2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
35 #include <sys/types.h>
37 #include <sys/param.h>
43 #ifdef HAVE_SYS_FILE_H
66 /** @defgroup internal MDB Internals
69 /** @defgroup compat Windows Compatibility Macros
70 * A bunch of macros to minimize the amount of platform-specific ifdefs
71 * needed throughout the rest of the code. When the features this library
72 * needs are similar enough to POSIX to be hidden in a one-or-two line
73 * replacement, this macro approach is used.
77 #define pthread_t DWORD
78 #define pthread_mutex_t HANDLE
79 #define pthread_key_t DWORD
80 #define pthread_self() GetCurrentThreadId()
81 #define pthread_key_create(x,y) *(x) = TlsAlloc()
82 #define pthread_key_delete(x) TlsFree(x)
83 #define pthread_getspecific(x) TlsGetValue(x)
84 #define pthread_setspecific(x,y) TlsSetValue(x,y)
85 #define pthread_mutex_unlock(x) ReleaseMutex(x)
86 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
87 #define LOCK_MUTEX_R(env) pthread_mutex_lock(env->me_rmutex)
88 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(env->me_rmutex)
89 #define LOCK_MUTEX_W(env) pthread_mutex_lock(env->me_wmutex)
90 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(env->me_wmutex)
91 #define getpid() GetCurrentProcessId()
92 #define fdatasync(fd) !FlushFileBuffers(fd)
93 #define ErrCode() GetLastError()
94 #define GetPageSize(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
95 #define close(fd) CloseHandle(fd)
96 #define munmap(ptr,len) UnmapViewOfFile(ptr)
98 /** Lock the reader mutex.
100 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&env->me_txns->mti_mutex)
101 /** Unlock the reader mutex.
103 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&env->me_txns->mti_mutex)
105 /** Lock the writer mutex.
106 * Only a single write transaction is allowed at a time. Other writers
107 * will block waiting for this mutex.
109 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&env->me_txns->mti_wmutex)
110 /** Unlock the writer mutex.
112 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&env->me_txns->mti_wmutex)
114 /** Get the error code for the last failed system function.
116 #define ErrCode() errno
118 /** An abstraction for a file handle.
119 * On POSIX systems file handles are small integers. On Windows
120 * they're opaque pointers.
124 /** A value for an invalid file handle.
125 * Mainly used to initialize file variables and signify that they are
128 #define INVALID_HANDLE_VALUE -1
130 /** Get the size of a memory page for the system.
131 * This is the basic size that the platform's memory manager uses, and is
132 * fundamental to the use of memory-mapped files.
134 #define GetPageSize(x) (x) = sysconf(_SC_PAGE_SIZE)
140 /** A flag for opening a file and requesting synchronous data writes.
141 * This is only used when writing a meta page. It's not strictly needed;
142 * we could just do a normal write and then immediately perform a flush.
143 * But if this flag is available it saves us an extra system call.
145 * @note If O_DSYNC is undefined but exists in /usr/include,
146 * preferably set some compiler flag to get the definition.
147 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
150 # define MDB_DSYNC O_DSYNC
154 /** A page number in the database.
155 * Note that 64 bit page numbers are overkill, since pages themselves
156 * already represent 12-13 bits of addressable memory, and the OS will
157 * always limit applications to a maximum of 63 bits of address space.
159 * @note In the #MDB_node structure, we only store 48 bits of this value,
160 * which thus limits us to only 60 bits of addressable data.
162 typedef ULONG pgno_t;
164 /** @defgroup debug Debug Macros
168 /** Enable debug output.
169 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
170 * read from and written to the database (used for free space management).
175 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
176 # define DPRINTF (void) /* Vararg macros may be unsupported */
178 /** Print a debug message with printf formatting. */
179 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
180 fprintf(stderr, "%s:%d:(%p) " fmt "\n", __func__, __LINE__, pthread_self(), __VA_ARGS__)
182 # define DPRINTF(fmt, ...) ((void) 0)
184 /** Print a debug string.
185 * The string is printed literally, with no format processing.
187 #define DPUTS(arg) DPRINTF("%s", arg)
190 /** A default memory page size.
191 * The actual size is platform-dependent, but we use this for
192 * boot-strapping. We probably should not be using this any more.
193 * The #GetPageSize() macro is used to get the actual size.
195 * Note that we don't currently support Huge pages. On Linux,
196 * regular data files cannot use Huge pages, and in general
197 * Huge pages aren't actually pageable. We rely on the OS
198 * demand-pager to read our data and page it out when memory
199 * pressure from other processes is high. So until OSs have
200 * actual paging support for Huge pages, they're not viable.
202 #define PAGESIZE 4096
204 /** The minimum number of keys required in a database page.
