np->mp_next;
1039 if ((np = malloc(txn->mt_env->me_psize * num )) == NULL)
1042 if (pgno == P_INVALID) {
1043 np->mp_pgno = txn->mt_next_pgno;
1044 txn->mt_next_pgno += num;
1048 mid.mid = np->mp_pgno;
1050 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1055 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1056 * @param[in] mc cursor pointing to the page to be touched
1057 * @return 0 on success, non-zero on failure.
1060 mdb_touch(MDB_cursor *mc)
1062 MDB_page *mp = mc->mc_pg[mc->mc_top];
1065 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1067 if ((np = mdb_alloc_page(mc, 1)) == NULL)
1069 DPRINTF("touched db %u page %lu -> %lu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1070 assert(mp->mp_pgno != np->mp_pgno);
1071 mdb_midl_append(mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1073 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1076 mp->mp_flags |= P_DIRTY;
1078 mc->mc_pg[mc->mc_top] = mp;
1079 /** If this page has a parent, update the parent to point to
1083 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1089 mdb_env_sync(MDB_env *env, int force)
1092 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1093 if (fdatasync(env->me_fd))
1100 mdb_txn_reset0(MDB_txn *txn);
1102 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1103 * @param[in] txn the transaction handle to initialize
1104 * @return 0 on success, non-zero on failure. This can only
1105 * fail for read-only transactions, and then only if the
1106 * reader table is full.
1109 mdb_txn_renew0(MDB_txn *txn)
1111 MDB_env *env = txn->mt_env;
1113 if (txn->mt_flags & MDB_TXN_RDONLY) {
1114 MDB_reader *r = pthread_getspecific(env->me_txkey);
1117 pid_t pid = getpid();
1118 pthread_t tid = pthread_self();
1121 for (i=0; i<env->me_txns->mti_numreaders; i++)
1122 if (env->me_txns->mti_readers[i].mr_pid == 0)
1124 if (i == env->me_maxreaders) {
1125 UNLOCK_MUTEX_R(env);
1128 env->me_txns->mti_readers[i].mr_pid = pid;
1129 env->me_txns->mti_readers[i].mr_tid = tid;
1130 if (i >= env->me_txns->mti_numreaders)
1131 env->me_txns->mti_numreaders = i+1;
1132 UNLOCK_MUTEX_R(env);
1133 r = &env->me_txns->mti_readers[i];
1134 pthread_setspecific(env->me_txkey, r);
1136 txn->mt_txnid = env->me_txns->mti_txnid;
1137 txn->mt_toggle = env->me_txns->mti_me_toggle;
1138 r->mr_txnid = txn->mt_txnid;
1139 txn->mt_u.reader = r;
1143 txn->mt_txnid = env->me_txns->mti_txnid+1;
1144 txn->mt_toggle = env->me_txns->mti_me_toggle;
1145 txn->mt_u.dirty_list = env->me_dirty_list;
1146 txn->mt_u.dirty_list[0].mid = 0;
1147 txn->mt_free_pgs = env->me_free_pgs;
1148 txn->mt_free_pgs[0] = 0;
1149 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1153 /* Copy the DB arrays */
1154 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
1155 txn->mt_numdbs = env->me_numdbs;
1156 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1157 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1158 if (txn->mt_numdbs > 2)
1159 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
1160 (txn->mt_numdbs - 2) * sizeof(MDB_db));
1161 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1167 mdb_txn_renew(MDB_txn *txn)
1174 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1175 DPUTS("environment had fatal error, must shutdown!");
1179 rc = mdb_txn_renew0(txn);
1180 if (rc == MDB_SUCCESS) {
1181 DPRINTF("renew txn %lu%c %p on mdbenv %p, root page %lu",
1182 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1183 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1189 mdb_txn_begin(MDB_env *env, unsigned int flags, MDB_txn **ret)
1194 if (env->me_flags & MDB_FATAL_ERROR) {
1195 DPUTS("environment had fatal error, must shutdown!");
1198 if ((txn = calloc(1, sizeof(MDB_txn) + env->me_maxdbs * sizeof(MDB_db))) == NULL) {
1199 DPRINTF("calloc: %s", strerror(ErrCode()));
1202 txn->mt_dbs = (MDB_db *)(txn+1);
1203 if (flags & MDB_RDONLY) {
1204 txn->mt_flags |= MDB_TXN_RDONLY;
1208 rc = mdb_txn_renew0(txn);
1213 DPRINTF("begin txn %lu%c %p on mdbenv %p, root page %lu",
1214 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1215 (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1221 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1222 * @param[in] txn the transaction handle to reset
1225 mdb_txn_reset0(MDB_txn *txn)
1227 MDB_env *env = txn->mt_env;
1229 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1230 txn->mt_u.reader->mr_txnid = 0;
1236 /* return all dirty pages to dpage list */
1237 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1238 dp = txn->mt_u.dirty_list[i].mptr;
1239 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1240 dp->mp_next = txn->mt_env->me_dpages;
1241 txn->mt_env->me_dpages = dp;
1243 /* large pages just get freed directly */
1248 while ((mop = txn->mt_env->me_pghead)) {
1249 txn->mt_env->me_pghead = mop->mo_next;
1254 for (i=2; i<env->me_numdbs; i++)
1255 env->me_dbxs[i].md_dirty = 0;
1256 /* The writer mutex was locked in mdb_txn_begin. */
1257 UNLOCK_MUTEX_W(env);
1262 mdb_txn_reset(MDB_txn *txn)
1267 DPRINTF("reset txn %lu%c %p on mdbenv %p, root page %lu",
1268 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1269 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1271 mdb_txn_reset0(txn);
1275 mdb_txn_abort(MDB_txn *txn)
1280 DPRINTF("abort txn %lu%c %p on mdbenv %p, root page %lu",
1281 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1282 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1284 mdb_txn_reset0(txn);
1289 mdb_txn_commit(MDB_txn *txn)
1300 assert(txn != NULL);
1301 assert(txn->mt_env != NULL);
1305 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1310 if (txn != env->me_txn) {
1311 DPUTS("attempt to commit unknown transaction");
1316 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1317 DPUTS("error flag is set, can't commit");
1322 if (!txn->mt_u.dirty_list[0].mid)
1325 DPRINTF("committing txn %lu %p on mdbenv %p, root page %lu",
1326 txn->mt_txnid, txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1328 mdb_cursor_init(&mc, txn, FREE_DBI);
1330 /* should only be one record now */
1331 if (env->me_pghead) {
1332 /* make sure first page of freeDB is touched and on freelist */
1333 mdb_search_page(&mc, NULL, 1);
1335 /* save to free list */
1336 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1340 /* make sure last page of freeDB is touched and on freelist */
1341 key.mv_size = MAXKEYSIZE+1;
1343 mdb_search_page(&mc, &key, 1);
1345 mdb_midl_sort(txn->mt_free_pgs);
1349 ULONG *idl = txn->mt_free_pgs;
1350 DPRINTF("IDL write txn %lu root %lu num %lu",
1351 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1352 for (i=0; i<idl[0]; i++) {
1353 DPRINTF("IDL %lu", idl[i+1]);
1357 /* write to last page of freeDB */
1358 key.mv_size = sizeof(pgno_t);
1359 key.mv_data = (char *)&txn->mt_txnid;
1360 data.mv_data = txn->mt_free_pgs;
1361 /* The free list can still grow during this call,
1362 * despite the pre-emptive touches above. So check
1363 * and make sure the entire thing got written.
1366 i = txn->mt_free_pgs[0];
1367 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1368 rc = mdb_cursor_put(&mc, &key, &data, 0);
1373 } while (i != txn->mt_free_pgs[0]);
1375 /* should only be one record now */
1376 if (env->me_pghead) {
1380 mop = env->me_pghead;
1381 key.mv_size = sizeof(pgno_t);
1382 key.mv_data = (char *)&mop->mo_txnid;
1383 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1384 data.mv_data = mop->mo_pages;
1385 mdb_cursor_put(&mc, &key, &data, 0);
1386 free(env->me_pghead);
1387 env->me_pghead = NULL;
1390 /* Update DB root pointers. Their pages have already been
1391 * touched so this is all in-place and cannot fail.
1395 data.mv_size = sizeof(MDB_db);
1397 mdb_cursor_init(&mc, txn, MAIN_DBI);
1398 for (i = 2; i < txn->mt_numdbs; i++) {
1399 if (txn->mt_dbxs[i].md_dirty) {
1400 data.mv_data = &txn->mt_dbs[i];
1401 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1406 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
1412 /* Windows actually supports scatter/gather I/O, but only on
1413 * unbuffered file handles. Since we're relying on the OS page
1414 * cache for all our data, that's self-defeating. So we just
1415 * write pages one at a time. We use the ov structure to set
1416 * the write offset, to at least save the overhead of a Seek
1420 memset(&ov, 0, sizeof(ov));
1421 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1423 dp = txn->mt_u.dirty_list[i].mptr;
1424 DPRINTF("committing page %lu", dp->mp_pgno);
1425 size = dp->mp_pgno * env->me_psize;
1426 ov.Offset = size & 0xffffffff;
1427 ov.OffsetHigh = size >> 16;
1428 ov.OffsetHigh >>= 16;
1429 /* clear dirty flag */
1430 dp->mp_flags &= ~P_DIRTY;
1431 wsize = env->me_psize;
1432 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
1433 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
1436 DPRINTF("WriteFile: %d", n);
1443 struct iovec iov[MDB_COMMIT_PAGES];
1447 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1448 dp = txn->mt_u.dirty_list[i].mptr;
1449 if (dp->mp_pgno != next) {
1451 DPRINTF("committing %u dirty pages", n);
1452 rc = writev(env->me_fd, iov, n);
1456 DPUTS("short write, filesystem full?");
1458 DPRINTF("writev: %s", strerror(n));
1465 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
1468 DPRINTF("committing page %lu", dp->mp_pgno);
1469 iov[n].iov_len = env->me_psize;
1470 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
1471 iov[n].iov_base = dp;
1472 size += iov[n].iov_len;
1473 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
1474 /* clear dirty flag */
1475 dp->mp_flags &= ~P_DIRTY;
1476 if (++n >= MDB_COMMIT_PAGES) {
1486 DPRINTF("committing %u dirty pages", n);
1487 rc = writev(env->me_fd, iov, n);
1491 DPUTS("short write, filesystem full?");
1493 DPRINTF("writev: %s", strerror(n));
1500 /* Drop the dirty pages.