205 * Setting this to a larger value will place a smaller bound on the
206 * maximum size of a data item. Data items larger than this size will
207 * be pushed into overflow pages instead of being stored directly in
208 * the B-tree node. This value used to default to 4. With a page size
209 * of 4096 bytes that meant that any item larger than 1024 bytes would
210 * go into an overflow page. That also meant that on average 2-3KB of
211 * each overflow page was wasted space. The value cannot be lower than
212 * 2 because then there would no longer be a tree structure. With this
213 * value, items larger than 2KB will go into overflow pages, and on
214 * average only 1KB will be wasted.
216 #define MDB_MINKEYS 2
218 /** A stamp that identifies a file as an MDB file.
219 * There's nothing special about this value other than that it is easily
220 * recognizable, and it will reflect any byte order mismatches.
222 #define MDB_MAGIC 0xBEEFC0DE
224 /** The version number for a database's file format. */
225 #define MDB_VERSION 1
227 /** The maximum size of a key in the database.
228 * While data items have essentially unbounded size, we require that
229 * keys all fit onto a regular page. This limit could be raised a bit
230 * further if needed; to something just under #PAGESIZE / #MDB_MINKEYS.
232 #define MAXKEYSIZE 511
237 * This is used for printing a hex dump of a key's contents.
239 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
240 /** Display a key in hex.
242 * Invoke a function to display a key in hex.
244 #define DKEY(x) mdb_dkey(x, kbuf)
250 /** @defgroup lazylock Lazy Locking
251 * Macros for locks that are't actually needed.
252 * The DB view is always consistent because all writes are wrapped in
253 * the wmutex. Finer-grained locks aren't necessary.
257 /** Use lazy locking. I.e., don't lock these accesses at all. */
261 /** Grab the reader lock */
262 #define LAZY_MUTEX_LOCK(x)
263 /** Release the reader lock */
264 #define LAZY_MUTEX_UNLOCK(x)
265 /** Release the DB table reader/writer lock */
266 #define LAZY_RWLOCK_UNLOCK(x)
267 /** Grab the DB table write lock */
268 #define LAZY_RWLOCK_WRLOCK(x)
269 /** Grab the DB table read lock */
270 #define LAZY_RWLOCK_RDLOCK(x)
271 /** Declare the DB table rwlock */
272 #define LAZY_RWLOCK_DEF(x)
273 /** Initialize the DB table rwlock */
274 #define LAZY_RWLOCK_INIT(x,y)
275 /** Destroy the DB table rwlock */
276 #define LAZY_RWLOCK_DESTROY(x)
278 #define LAZY_MUTEX_LOCK(x) pthread_mutex_lock(x)
279 #define LAZY_MUTEX_UNLOCK(x) pthread_mutex_unlock(x)
280 #define LAZY_RWLOCK_UNLOCK(x) pthread_rwlock_unlock(x)
281 #define LAZY_RWLOCK_WRLOCK(x) pthread_rwlock_wrlock(x)
282 #define LAZY_RWLOCK_RDLOCK(x) pthread_rwlock_rdlock(x)
283 #define LAZY_RWLOCK_DEF(x) pthread_rwlock_t x
284 #define LAZY_RWLOCK_INIT(x,y) pthread_rwlock_init(x,y)
285 #define LAZY_RWLOCK_DESTROY(x) pthread_rwlock_destroy(x)
289 /** An invalid page number.
290 * Mainly used to denote an empty tree.
292 #define P_INVALID (~0UL)
294 /** Test if a flag \b f is set in a flag word \b w. */
295 #define F_ISSET(w, f) (((w) & (f)) == (f))
297 /** Used for offsets within a single page.
298 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
301 typedef uint16_t indx_t;
303 /** Default size of memory map.
304 * This is certainly too small for any actual applications. Apps should always set
305 * the size explicitly using #mdb_env_set_mapsize().
307 #define DEFAULT_MAPSIZE 1048576
309 /** @defgroup readers Reader Lock Table
310 * Readers don't acquire any locks for their data access. Instead, they
311 * simply record their transaction ID in the reader table. The reader
312 * mutex is needed just to find an empty slot in the reader table. The
313 * slot's address is saved in thread-specific data so that subsequent read
314 * transactions started by the same thread need no further locking to proceed.