1502 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1503 dp = txn->mt_u.dirty_list[i].mptr;
1504 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1505 dp->mp_next = txn->mt_env->me_dpages;
1506 txn->mt_env->me_dpages = dp;
1510 txn->mt_u.dirty_list[i].mid = 0;
1512 txn->mt_u.dirty_list[0].mid = 0;
1514 if ((n = mdb_env_sync(env, 0)) != 0 ||
1515 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
1522 /* update the DB tables */
1524 int toggle = !env->me_db_toggle;
1527 ip = &env->me_dbs[toggle][2];
1528 jp = &txn->mt_dbs[2];
1529 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1530 for (i = 2; i < txn->mt_numdbs; i++) {
1531 if (ip->md_root != jp->md_root)
1536 for (i = 2; i < txn->mt_numdbs; i++) {
1537 if (txn->mt_dbxs[i].md_dirty)
1538 txn->mt_dbxs[i].md_dirty = 0;
1540 env->me_db_toggle = toggle;
1541 env->me_numdbs = txn->mt_numdbs;
1542 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1545 UNLOCK_MUTEX_W(env);
1551 /** Read the environment parameters of a DB environment before
1552 * mapping it into memory.
1553 * @param[in] env the environment handle
1554 * @param[out] meta address of where to store the meta information
1555 * @return 0 on success, non-zero on failure.
1558 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
1560 char page[PAGESIZE];
1565 /* We don't know the page size yet, so use a minimum value.
1569 if (!ReadFile(env->me_fd, page, PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
1571 if ((rc = read(env->me_fd, page, PAGESIZE)) == 0)
1576 else if (rc != PAGESIZE) {
1580 DPRINTF("read: %s", strerror(err));
1584 p = (MDB_page *)page;
1586 if (!F_ISSET(p->mp_flags, P_META)) {
1587 DPRINTF("page %lu not a meta page", p->mp_pgno);
1592 if (m->mm_magic != MDB_MAGIC) {
1593 DPUTS("meta has invalid magic");
1597 if (m->mm_version != MDB_VERSION) {
1598 DPRINTF("database is version %u, expected version %u",
1599 m->mm_version, MDB_VERSION);
1600 return MDB_VERSION_MISMATCH;
1603 memcpy(meta, m, sizeof(*m));
1607 /** Write the environment parameters of a freshly created DB environment.
1608 * @param[in] env the environment handle
1609 * @param[out] meta address of where to store the meta information
1610 * @return 0 on success, non-zero on failure.
1613 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
1620 DPUTS("writing new meta page");
1624 meta->mm_magic = MDB_MAGIC;
1625 meta->mm_version = MDB_VERSION;
1626 meta->mm_psize = psize;
1627 meta->mm_last_pg = 1;
1628 meta->mm_flags = env->me_flags & 0xffff;
1629 meta->mm_flags |= MDB_INTEGERKEY;
1630 meta->mm_dbs[0].md_root = P_INVALID;
1631 meta->mm_dbs[1].md_root = P_INVALID;
1633 p = calloc(2, psize);
1635 p->mp_flags = P_META;
1638 memcpy(m, meta, sizeof(*meta));
1640 q = (MDB_page *)((char *)p + psize);
1643 q->mp_flags = P_META;
1646 memcpy(m, meta, sizeof(*meta));
1651 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
1652 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
1655 rc = write(env->me_fd, p, psize * 2);
1656 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
1662 /** Update the environment info to commit a transaction.
1663 * @param[in] txn the transaction that's being committed
1664 * @return 0 on success, non-zero on failure.
1667 mdb_env_write_meta(MDB_txn *txn)
1670 MDB_meta meta, metab;
1672 int rc, len, toggle;
1678 assert(txn != NULL);
1679 assert(txn->mt_env != NULL);
1681 toggle = !txn->mt_toggle;
1682 DPRINTF("writing meta page %d for root page %lu",
1683 toggle, txn->mt_dbs[MAIN_DBI].md_root);
1687 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
1688 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
1690 ptr = (char *)&meta;
1691 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
1692 len = sizeof(MDB_meta) - off;
1695 meta.mm_dbs[0] = txn->mt_dbs[0];
1696 meta.mm_dbs[1] = txn->mt_dbs[1];
1697 meta.mm_last_pg = txn->mt_next_pgno - 1;
1698 meta.mm_txnid = txn->mt_txnid;
1701 off += env->me_psize;
1704 /* Write to the SYNC fd */
1707 memset(&ov, 0, sizeof(ov));
1709 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
1712 rc = pwrite(env->me_mfd, ptr, len, off);
1717 DPUTS("write failed, disk error?");
1718 /* On a failure, the pagecache still contains the new data.
1719 * Write some old data back, to prevent it from being used.
1720 * Use the non-SYNC fd; we know it will fail anyway.
1722 meta.mm_last_pg = metab.mm_last_pg;
1723 meta.mm_txnid = metab.mm_txnid;
1725 WriteFile(env->me_fd, ptr, len, NULL, &ov);
1727 r2 = pwrite(env->me_fd, ptr, len, off);
1729 env->me_flags |= MDB_FATAL_ERROR;
1732 /* Memory ordering issues are irrelevant; since the entire writer
1733 * is wrapped by wmutex, all of these changes will become visible
1734 * after the wmutex is unlocked. Since the DB is multi-version,
1735 * readers will get consistent data regardless of how fresh or
1736 * how stale their view of these values is.
1738 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
1739 txn->mt_env->me_txns->mti_me_toggle = toggle;
1740 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
1741 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
1746 /** Check both meta pages to see which one is newer.
1747 * @param[in] env the environment handle
1748 * @param[out] which address of where to store the meta toggle ID
1749 * @return 0 on success, non-zero on failure.
1752 mdb_env_read_meta(MDB_env *env, int *which)
1756 assert(env != NULL);
1758 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1761 DPRINTF("Using meta page %d", toggle);
1768 mdb_env_create(MDB_env **env)
1772 e = calloc(1, sizeof(MDB_env));
1776 e->me_maxreaders = DEFAULT_READERS;
1778 e->me_fd = INVALID_HANDLE_VALUE;
1779 e->me_lfd = INVALID_HANDLE_VALUE;
1780 e->me_mfd = INVALID_HANDLE_VALUE;
1786 mdb_env_set_mapsize(MDB_env *env, size_t size)
1790 env->me_mapsize = size;
1795 mdb_env_set_maxdbs(MDB_env *env, int dbs)
1799 env->me_maxdbs = dbs;
1804 mdb_env_set_maxreaders(MDB_env *env, int readers)
1808 env->me_maxreaders = readers;
1813 mdb_env_get_maxreaders(MDB_env *env, int *readers)
1815 if (!env || !readers)
1817 *readers = env->me_maxreaders;
1821 /** Further setup required for opening an MDB environment
1824 mdb_env_open2(MDB_env *env, unsigned int flags)
1826 int i, newenv = 0, toggle;
1830 env->me_flags = flags;
1832 memset(&meta, 0, sizeof(meta));
1834 if ((i = mdb_env_read_header(env, &meta)) != 0) {
1837 DPUTS("new mdbenv");
1841 if (!env->me_mapsize) {
1842 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
1848 LONG sizelo, sizehi;
1849 sizelo = env->me_mapsize & 0xffffffff;
1850 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
1852 /* Windows won't create mappings for zero length files.
1853 * Just allocate the maxsize right now.
1856 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
1857 if (!SetEndOfFile(env->me_fd))
1859 SetFilePointer(env->me_fd, 0, NULL, 0);
1861 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
1862 sizehi, sizelo, NULL);
1865 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
1873 if (meta.mm_address && (flags & MDB_FIXEDMAP))
1875 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
1877 if (env->me_map == MAP_FAILED)
1882 meta.mm_mapsize = env->me_mapsize;
1883 if (flags & MDB_FIXEDMAP)
1884 meta.mm_address = env->me_map;
1885 i = mdb_env_init_meta(env, &meta);
1886 if (i != MDB_SUCCESS) {
1887 munmap(env->me_map, env->me_mapsize);
1891 env->me_psize = meta.mm_psize;
1893 env->me_maxpg = env->me_mapsize / env->me_psize;
1895 p = (MDB_page *)env->me_map;
1896 env->me_metas[0] = METADATA(p);
1897 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
1899 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
1902 DPRINTF("opened database version %u, pagesize %u",
1903 env->me_metas[toggle]->mm_version, env->me_psize);
1904 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
1905 DPRINTF("entries: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
1906 DPRINTF("branch pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
1907 DPRINTF("leaf pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
1908 DPRINTF("overflow pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
1909 DPRINTF("root: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
1915 /* Windows doesn't support destructor callbacks for thread-specific storage */
1917 mdb_env_reader_dest(void *ptr)
1919 MDB_reader *reader = ptr;
1921 reader->mr_txnid = 0;
1927 /* downgrade the exclusive lock on the region back to shared */
1929 mdb_env_share_locks(MDB_env *env)
1933 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1935 env->me_txns->mti_me_toggle = toggle;
1936 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
1941 /* First acquire a shared lock. The Unlock will
1942 * then release the existing exclusive lock.
1944 memset(&ov, 0, sizeof(ov));
1945 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
1946 UnlockFile(env->me_lfd, 0, 0, 1, 0);
1950 struct flock lock_info;
1951 /* The shared lock replaces the existing lock */
1952 memset((void *)&lock_info, 0, sizeof(lock_info));
1953 lock_info.l_type = F_RDLCK;
1954 lock_info.l_whence = SEEK_SET;
1955 lock_info.l_start = 0;
1956 lock_info.l_len = 1;
1957 fcntl(env->me_lfd, F_SETLK, &lock_info);
1963 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
1971 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
1972 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
1973 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
1977 /* Try to get exclusive lock. If we succeed, then
1978 * nobody is using the lock region and we should initialize it.
1981 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
1985 memset(&ov, 0, sizeof(ov));
1986 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
1992 size = GetFileSize(env->me_lfd, NULL);
1994 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
1998 /* Try to get exclusive lock. If we succeed, then
1999 * nobody is using the lock region and we should initialize it.
2002 struct flock lock_info;
2003 memset((void *)&lock_info, 0, sizeof(lock_info));
2004 lock_info.l_type = F_WRLCK;
2005 lock_info.l_whence = SEEK_SET;
2006 lock_info.l_start = 0;
2007 lock_info.l_len = 1;
2008 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2012 lock_info.l_type = F_RDLCK;
2013 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2020 size = lseek(env->me_lfd, 0, SEEK_END);
2022 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2023 if (size < rsize && *excl) {
2025 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2026 if (!SetEndOfFile(env->me_lfd)) {
2031 if (ftruncate(env->me_lfd, rsize) != 0) {
2038 size = rsize - sizeof(MDB_txninfo);
2039 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2044 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2050 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2052 if (!env->me_txns) {
2058 env->me_txns = mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2060 if (env->me_txns == MAP_FAILED) {
2068 if (!mdb_sec_inited) {
2069 InitializeSecurityDescriptor(&mdb_null_sd,
2070 SECURITY_DESCRIPTOR_REVISION);
2071 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2072 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2073 mdb_all_sa.bInheritHandle = FALSE;
2074 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2077 /* FIXME: only using up to 20 characters of the env path here,
2078 * probably not enough to assure uniqueness...
2080 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%.20s", lpath);
2081 ptr = env->me_txns->mti_rmname + sizeof("Global\\MDBr");
2082 while ((ptr = strchr(ptr, '\\')))
2084 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2085 if (!env->me_rmutex) {
2089 sprintf(env->me_txns->mti_rmname, "Global\\MDBw%.20s", lpath);
2090 ptr = env->me_txns->mti_rmname + sizeof("Global\\MDBw");
2091 while ((ptr = strchr(ptr, '\\')))
2093 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2094 if (!env->me_wmutex) {
2099 pthread_mutexattr_t mattr;
2101 pthread_mutexattr_init(&mattr);
2102 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2106 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2107 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2109 env->me_txns->mti_version = MDB_VERSION;
2110 env->me_txns->mti_magic = MDB_MAGIC;
2111 env->me_txns->mti_txnid = 0;
2112 env->me_txns->mti_numreaders = 0;
2113 env->me_txns->mti_me_toggle = 0;
2116 if (env->me_txns->mti_magic != MDB_MAGIC) {
2117 DPUTS("lock region has invalid magic");
2121 if (env->me_txns->mti_version != MDB_VERSION) {
2122 DPRINTF("lock region is version %u, expected version %u",
2123 env->me_txns->mti_version, MDB_VERSION);
2124 rc = MDB_VERSION_MISMATCH;
2128 if (rc != EACCES && rc != EAGAIN) {
2132 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2133 if (!env->me_rmutex) {
2137 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2138 if (!env->me_wmutex) {
2148 env->me_lfd = INVALID_HANDLE_VALUE;
2153 /** The name of the lock file in the DB environment */
2154 #define LOCKNAME "/lock.mdb"
2155 /** The name of the data file in the DB environment */
2156 #define DATANAME "/data.mdb"
2158 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2160 int oflags, rc, len, excl;
2161 char *lpath, *dpath;
2164 lpath = malloc(len + sizeof(LOCKNAME) + len + sizeof(DATANAME));
2167 dpath = lpath + len + sizeof(LOCKNAME);
2168 sprintf(lpath, "%s" LOCKNAME, path);
2169 sprintf(dpath, "%s" DATANAME, path);
2171 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
2176 if (F_ISSET(flags, MDB_RDONLY)) {
2177 oflags = GENERIC_READ;
2178 len = OPEN_EXISTING;
2180 oflags = GENERIC_READ|GENERIC_WRITE;
2183 mode = FILE_ATTRIBUTE_NORMAL;
2184 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2185 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2190 if (F_ISSET(flags, MDB_RDONLY))
2193 oflags = O_RDWR | O_CREAT;
2195 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
2201 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
2202 /* synchronous fd for meta writes */
2204 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2205 mode |= FILE_FLAG_WRITE_THROUGH;
2206 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2207 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2212 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2213 oflags |= MDB_DSYNC;
2214 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
2219 env->me_path = strdup(path);
2220 DPRINTF("opened dbenv %p", (void *) env);
2221 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
2222 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
2224 mdb_env_share_locks(env);
2225 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
2226 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
2227 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
2233 if (env->me_fd != INVALID_HANDLE_VALUE) {
2235 env->me_fd = INVALID_HANDLE_VALUE;
2237 if (env->me_lfd != INVALID_HANDLE_VALUE) {
2239 env->me_lfd = INVALID_HANDLE_VALUE;
2247 mdb_env_close(MDB_env *env)
2254 while (env->me_dpages) {
2255 dp = env->me_dpages;
2256 env->me_dpages = dp->mp_next;
2260 free(env->me_dbs[1]);
2261 free(env->me_dbs[0]);
2265 LAZY_RWLOCK_DESTROY(&env->me_dblock);
2266 pthread_key_delete(env->me_txkey);
2269 munmap(env->me_map, env->me_mapsize);
2274 pid_t pid = getpid();
2276 for (i=0; i<env->me_txns->mti_numreaders; i++)
2277 if (env->me_txns->mti_readers[i].mr_pid == pid)
2278 env->me_txns->mti_readers[i].mr_pid = 0;
2279 munmap(env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
2285 /* only for aligned ints */
2287 intcmp(const MDB_val *a, const MDB_val *b)
2289 if (a->mv_size == sizeof(long))
2291 unsigned long *la, *lb;
2296 unsigned int *ia, *ib;
2303 /* ints must always be the same size */
2305 cintcmp(const MDB_val *a, const MDB_val *b)
2307 #if __BYTE_ORDER == __LITTLE_ENDIAN
2308 unsigned short *u, *c;
2311 u = a->mv_data + a->mv_size;
2312 c = b->mv_data + a->mv_size;
2315 } while(!x && u > (unsigned short *)a->mv_data);
2318 return memcmp(a->mv_data, b->mv_data, a->mv_size);
2323 memncmp(const MDB_val *a, const MDB_val *b)
2329 len_diff = a->mv_size - b->mv_size;
2332 diff = memcmp(a->mv_data, b->mv_data, len);
2333 return diff ? diff : len_diff;
2337 memnrcmp(const MDB_val *a, const MDB_val *b)
2339 const unsigned char *p1, *p2, *p1_lim;
2342 if (b->mv_size == 0)
2343 return a->mv_size != 0;
2344 if (a->mv_size == 0)
2347 p1 = (const unsigned char *)a->mv_data + a->mv_size - 1;
2348 p2 = (const unsigned char *)b->mv_data + b->mv_size - 1;
2350 len_diff = a->mv_size - b->mv_size;
2352 p1_lim = p1 - a->mv_size;
2354 p1_lim = p1 - b->mv_size;
2356 while (p1 > p1_lim) {
2366 /* Search for key within a leaf page, using binary search.
2367 * Returns the smallest entry larger or equal to the key.
2368 * If exactp is non-null, stores whether the found entry was an exact match
2369 * in *exactp (1 or 0).
2370 * If kip is non-null, stores the index of the found entry in *kip.
2371 * If no entry larger or equal to the key is found, returns NULL.
2374 mdb_search_node(MDB_cursor *mc, MDB_val *key, int *exactp)
2376 unsigned int i = 0, nkeys;
2379 MDB_page *mp = mc->mc_pg[mc->mc_top];
2380 MDB_node *node = NULL;
2385 nkeys = NUMKEYS(mp);
2387 DPRINTF("searching %u keys in %s page %lu",
2388 nkeys, IS_LEAF(mp) ? "leaf" : "branch",
2393 low = IS_LEAF(mp) ? 0 : 1;
2395 cmp = mc->mc_dbx->md_cmp;
2397 nodekey.mv_size = mc->mc_db->md_pad;
2398 node = NODEPTR(mp, 0); /* fake */
2400 while (low <= high) {
2401 i = (low + high) >> 1;
2404 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
2406 node = NODEPTR(mp, i);
2408 nodekey.mv_size = node->mn_ksize;
2409 nodekey.mv_data = NODEKEY(node);
2412 rc = cmp(key, &nodekey);
2416 DPRINTF("found leaf index %u [%s], rc = %i",
2417 i, DKEY(&nodekey), rc);
2419 DPRINTF("found branch index %u [%s -> %lu], rc = %i",
2420 i, DKEY(&nodekey), NODEPGNO(node), rc);
2431 if (rc > 0) { /* Found entry is less than the key. */
2432 i++; /* Skip to get the smallest entry larger than key. */
2434 node = NODEPTR(mp, i);
2437 *exactp = (rc == 0);
2438 /* store the key index */
2439 mc->mc_ki[mc->mc_top] = i;
2441 /* There is no entry larger or equal to the key. */
2444 /* nodeptr is fake for LEAF2 */
2449 cursor_pop_page(MDB_cursor *mc)
2454 top = mc->mc_pg[mc->mc_top];
2459 DPRINTF("popped page %lu off db %u cursor %p", top->mp_pgno,
2460 mc->mc_dbi, (void *) mc);
2465 cursor_push_page(MDB_cursor *mc, MDB_page *mp)
2467 DPRINTF("pushing page %lu on db %u cursor %p", mp->mp_pgno,
2468 mc->mc_dbi, (void *) mc);
2470 if (mc->mc_snum >= CURSOR_STACK) {
2471 assert(mc->mc_snum < CURSOR_STACK);
2475 mc->mc_top = mc->mc_snum++;
2476 mc->mc_pg[mc->mc_top] = mp;
2477 mc->mc_ki[mc->mc_top] = 0;
2483 mdb_get_page(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
2487 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
2489 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
2490 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
2491 p = txn->mt_u.dirty_list[x].mptr;
2495 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
2496 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
2500 DPRINTF("page %lu not found", pgno);
2503 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
2507 mdb_search_page_root(MDB_cursor *mc, MDB_val *key, int modify)
2509 MDB_page *mp = mc->mc_pg[mc->mc_top];
2514 while (IS_BRANCH(mp)) {
2517 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
2518 assert(NUMKEYS(mp) > 1);
2519 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
2521 if (key == NULL) /* Initialize cursor to first page. */
2522 mc->mc_ki[mc->mc_top] = 0;
2523 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
2524 /* cursor to last page */
2525 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
2528 node = mdb_search_node(mc, key, &exact);
2530 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
2532 assert(mc->mc_ki[mc->mc_top] > 0);
2533 mc->mc_ki[mc->mc_top]--;
2538 DPRINTF("following index %u for key [%s]",
2539 mc->mc_ki[mc->mc_top], DKEY(key));
2540 assert(mc->mc_ki[mc->mc_top] < NUMKEYS(mp));
2541 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2543 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mp)))
2546 if ((rc = cursor_push_page(mc, mp)))
2550 if ((rc = mdb_touch(mc)) != 0)
2552 mp = mc->mc_pg[mc->mc_top];
2557 DPRINTF("internal error, index points to a %02X page!?",
2559 return MDB_CORRUPTED;
2562 DPRINTF("found leaf page %lu for key [%s]", mp->mp_pgno,
2563 key ? DKEY(key) : NULL);
2568 /* Search for the page a given key should be in.
2569 * Pushes parent pages on the cursor stack.
2570 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
2571 * If modify is true, visited pages are updated with new page numbers.
2574 mdb_search_page(MDB_cursor *mc, MDB_val *key, int modify)
2579 /* Make sure the txn is still viable, then find the root from
2580 * the txn's db table.
2582 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
2583 DPUTS("transaction has failed, must abort");
2586 root = mc->mc_db->md_root;
2588 if (root == P_INVALID) { /* Tree is empty. */
2589 DPUTS("tree is empty");
2590 return MDB_NOTFOUND;
2593 if ((rc = mdb_get_page(mc->mc_txn, root, &mc->mc_pg[0])))
2599 DPRINTF("db %u root page %lu has flags 0x%X",
2600 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
2603 /* For sub-databases, update main root first */
2604 if (mc->mc_dbi > MAIN_DBI && !mc->mc_dbx->md_dirty) {
2606 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI);
2607 rc = mdb_search_page(&mc2, &mc->mc_dbx->md_name, 1);
2610 mc->mc_dbx->md_dirty = 1;
2612 if (!F_ISSET(mc->mc_pg[0]->mp_flags, P_DIRTY)) {
2613 if ((rc = mdb_touch(mc)))
2615 mc->mc_db->md_root = mc->mc_pg[0]->mp_pgno;
2619 return mdb_search_page_root(mc, key, modify);
2623 mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
2625 MDB_page *omp; /* overflow mpage */
2629 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
2630 data->mv_size = NODEDSZ(leaf);
2631 data->mv_data = NODEDATA(leaf);
2635 /* Read overflow data.
2637 data->mv_size = NODEDSZ(leaf);
2638 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
2639 if ((rc = mdb_get_page(txn, pgno, &omp))) {
2640 DPRINTF("read overflow page %lu failed", pgno);
2643 data->mv_data = METADATA(omp);
2649 mdb_get(MDB_txn *txn, MDB_dbi dbi,
2650 MDB_val *key, MDB_val *data)
2659 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
2661 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
2664 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2668 mdb_cursor_init(&mc, txn, dbi);
2669 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
2670 mc.mc_xcursor = &mx;
2671 mdb_xcursor_init0(&mc);
2673 mc.mc_xcursor = NULL;
2675 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
2679 mdb_sibling(MDB_cursor *mc, int move_right)
2685 if (mc->mc_snum < 2) {
2686 return MDB_NOTFOUND; /* root has no siblings */
2689 cursor_pop_page(mc);
2690 DPRINTF("parent page is page %lu, index %u",
2691 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
2693 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
2694 : (mc->mc_ki[mc->mc_top] == 0)) {
2695 DPRINTF("no more keys left, moving to %s sibling",
2696 move_right ? "right" : "left");
2697 if ((rc = mdb_sibling(mc, move_right)) != MDB_SUCCESS)
2701 mc->mc_ki[mc->mc_top]++;
2703 mc->mc_ki[mc->mc_top]--;
2704 DPRINTF("just moving to %s index key %u",
2705 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
2707 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
2709 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
2710 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(indx), &mp)))
2713 cursor_push_page(mc, mp);
2719 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
2725 if (mc->mc_flags & C_EOF) {
2726 return MDB_NOTFOUND;
2729 assert(mc->mc_flags & C_INITIALIZED);
2731 mp = mc->mc_pg[mc->mc_top];
2733 if (mc->mc_db->md_flags & MDB_DUPSORT) {
2734 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2735 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2736 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
2737 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
2738 if (op != MDB_NEXT || rc == MDB_SUCCESS)
2742 mc->mc_xcursor->mx_cursor.mc_flags = 0;
2743 if (op == MDB_NEXT_DUP)
2744 return MDB_NOTFOUND;
2748 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) mc);
2750 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
2751 DPUTS("=====> move to next sibling page");
2752 if (mdb_sibling(mc, 1) != MDB_SUCCESS) {
2753 mc->mc_flags |= C_EOF;
2754 return MDB_NOTFOUND;
2756 mp = mc->mc_pg[mc->mc_top];
2757 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
2759 mc->mc_ki[mc->mc_top]++;
2761 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
2762 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
2765 key->mv_size = mc->mc_db->md_pad;
2766 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
2770 assert(IS_LEAF(mp));
2771 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2773 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2774 mdb_xcursor_init1(mc, leaf);
2777 if ((rc = mdb_read_data(mc->mc_txn, leaf, data) != MDB_SUCCESS))
2780 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2781 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
2782 if (rc != MDB_SUCCESS)
2787 MDB_SET_KEY(leaf, key);
2792 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
2798 assert(mc->mc_flags & C_INITIALIZED);
2800 mp = mc->mc_pg[mc->mc_top];
2802 if (mc->mc_db->md_flags & MDB_DUPSORT) {
2803 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2804 if (op == MDB_PREV || op == MDB_PREV_DUP) {
2805 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2806 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
2807 if (op != MDB_PREV || rc == MDB_SUCCESS)
2810 mc->mc_xcursor->mx_cursor.mc_flags = 0;
2811 if (op == MDB_PREV_DUP)
2812 return MDB_NOTFOUND;
2817 DPRINTF("cursor_prev: top page is %lu in cursor %p", mp->mp_pgno, (void *) mc);
2819 if (mc->mc_ki[mc->mc_top] == 0) {
2820 DPUTS("=====> move to prev sibling page");
2821 if (mdb_sibling(mc, 0) != MDB_SUCCESS) {
2822 mc->mc_flags &= ~C_INITIALIZED;
2823 return MDB_NOTFOUND;
2825 mp = mc->mc_pg[mc->mc_top];
2826 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
2827 DPRINTF("prev page is %lu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
2829 mc->mc_ki[mc->mc_top]--;
2831 mc->mc_flags &= ~C_EOF;
2833 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
2834 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
2837 key->mv_size = mc->mc_db->md_pad;
2838 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
2842 assert(IS_LEAF(mp));
2843 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2845 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2846 mdb_xcursor_init1(mc, leaf);
2849 if ((rc = mdb_read_data(mc->mc_txn, leaf, data) != MDB_SUCCESS))
2852 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2853 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
2854 if (rc != MDB_SUCCESS)
2859 MDB_SET_KEY(leaf, key);
2864 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
2865 MDB_cursor_op op, int *exactp)
2873 assert(key->mv_size > 0);
2875 /* See if we're already on the right page */
2876 if (mc->mc_flags & C_INITIALIZED) {
2879 if (mc->mc_pg[mc->mc_top]->mp_flags & P_LEAF2) {
2880 nodekey.mv_size = mc->mc_db->md_pad;
2881 nodekey.mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, nodekey.mv_size);
2883 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
2884 MDB_SET_KEY(leaf, &nodekey);
2886 rc = mc->mc_dbx->md_cmp(key, &nodekey);
2888 /* Probably happens rarely, but first node on the page
2889 * was the one we wanted.
2891 mc->mc_ki[mc->mc_top] = 0;
2895 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
2900 if (NUMKEYS(mc->mc_pg[mc->mc_top]) > 1) {
2901 if (mc->mc_pg[mc->mc_top]->mp_flags & P_LEAF2) {
2902 nodekey.mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top],
2903 NUMKEYS(mc->mc_pg[mc->mc_top])-1, nodekey.mv_size);
2905 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
2906 MDB_SET_KEY(leaf, &nodekey);
2908 rc = mc->mc_dbx->md_cmp(key, &nodekey);
2910 /* last node was the one we wanted */
2911 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top])-1;
2915 /* This is definitely the right page, skip search_page */
2920 /* If any parents have right-sibs, search.
2921 * Otherwise, there's nothing further.
2923 for (i=0; i<mc->mc_top; i++)
2925 NUMKEYS(mc->mc_pg[i])-1)
2927 if (i == mc->mc_top) {
2928 /* There are no other pages */
2929 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
2930 return MDB_NOTFOUND;
2935 rc = mdb_search_page(mc, key, 0);
2936 if (rc != MDB_SUCCESS)
2939 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
2942 leaf = mdb_search_node(mc, key, exactp);
2943 if (exactp != NULL && !*exactp) {
2944 /* MDB_SET specified and not an exact match. */
2945 return MDB_NOTFOUND;
2949 DPUTS("===> inexact leaf not found, goto sibling");
2950 if ((rc = mdb_sibling(mc, 1)) != MDB_SUCCESS)
2951 return rc; /* no entries matched */
2952 mc->mc_ki[mc->mc_top] = 0;
2953 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
2954 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
2958 mc->mc_flags |= C_INITIALIZED;
2959 mc->mc_flags &= ~C_EOF;
2961 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
2962 key->mv_size = mc->mc_db->md_pad;
2963 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
2967 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2968 mdb_xcursor_init1(mc, leaf);
2971 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2972 if (op == MDB_SET || op == MDB_SET_RANGE) {
2973 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
2976 if (op == MDB_GET_BOTH) {
2982 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
2983 if (rc != MDB_SUCCESS)
2986 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
2988 if ((rc = mdb_read_data(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
2990 rc = mc->mc_dbx->md_dcmp(data, &d2);
2992 if (op == MDB_GET_BOTH || rc > 0)
2993 return MDB_NOTFOUND;
2997 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3002 /* The key already matches in all other cases */
3003 if (op == MDB_SET_RANGE)
3004 MDB_SET_KEY(leaf, key);
3005 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3011 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3016 rc = mdb_search_page(mc, NULL, 0);
3017 if (rc != MDB_SUCCESS)
3019 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3021 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3022 mc->mc_flags |= C_INITIALIZED;
3023 mc->mc_flags &= ~C_EOF;
3025 mc->mc_ki[mc->mc_top] = 0;
3027 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3028 key->mv_size = mc->mc_db->md_pad;
3029 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
3034 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3035 mdb_xcursor_init1(mc, leaf);
3036 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3041 mc->mc_xcursor->mx_cursor.mc_flags = 0;
3042 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3046 MDB_SET_KEY(leaf, key);
3051 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3057 lkey.mv_size = MAXKEYSIZE+1;
3058 lkey.mv_data = NULL;
3060 rc = mdb_search_page(mc, &lkey, 0);
3061 if (rc != MDB_SUCCESS)
3063 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3065 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
3066 mc->mc_flags |= C_INITIALIZED;
3067 mc->mc_flags &= ~C_EOF;
3069 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
3071 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3072 key->mv_size = mc->mc_db->md_pad;
3073 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
3078 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3079 mdb_xcursor_init1(mc, leaf);
3080 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3084 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3089 MDB_SET_KEY(leaf, key);
3094 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3104 case MDB_GET_BOTH_RANGE:
3105 if (data == NULL || mc->mc_xcursor == NULL) {
3112 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3114 } else if (op == MDB_SET_RANGE)
3115 rc = mdb_cursor_set(mc, key, data, op, NULL);
3117 rc = mdb_cursor_set(mc, key, data, op, &exact);
3119 case MDB_GET_MULTIPLE:
3121 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
3122 !(mc->mc_flags & C_INITIALIZED)) {
3127 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
3128 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
3131 case MDB_NEXT_MULTIPLE:
3133 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
3137 if (!(mc->mc_flags & C_INITIALIZED))
3138 rc = mdb_cursor_first(mc, key, data);
3140 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
3141 if (rc == MDB_SUCCESS) {
3142 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
3145 mx = &mc->mc_xcursor->mx_cursor;
3146 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
3148 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
3149 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
3157 case MDB_NEXT_NODUP:
3158 if (!(mc->mc_flags & C_INITIALIZED))
3159 rc = mdb_cursor_first(mc, key, data);
3161 rc = mdb_cursor_next(mc, key, data, op);
3165 case MDB_PREV_NODUP:
3166 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
3167 rc = mdb_cursor_last(mc, key, data);
3169 rc = mdb_cursor_prev(mc, key, data, op);
3172 rc = mdb_cursor_first(mc, key, data);
3176 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
3177 !(mc->mc_flags & C_INITIALIZED) ||
3178 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3182 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3185 rc = mdb_cursor_last(mc, key, data);
3189 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
3190 !(mc->mc_flags & C_INITIALIZED) ||
3191 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3195 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3198 DPRINTF("unhandled/unimplemented cursor operation %u", op);
3207 mdb_cursor_touch(MDB_cursor *mc)
3211 if (mc->mc_dbi > MAIN_DBI && !mc->mc_dbx->md_dirty) {
3213 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI);
3214 rc = mdb_search_page(&mc2, &mc->mc_dbx->md_name, 1);
3217 mc->mc_dbx->md_dirty = 1;
3219 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
3220 if (!F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) {
3225 mc->mc_db->md_root =
3226 mc->mc_pg[mc->mc_top]->mp_pgno;
3230 mc->mc_top = mc->mc_snum-1;
3235 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3239 MDB_val xdata, *rdata, dkey;
3241 char dbuf[PAGESIZE];
3247 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
3250 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
3251 mc->mc_dbi, DKEY(key), key->mv_size, data->mv_size);
3255 if (flags == MDB_CURRENT) {
3256 if (!(mc->mc_flags & C_INITIALIZED))
3259 } else if (mc->mc_db->md_root == P_INVALID) {
3261 /* new database, write a root leaf page */
3262 DPUTS("allocating new root leaf page");
3263 if ((np = mdb_new_page(mc, P_LEAF, 1)) == NULL) {
3267 cursor_push_page(mc, np);
3268 mc->mc_db->md_root = np->mp_pgno;
3269 mc->mc_db->md_depth++;
3270 mc->mc_dbx->md_dirty = 1;
3271 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
3273 np->mp_flags |= P_LEAF2;
3274 mc->mc_flags |= C_INITIALIZED;
3280 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
3281 if (flags == MDB_NOOVERWRITE && rc == 0) {
3282 DPRINTF("duplicate key [%s]", DKEY(key));
3284 return MDB_KEYEXIST;
3286 if (rc && rc != MDB_NOTFOUND)
3290 /* Cursor is positioned, now make sure all pages are writable */
3291 rc2 = mdb_cursor_touch(mc);
3296 /* The key already exists */
3297 if (rc == MDB_SUCCESS) {
3298 /* there's only a key anyway, so this is a no-op */
3299 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3300 unsigned int ksize = mc->mc_db->md_pad;
3301 if (key->mv_size != ksize)
3303 if (flags == MDB_CURRENT) {
3304 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
3305 memcpy(ptr, key->mv_data, ksize);
3310 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3313 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
3314 /* Was a single item before, must convert now */
3315 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3316 dkey.mv_size = NODEDSZ(leaf);
3317 dkey.mv_data = dbuf;
3318 memcpy(dbuf, NODEDATA(leaf), dkey.mv_size);
3319 /* data matches, ignore it */
3320 if (!mc->mc_dbx->md_dcmp(data, &dkey))
3321 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
3322 memset(&dummy, 0, sizeof(dummy));
3323 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
3324 dummy.md_pad = data->mv_size;
3325 dummy.md_flags = MDB_DUPFIXED;
3326 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
3327 dummy.md_flags |= MDB_INTEGERKEY;
3329 dummy.md_flags |= MDB_SUBDATA;
3330 dummy.md_root = P_INVALID;
3331 if (dkey.mv_size == sizeof(MDB_db)) {
3332 memcpy(NODEDATA(leaf), &dummy, sizeof(dummy));
3335 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
3338 xdata.mv_size = sizeof(MDB_db);
3339 xdata.mv_data = &dummy;
3340 /* new sub-DB, must fully init xcursor */
3341 if (flags == MDB_CURRENT)
3347 /* same size, just replace it */
3348 if (!F_ISSET(leaf->mn_flags, F_BIGDATA) &&
3349 NODEDSZ(leaf) == data->mv_size) {
3350 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
3353 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
3355 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
3361 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
3362 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
3363 rc = mdb_split(mc, key, rdata, P_INVALID);
3365 /* There is room already in this leaf page. */
3366 rc = mdb_add_node(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, 0);
3369 if (rc != MDB_SUCCESS)
3370 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
3372 /* Remember if we just added a subdatabase */
3373 if (flags & F_SUBDATA) {
3374 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3375 leaf->mn_flags |= F_SUBDATA;
3378 /* Now store the actual data in the child DB. Note that we're
3379 * storing the user data in the keys field, so there are strict
3380 * size limits on dupdata. The actual data fields of the child
3381 * DB are all zero size.
3385 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3387 if (flags != MDB_CURRENT)
3388 mdb_xcursor_init1(mc, leaf);
3391 if (flags == MDB_NODUPDATA)
3392 flags = MDB_NOOVERWRITE;
3393 /* converted, write the original data first */
3395 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, flags);
3398 leaf->mn_flags |= F_DUPDATA;
3400 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, flags);
3401 db = NODEDATA(leaf);
3402 assert((db->md_flags & MDB_SUBDATA) == MDB_SUBDATA);
3403 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
3405 mc->mc_db->md_entries++;
3412 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
3417 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
3420 if (!mc->mc_flags & C_INITIALIZED)
3423 rc = mdb_cursor_touch(mc);
3427 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3429 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3430 if (flags != MDB_NODUPDATA) {
3431 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
3432 /* If sub-DB still has entries, we're done */
3433 if (mc->mc_xcursor->mx_db.md_root != P_INVALID) {
3434 MDB_db *db = NODEDATA(leaf);
3435 assert((db->md_flags & MDB_SUBDATA) == MDB_SUBDATA);
3436 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
3437 mc->mc_db->md_entries--;
3440 /* otherwise fall thru and delete the sub-DB */
3443 /* add all the child DB's pages to the free list */
3444 rc = mdb_search_page(&mc->mc_xcursor->mx_cursor, NULL, 0);
3445 if (rc == MDB_SUCCESS) {
3450 mx = &mc->mc_xcursor->mx_cursor;
3451 mc->mc_db->md_entries -=
3452 mx->mc_db->md_entries;
3454 cursor_pop_page(mx);
3455 while (mx->mc_snum > 1) {
3456 for (i=0; i<NUMKEYS(mx->mc_pg[mx->mc_top]); i++) {
3459 ni = NODEPTR(mx->mc_pg[mx->mc_top], i);
3461 if ((rc = mdb_get_page(mc->mc_txn, pg, &mp)))
3464 mdb_midl_append(mc->mc_txn->mt_free_pgs, pg);
3466 rc = mdb_sibling(mx, 1);
3471 mdb_midl_append(mc->mc_txn->mt_free_pgs,
3472 mx->mc_db->md_root);
3476 return mdb_del0(mc, leaf);
3479 /* Allocate a page and initialize it
3482 mdb_new_page(MDB_cursor *mc, uint32_t flags, int num)
3486 if ((np = mdb_alloc_page(mc, num)) == NULL)
3488 DPRINTF("allocated new mpage %lu, page size %u",
3489 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
3490 np->mp_flags = flags | P_DIRTY;
3491 np->mp_lower = PAGEHDRSZ;
3492 np->mp_upper = mc->mc_txn->mt_env->me_psize;
3495 mc->mc_db->md_branch_pages++;
3496 else if (IS_LEAF(np))
3497 mc->mc_db->md_leaf_pages++;
3498 else if (IS_OVERFLOW(np)) {
3499 mc->mc_db->md_overflow_pages += num;
3507 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
3511 sz = LEAFSIZE(key, data);
3512 if (data->mv_size >= env->me_psize / MDB_MINKEYS) {
3513 /* put on overflow page */
3514 sz -= data->mv_size - sizeof(pgno_t);
3518 return sz + sizeof(indx_t);
3522 mdb_branch_size(MDB_env *env, MDB_val *key)
3527 if (sz >= env->me_psize / MDB_MINKEYS) {
3528 /* put on overflow page */
3529 /* not implemented */
3530 /* sz -= key->size - sizeof(pgno_t); */
3533 return sz + sizeof(indx_t);
3537 mdb_add_node(MDB_cursor *mc, indx_t indx,
3538 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
3541 size_t node_size = NODESIZE;
3544 MDB_page *mp = mc->mc_pg[mc->mc_top];
3545 MDB_page *ofp = NULL; /* overflow page */
3548 assert(mp->mp_upper >= mp->mp_lower);
3550 DPRINTF("add to %s page %lu index %i, data size %zu key size %zu [%s]",
3551 IS_LEAF(mp) ? "leaf" : "branch",
3552 mp->mp_pgno, indx, data ? data->mv_size : 0,
3553 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
3556 /* Move higher keys up one slot. */
3557 int ksize = mc->mc_db->md_pad, dif;
3558 char *ptr = LEAF2KEY(mp, indx, ksize);
3559 dif = NUMKEYS(mp) - indx;
3561 memmove(ptr+ksize, ptr, dif*ksize);
3562 /* insert new key */
3563 memcpy(ptr, key->mv_data, ksize);
3565 /* Just using these for counting */
3566 mp->mp_lower += sizeof(indx_t);
3567 mp->mp_upper -= ksize - sizeof(indx_t);
3572 node_size += key->mv_size;
3576 if (F_ISSET(flags, F_BIGDATA)) {
3577 /* Data already on overflow page. */
3578 node_size += sizeof(pgno_t);
3579 } else if (data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
3580 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
3581 /* Put data on overflow page. */
3582 DPRINTF("data size is %zu, put on overflow page",
3584 node_size += sizeof(pgno_t);
3585 if ((ofp = mdb_new_page(mc, P_OVERFLOW, ovpages)) == NULL)
3587 DPRINTF("allocated overflow page %lu", ofp->mp_pgno);
3590 node_size += data->mv_size;
3593 node_size += node_size & 1;
3595 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
3596 DPRINTF("not enough room in page %lu, got %u ptrs",
3597 mp->mp_pgno, NUMKEYS(mp));
3598 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
3599 mp->mp_upper - mp->mp_lower);
3600 DPRINTF("node size = %zu", node_size);
3604 /* Move higher pointers up one slot. */
3605 for (i = NUMKEYS(mp); i > indx; i--)
3606 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
3608 /* Adjust free space offsets. */
3609 ofs = mp->mp_upper - node_size;
3610 assert(ofs >= mp->mp_lower + sizeof(indx_t));
3611 mp->mp_ptrs[indx] = ofs;
3613 mp->mp_lower += sizeof(indx_t);
3615 /* Write the node data. */
3616 node = NODEPTR(mp, indx);
3617 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
3618 node->mn_flags = flags;
3620 SETDSZ(node,data->mv_size);
3625 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
3630 if (F_ISSET(flags, F_BIGDATA))
3631 memcpy(node->mn_data + key->mv_size, data->mv_data,
3634 memcpy(node->mn_data + key->mv_size, data->mv_data,
3637 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
3639 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
3647 mdb_del_node(MDB_page *mp, indx_t indx, int ksize)
3650 indx_t i, j, numkeys, ptr;
3654 DPRINTF("delete node %u on %s page %lu", indx,
3655 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
3656 assert(indx < NUMKEYS(mp));
3659 int x = NUMKEYS(mp) - 1 - indx;
3660 base = LEAF2KEY(mp, indx, ksize);
3662 memmove(base, base + ksize, x * ksize);
3663 mp->mp_lower -= sizeof(indx_t);
3664 mp->mp_upper += ksize - sizeof(indx_t);
3668 node = NODEPTR(mp, indx);
3669 sz = NODESIZE + node->mn_ksize;
3671 if (F_ISSET(node->mn_flags, F_BIGDATA))
3672 sz += sizeof(pgno_t);
3674 sz += NODEDSZ(node);
3678 ptr = mp->mp_ptrs[indx];
3679 numkeys = NUMKEYS(mp);
3680 for (i = j = 0; i < numkeys; i++) {
3682 mp->mp_ptrs[j] = mp->mp_ptrs[i];
3683 if (mp->mp_ptrs[i] < ptr)
3684 mp->mp_ptrs[j] += sz;
3689 base = (char *)mp + mp->mp_upper;
3690 memmove(base + sz, base, ptr - mp->mp_upper);
3692 mp->mp_lower -= sizeof(indx_t);
3697 mdb_xcursor_init0(MDB_cursor *mc)
3699 MDB_xcursor *mx = mc->mc_xcursor;
3701 mx->mx_cursor.mc_xcursor = NULL;
3702 mx->mx_cursor.mc_txn = mc->mc_txn;
3703 mx->mx_cursor.mc_db = &mx->mx_db;
3704 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
3705 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
3706 mx->mx_dbx.md_parent = mc->mc_dbi;
3707 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
3708 mx->mx_dbx.md_dcmp = NULL;
3709 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
3710 mx->mx_dbx.md_dirty = 0;
3714 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
3716 MDB_db *db = NODEDATA(node);
3717 MDB_xcursor *mx = mc->mc_xcursor;
3718 assert((db->md_flags & MDB_SUBDATA) == MDB_SUBDATA);
3720 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
3722 if (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY))
3723 mx->mx_dbx.md_dirty = 1;
3724 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
3725 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
3726 mx->mx_cursor.mc_snum = 0;
3727 mx->mx_cursor.mc_flags = 0;
3731 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi)
3735 mc->mc_db = &txn->mt_dbs[dbi];
3736 mc->mc_dbx = &txn->mt_dbxs[dbi];
3742 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
3745 size_t size = sizeof(MDB_cursor);
3747 if (txn == NULL || ret == NULL || !dbi || dbi >= txn->mt_numdbs)
3750 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
3751 size += sizeof(MDB_xcursor);
3753 if ((mc = malloc(size)) != NULL) {
3754 mdb_cursor_init(mc, txn, dbi);
3755 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
3756 MDB_xcursor *mx = (MDB_xcursor *)(mc + 1);
3757 mc->mc_xcursor = mx;
3758 mdb_xcursor_init0(mc);
3769 /* Return the count of duplicate data items for the current key */
3771 mdb_cursor_count(MDB_cursor *mc, unsigned long *countp)
3775 if (mc == NULL || countp == NULL)
3778 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
3781 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3782 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3785 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
3788 *countp = mc->mc_xcursor->mx_db.md_entries;
3794 mdb_cursor_close(MDB_cursor *mc)
3802 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
3804 indx_t ptr, i, numkeys;
3811 node = NODEPTR(mp, indx);
3812 ptr = mp->mp_ptrs[indx];
3813 DPRINTF("update key %u (ofs %u) [%.*s] to [%s] on page %lu",
3815 (int)node->mn_ksize, (char *)NODEKEY(node),
3819 delta = key->mv_size - node->mn_ksize;
3821 if (delta > 0 && SIZELEFT(mp) < delta) {
3822 DPRINTF("OUCH! Not enough room, delta = %d", delta);
3826 numkeys = NUMKEYS(mp);
3827 for (i = 0; i < numkeys; i++) {
3828 if (mp->mp_ptrs[i] <= ptr)
3829 mp->mp_ptrs[i] -= delta;
3832 base = (char *)mp + mp->mp_upper;
3833 len = ptr - mp->mp_upper + NODESIZE;
3834 memmove(base - delta, base, len);
3835 mp->mp_upper -= delta;
3837 node = NODEPTR(mp, indx);
3838 node->mn_ksize = key->mv_size;
3841 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
3846 /* Move a node from csrc to cdst.
3849 mdb_move_node(MDB_cursor *csrc, MDB_cursor *cdst)
3856 /* Mark src and dst as dirty. */
3857 if ((rc = mdb_touch(csrc)) ||
3858 (rc = mdb_touch(cdst)))
3861 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3862 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
3863 key.mv_size = csrc->mc_db->md_pad;
3864 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3866 data.mv_data = NULL;
3868 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
3869 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
3870 unsigned int snum = csrc->mc_snum;
3872 /* must find the lowest key below src */
3873 mdb_search_page_root(csrc, NULL, 0);
3874 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
3875 key.mv_size = NODEKSZ(s2);
3876 key.mv_data = NODEKEY(s2);
3877 csrc->mc_snum = snum--;
3878 csrc->mc_top = snum;
3880 key.mv_size = NODEKSZ(srcnode);
3881 key.mv_data = NODEKEY(srcnode);
3883 data.mv_size = NODEDSZ(srcnode);
3884 data.mv_data = NODEDATA(srcnode);
3886 DPRINTF("moving %s node %u [%s] on page %lu to node %u on page %lu",
3887 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
3888 csrc->mc_ki[csrc->mc_top],
3890 csrc->mc_pg[csrc->mc_top]->mp_pgno,
3891 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
3893 /* Add the node to the destination page.
3895 rc = mdb_add_node(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, NODEPGNO(srcnode),
3897 if (rc != MDB_SUCCESS)
3900 /* Delete the node from the source page.
3902 mdb_del_node(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3904 /* Update the parent separators.
3906 if (csrc->mc_ki[csrc->mc_top] == 0) {
3907 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
3908 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3909 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
3911 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
3912 key.mv_size = NODEKSZ(srcnode);
3913 key.mv_data = NODEKEY(srcnode);
3915 DPRINTF("update separator for source page %lu to [%s]",
3916 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
3917 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
3918 &key)) != MDB_SUCCESS)
3921 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
3923 nullkey.mv_size = 0;
3924 assert(mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey) == MDB_SUCCESS);
3928 if (cdst->mc_ki[cdst->mc_top] == 0) {
3929 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
3930 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3931 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
3933 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
3934 key.mv_size = NODEKSZ(srcnode);
3935 key.mv_data = NODEKEY(srcnode);
3937 DPRINTF("update separator for destination page %lu to [%s]",
3938 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
3939 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
3940 &key)) != MDB_SUCCESS)
3943 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
3945 nullkey.mv_size = 0;
3946 assert(mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey) == MDB_SUCCESS);
3954 mdb_merge(MDB_cursor *csrc, MDB_cursor *cdst)
3961 DPRINTF("merging page %lu into %lu", csrc->mc_pg[csrc->mc_top]->mp_pgno, cdst->mc_pg[cdst->mc_top]->mp_pgno);
3963 assert(csrc->mc_snum > 1); /* can't merge root page */
3964 assert(cdst->mc_snum > 1);
3966 /* Mark dst as dirty. */
3967 if ((rc = mdb_touch(cdst)))
3970 /* Move all nodes from src to dst.
3972 j = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
3973 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3974 key.mv_size = csrc->mc_db->md_pad;
3975 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
3976 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
3977 rc = mdb_add_node(cdst, j, &key, NULL, 0, 0);
3978 if (rc != MDB_SUCCESS)
3980 key.mv_data = (char *)key.mv_data + key.mv_size;
3983 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
3984 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
3986 key.mv_size = srcnode->mn_ksize;
3987 key.mv_data = NODEKEY(srcnode);
3988 data.mv_size = NODEDSZ(srcnode);
3989 data.mv_data = NODEDATA(srcnode);
3990 rc = mdb_add_node(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
3991 if (rc != MDB_SUCCESS)
3996 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
3997 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);
3999 /* Unlink the src page from parent and add to free list.
4001 mdb_del_node(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
4002 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
4004 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
4008 mdb_midl_append(csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
4009 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
4010 csrc->mc_db->md_leaf_pages--;
4012 csrc->mc_db->md_branch_pages--;
4013 cursor_pop_page(csrc);
4015 return mdb_rebalance(csrc);
4019 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
4023 cdst->mc_txn = csrc->mc_txn;
4024 cdst->mc_dbi = csrc->mc_dbi;
4025 cdst->mc_db = csrc->mc_db;
4026 cdst->mc_dbx = csrc->mc_dbx;
4027 cdst->mc_snum = csrc->mc_snum;
4028 cdst->mc_top = csrc->mc_top;
4029 cdst->mc_flags = csrc->mc_flags;
4031 for (i=0; i<csrc->mc_snum; i++) {
4032 cdst->mc_pg[i] = csrc->mc_pg[i];
4033 cdst->mc_ki[i] = csrc->mc_ki[i];
4038 mdb_rebalance(MDB_cursor *mc)
4045 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
4046 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
4047 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);
4049 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
4050 DPRINTF("no need to rebalance page %lu, above fill threshold",
4051 mc->mc_pg[mc->mc_top]->mp_pgno);
4055 if (mc->mc_snum < 2) {
4056 if (NUMKEYS(mc->mc_pg[mc->mc_top]) == 0) {
4057 DPUTS("tree is completely empty");
4058 mc->mc_db->md_root = P_INVALID;
4059 mc->mc_db->md_depth = 0;
4060 mc->mc_db->md_leaf_pages = 0;
4061 mdb_midl_append(mc->mc_txn->mt_free_pgs, mc->mc_pg[mc->mc_top]->mp_pgno);
4063 } else if (IS_BRANCH(mc->mc_pg[mc->mc_top]) && NUMKEYS(mc->mc_pg[mc->mc_top]) == 1) {
4064 DPUTS("collapsing root page!");
4065 mdb_midl_append(mc->mc_txn->mt_free_pgs, mc->mc_pg[mc->mc_top]->mp_pgno);
4066 mc->mc_db->md_root = NODEPGNO(NODEPTR(mc->mc_pg[mc->mc_top], 0));
4067 if ((rc = mdb_get_page(mc->mc_txn, mc->mc_db->md_root,
4068 &mc->mc_pg[mc->mc_top])))
4070 mc->mc_db->md_depth--;
4071 mc->mc_db->md_branch_pages--;
4073 DPUTS("root page doesn't need rebalancing");
4077 /* The parent (branch page) must have at least 2 pointers,
4078 * otherwise the tree is invalid.
4080 ptop = mc->mc_top-1;
4081 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
4083 /* Leaf page fill factor is below the threshold.
4084 * Try to move keys from left or right neighbor, or
4085 * merge with a neighbor page.
4090 mdb_cursor_copy(mc, &mn);
4091 mn.mc_xcursor = NULL;
4093 if (mc->mc_ki[ptop] == 0) {
4094 /* We're the leftmost leaf in our parent.
4096 DPUTS("reading right neighbor");
4098 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
4099 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
4101 mn.mc_ki[mn.mc_top] = 0;
4102 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
4104 /* There is at least one neighbor to the left.
4106 DPUTS("reading left neighbor");
4108 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
4109 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
4111 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
4112 mc->mc_ki[mc->mc_top] = 0;
4115 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
4116 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);
4118 /* If the neighbor page is above threshold and has at least two
4119 * keys, move one key from it.
4121 * Otherwise we should try to merge them.
4123 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
4124 return mdb_move_node(&mn, mc);
4125 else { /* FIXME: if (has_enough_room()) */
4126 if (mc->mc_ki[ptop] == 0)
4127 return mdb_merge(&mn, mc);
4129 return mdb_merge(mc, &mn);
4134 mdb_del0(MDB_cursor *mc, MDB_node *leaf)
4138 /* add overflow pages to free list */
4139 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4143 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4144 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4145 for (i=0; i<ovpages; i++) {
4146 DPRINTF("freed ov page %lu", pg);
4147 mdb_midl_append(mc->mc_txn->mt_free_pgs, pg);
4151 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
4152 mc->mc_db->md_entries--;
4153 rc = mdb_rebalance(mc);
4154 if (rc != MDB_SUCCESS)
4155 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4161 mdb_del(MDB_txn *txn, MDB_dbi dbi,
4162 MDB_val *key, MDB_val *data)
4167 MDB_val rdata, *xdata;
4171 assert(key != NULL);
4173 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
4175 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4178 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
4182 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4186 mdb_cursor_init(&mc, txn, dbi);
4187 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4188 mc.mc_xcursor = &mx;
4189 mdb_xcursor_init0(&mc);
4191 mc.mc_xcursor = NULL;
4203 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
4205 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
4209 /* Split page <mc->top>, and insert <key,(data|newpgno)> in either left or
4210 * right sibling, at index <mc->ki> (as if unsplit). Updates mc->top and
4211 * mc->ki with the actual values after split, ie if mc->top and mc->ki
4212 * refer to a node in the new right sibling page.
4215 mdb_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
4218 int rc = MDB_SUCCESS, ins_new = 0;
4221 unsigned int i, j, split_indx, nkeys, pmax;
4223 MDB_val sepkey, rkey, rdata;
4225 MDB_page *mp, *rp, *pp;
4230 mp = mc->mc_pg[mc->mc_top];
4231 newindx = mc->mc_ki[mc->mc_top];
4233 DPRINTF("-----> splitting %s page %lu and adding [%s] at index %i",
4234 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
4235 DKEY(newkey), mc->mc_ki[mc->mc_top]);
4237 if (mc->mc_snum < 2) {
4238 if ((pp = mdb_new_page(mc, P_BRANCH, 1)) == NULL)
4240 /* shift current top to make room for new parent */
4241 mc->mc_pg[1] = mc->mc_pg[0];
4242 mc->mc_ki[1] = mc->mc_ki[0];
4245 mc->mc_db->md_root = pp->mp_pgno;
4246 DPRINTF("root split! new root = %zu", pp->mp_pgno);
4247 mc->mc_db->md_depth++;
4249 /* Add left (implicit) pointer. */
4250 if ((rc = mdb_add_node(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
4251 /* undo the pre-push */
4252 mc->mc_pg[0] = mc->mc_pg[1];
4253 mc->mc_ki[0] = mc->mc_ki[1];
4254 mc->mc_db->md_root = mp->mp_pgno;
4255 mc->mc_db->md_depth--;
4262 ptop = mc->mc_top-1;
4263 DPRINTF("parent branch page is %lu", mc->mc_pg[ptop]->mp_pgno);
4266 /* Create a right sibling. */
4267 if ((rp = mdb_new_page(mc, mp->mp_flags, 1)) == NULL)
4269 mdb_cursor_copy(mc, &mn);
4270 mn.mc_pg[mn.mc_top] = rp;
4271 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
4272 DPRINTF("new right sibling: page %lu", rp->mp_pgno);
4274 nkeys = NUMKEYS(mp);
4275 split_indx = nkeys / 2 + 1;
4280 unsigned int lsize, rsize, ksize;
4281 /* Move half of the keys to the right sibling */
4283 x = mc->mc_ki[mc->mc_top] - split_indx;
4284 ksize = mc->mc_db->md_pad;
4285 split = LEAF2KEY(mp, split_indx, ksize);
4286 rsize = (nkeys - split_indx) * ksize;
4287 lsize = (nkeys - split_indx) * sizeof(indx_t);
4288 mp->mp_lower -= lsize;
4289 rp->mp_lower += lsize;
4290 mp->mp_upper += rsize - lsize;
4291 rp->mp_upper -= rsize - lsize;
4292 sepkey.mv_size = ksize;
4293 if (newindx == split_indx) {
4294 sepkey.mv_data = newkey->mv_data;
4296 sepkey.mv_data = split;
4299 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
4300 memcpy(rp->mp_ptrs, split, rsize);
4301 sepkey.mv_data = rp->mp_ptrs;
4302 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
4303 memcpy(ins, newkey->mv_data, ksize);
4304 mp->mp_lower += sizeof(indx_t);
4305 mp->mp_upper -= ksize - sizeof(indx_t);
4308 memcpy(rp->mp_ptrs, split, x * ksize);
4309 ins = LEAF2KEY(rp, x, ksize);
4310 memcpy(ins, newkey->mv_data, ksize);
4311 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
4312 rp->mp_lower += sizeof(indx_t);
4313 rp->mp_upper -= ksize - sizeof(indx_t);
4314 mc->mc_ki[mc->mc_top] = x;
4315 mc->mc_pg[mc->mc_top] = rp;
4320 /* For leaf pages, check the split point based on what
4321 * fits where, since otherwise add_node can fail.
4324 unsigned int psize, nsize;
4325 /* Maximum free space in an empty page */
4326 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
4327 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
4328 if (newindx < split_indx) {
4330 for (i=0; i<split_indx; i++) {
4331 node = NODEPTR(mp, i);
4332 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
4333 if (F_ISSET(node->mn_flags, F_BIGDATA))
4334 psize += sizeof(pgno_t);
4336 psize += NODEDSZ(node);
4345 for (i=nkeys-1; i>=split_indx; i--) {
4346 node = NODEPTR(mp, i);
4347 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
4348 if (F_ISSET(node->mn_flags, F_BIGDATA))
4349 psize += sizeof(pgno_t);
4351 psize += NODEDSZ(node);
4361 /* First find the separating key between the split pages.
4363 if (newindx == split_indx) {
4364 sepkey.mv_size = newkey->mv_size;
4365 sepkey.mv_data = newkey->mv_data;
4367 node = NODEPTR(mp, split_indx);
4368 sepkey.mv_size = node->mn_ksize;
4369 sepkey.mv_data = NODEKEY(node);
4373 DPRINTF("separator is [%s]", DKEY(&sepkey));
4375 /* Copy separator key to the parent.
4377 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
4380 rc = mdb_split(&mn, &sepkey, NULL, rp->mp_pgno);
4382 /* Right page might now have changed parent.
4383 * Check if left page also changed parent.
4385 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
4386 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
4387 mc->mc_pg[ptop] = mn.mc_pg[ptop];
4388 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
4392 rc = mdb_add_node(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
4398 if (rc != MDB_SUCCESS) {
4402 /* Move half of the keys to the right sibling. */
4404 /* grab a page to hold a temporary copy */
4405 if (mc->mc_txn->mt_env->me_dpages) {
4406 copy = mc->mc_txn->mt_env->me_dpages;
4407 mc->mc_txn->mt_env->me_dpages = copy->mp_next;
4409 if ((copy = malloc(mc->mc_txn->mt_env->me_psize)) == NULL)
4413 copy->mp_pgno = mp->mp_pgno;
4414 copy->mp_flags = mp->mp_flags;
4415 copy->mp_lower = PAGEHDRSZ;
4416 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
4417 mc->mc_pg[mc->mc_top] = copy;
4418 for (i = j = 0; i <= nkeys; j++) {
4419 if (i == split_indx) {
4420 /* Insert in right sibling. */
4421 /* Reset insert index for right sibling. */
4422 j = (i == newindx && ins_new);
4423 mc->mc_pg[mc->mc_top] = rp;
4426 if (i == newindx && !ins_new) {
4427 /* Insert the original entry that caused the split. */
4428 rkey.mv_data = newkey->mv_data;
4429 rkey.mv_size = newkey->mv_size;
4431 rdata.mv_data = newdata->mv_data;
4432 rdata.mv_size = newdata->mv_size;
4439 /* Update page and index for the new key. */
4440 mc->mc_ki[mc->mc_top] = j;
4441 } else if (i == nkeys) {
4444 node = NODEPTR(mp, i);
4445 rkey.mv_data = NODEKEY(node);
4446 rkey.mv_size = node->mn_ksize;
4448 rdata.mv_data = NODEDATA(node);
4449 rdata.mv_size = NODEDSZ(node);
4451 pgno = NODEPGNO(node);
4452 flags = node->mn_flags;
4457 if (!IS_LEAF(mp) && j == 0) {
4458 /* First branch index doesn't need key data. */
4462 rc = mdb_add_node(mc, j, &rkey, &rdata, pgno, flags);
4465 /* reset back to original page */
4466 if (newindx < split_indx)
4467 mc->mc_pg[mc->mc_top] = mp;
4469 nkeys = NUMKEYS(copy);
4470 for (i=0; i<nkeys; i++)
4471 mp->mp_ptrs[i] = copy->mp_ptrs[i];
4472 mp->mp_lower = copy->mp_lower;
4473 mp->mp_upper = copy->mp_upper;
4474 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
4475 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
4477 /* return tmp page to freelist */
4478 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
4479 mc->mc_txn->mt_env->me_dpages = copy;
4484 mdb_put(MDB_txn *txn, MDB_dbi dbi,
4485 MDB_val *key, MDB_val *data, unsigned int flags)
4490 assert(key != NULL);
4491 assert(data != NULL);
4493 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4496 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
4500 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4504 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA)) != flags)
4507 mdb_cursor_init(&mc, txn, dbi);
4508 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4509 mc.mc_xcursor = &mx;
4510 mdb_xcursor_init0(&mc);
4512 mc.mc_xcursor = NULL;
4514 return mdb_cursor_put(&mc, key, data, flags);
4518 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
4520 /** Only a subset of the @ref mdb_env flags can be changed
4521 * at runtime. Changing other flags requires closing the environment
4522 * and re-opening it with the new flags.
4524 #define CHANGEABLE (MDB_NOSYNC)
4525 if ((flag & CHANGEABLE) != flag)
4528 env->me_flags |= flag;
4530 env->me_flags &= ~flag;
4535 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
4540 *arg = env->me_flags;
4545 mdb_env_get_path(MDB_env *env, const char **arg)
4550 *arg = env->me_path;
4555 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
4557 arg->ms_psize = env->me_psize;
4558 arg->ms_depth = db->md_depth;
4559 arg->ms_branch_pages = db->md_branch_pages;
4560 arg->ms_leaf_pages = db->md_leaf_pages;
4561 arg->ms_overflow_pages = db->md_overflow_pages;
4562 arg->ms_entries = db->md_entries;
4567 mdb_env_stat(MDB_env *env, MDB_stat *arg)
4571 if (env == NULL || arg == NULL)
4574 mdb_env_read_meta(env, &toggle);
4576 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
4580 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
4582 if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEKEY)
4583 txn->mt_dbxs[dbi].md_cmp = memnrcmp;
4584 else if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERKEY)
4585 txn->mt_dbxs[dbi].md_cmp = cintcmp;
4587 txn->mt_dbxs[dbi].md_cmp = memncmp;
4589 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4590 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP) {
4591 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED)
4592 txn->mt_dbxs[dbi].md_dcmp = intcmp;
4594 txn->mt_dbxs[dbi].md_dcmp = cintcmp;
4595 } else if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEDUP) {
4596 txn->mt_dbxs[dbi].md_dcmp = memnrcmp;
4598 txn->mt_dbxs[dbi].md_dcmp = memncmp;
4601 txn->mt_dbxs[dbi].md_dcmp = NULL;
4605 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
4612 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
4613 mdb_default_cmp(txn, FREE_DBI);
4619 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
4620 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
4621 mdb_default_cmp(txn, MAIN_DBI);
4625 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
4626 mdb_default_cmp(txn, MAIN_DBI);
4629 /* Is the DB already open? */
4631 for (i=2; i<txn->mt_numdbs; i++) {
4632 if (len == txn->mt_dbxs[i].md_name.mv_size &&
4633 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
4639 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
4642 /* Find the DB info */
4644 key.mv_data = (void *)name;
4645 rc = mdb_get(txn, MAIN_DBI, &key, &data);
4647 /* Create if requested */
4648 if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
4651 data.mv_size = sizeof(MDB_db);
4652 data.mv_data = &dummy;
4653 memset(&dummy, 0, sizeof(dummy));
4654 dummy.md_root = P_INVALID;
4655 dummy.md_flags = flags & 0xffff;
4656 mdb_cursor_init(&mc, txn, MAIN_DBI);
4657 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
4661 /* OK, got info, add to table */
4662 if (rc == MDB_SUCCESS) {
4663 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
4664 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
4665 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
4666 txn->mt_dbxs[txn->mt_numdbs].md_parent = MAIN_DBI;
4667 txn->mt_dbxs[txn->mt_numdbs].md_dirty = dirty;
4668 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
4669 *dbi = txn->mt_numdbs;
4670 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
4671 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
4672 mdb_default_cmp(txn, txn->mt_numdbs);
4679 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
4681 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
4684 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
4687 void mdb_close(MDB_txn *txn, MDB_dbi dbi)
4690 if (dbi <= MAIN_DBI || dbi >= txn->mt_numdbs)
4692 ptr = txn->mt_dbxs[dbi].md_name.mv_data;
4693 txn->mt_dbxs[dbi].md_name.mv_data = NULL;
4694 txn->mt_dbxs[dbi].md_name.mv_size = 0;
4698 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
4700 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4703 txn->mt_dbxs[dbi].md_cmp = cmp;
4707 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
4709 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4712 txn->mt_dbxs[dbi].md_dcmp = cmp;
4716 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
4718 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4721 txn->mt_dbxs[dbi].md_rel = rel;