316 * Since the database uses multi-version concurrency control, readers don't
317 * actually need any locking. This table is used to keep track of which
318 * readers are using data from which old transactions, so that we'll know
319 * when a particular old transaction is no longer in use. Old transactions
320 * that have 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 /** Meta page content. */
619 typedef struct MDB_meta {
620 /** Stamp identifying this as an MDB data file. It must be set
623 /** Version number of this lock file. Must be set to #MDB_VERSION. */
625 void *mm_address; /**< address for fixed mapping */
626 size_t mm_mapsize; /**< size of mmap region */
627 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
628 /** The size of pages used in this DB */
629 #define mm_psize mm_dbs[0].md_pad
630 /** Any persistent environment flags. @ref mdb_env */
631 #define mm_flags mm_dbs[0].md_flags
632 pgno_t mm_last_pg; /**< last used page in file */
633 ULONG mm_txnid; /**< txnid that committed this page */
636 /** Auxiliary DB info.
637 * The information here is mostly static/read-only. There is
638 * only a single copy of this record in the environment.
639 * The \b md_dirty flag is not read-only, but only a write
640 * transaction can ever update it, and only write transactions
641 * need to worry about it.
643 typedef struct MDB_dbx {
644 MDB_val md_name; /**< name of the database */
645 MDB_cmp_func *md_cmp; /**< function for comparing keys */
646 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
647 MDB_rel_func *md_rel; /**< user relocate function */
648 MDB_dbi md_parent; /**< parent DB of a sub-DB */
649 unsigned int md_dirty; /**< TRUE if DB was written in this txn */
652 /** A database transaction.
653 * Every operation requires a transaction handle.
656 pgno_t mt_next_pgno; /**< next unallocated page */
657 /** The ID of this transaction. IDs are integers incrementing from 1.
658 * Only committed write transactions increment the ID. If a transaction
659 * aborts, the ID may be re-used by the next writer.
662 MDB_env *mt_env; /**< the DB environment */
663 /** The list of pages that became unused during this transaction.
668 ID2L dirty_list; /**< modified pages */
669 MDB_reader *reader; /**< this thread's slot in the reader table */
671 /** Array of records for each DB known in the environment. */
673 /** Array of MDB_db records for each known DB */
675 /** Number of DB records in use. This number only ever increments;
676 * we don't decrement it when individual DB handles are closed.
678 unsigned int mt_numdbs;
680 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
681 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
682 unsigned int mt_flags;
683 /** Tracks which of the two meta pages was used at the start
684 * of this transaction.
686 unsigned int mt_toggle;
689 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
690 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
691 * raise this on a 64 bit machine.
693 #define CURSOR_STACK 32
697 /** Cursors are used for all DB operations */
699 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
700 struct MDB_xcursor *mc_xcursor;
701 /** The transaction that owns this cursor */
703 /** The database handle this cursor operates on */
705 unsigned short mc_snum; /**< number of pushed pages */
706 unsigned short mc_top; /**< index of top page, mc_snum-1 */
707 unsigned int mc_flags;
708 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
709 #define C_EOF 0x02 /**< No more data */
710 #define C_XDIRTY 0x04 /**< @deprecated mc_xcursor needs to be flushed */
711 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
712 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
715 /** Context for sorted-dup records.
716 * We could have gone to a fully recursive design, with arbitrarily
717 * deep nesting of sub-databases. But for now we only handle these
718 * levels - main DB, optional sub-DB, sorted-duplicate DB.
720 typedef struct MDB_xcursor {
721 /** A sub-cursor for traversing the Dup DB */
722 MDB_cursor mx_cursor;
723 /** A fake transaction struct for pointing to our own table
727 /** Our private DB information tables. Slots 0 and 1 are always
728 * copies of the corresponding slots in the main transaction. These
729 * hold the FREEDB and MAINDB, respectively. If the main cursor is
730 * on a sub-database, that will be copied to slot 2, and the duplicate
731 * database info will be in slot 3. If the main cursor is on the MAINDB
732 * then the duplicate DB info will be in slot 2 and slot 3 will be unused.
739 /** A set of pages freed by an earlier transaction. */
740 typedef struct MDB_oldpages {
741 /** Usually we only read one record from the FREEDB at a time, but
742 * in case we read more, this will chain them together.
744 struct MDB_oldpages *mo_next;
745 /** The ID of the transaction in which these pages were freed. */
747 /** An #IDL of the pages */
748 pgno_t mo_pages[1]; /* dynamic */
751 /** The database environment. */
753 HANDLE me_fd; /**< The main data file */
754 HANDLE me_lfd; /**< The lock file */
755 HANDLE me_mfd; /**< just for writing the meta pages */
756 /** Failed to update the meta page. Probably an I/O error. */
757 #define MDB_FATAL_ERROR 0x80000000U
759 uint32_t me_extrapad; /**< unused for now */
760 unsigned int me_maxreaders; /**< size of the reader table */
761 unsigned int me_numdbs; /**< number of DBs opened */
762 unsigned int me_maxdbs; /**< size of the DB table */
763 char *me_path; /**< path to the DB files */
764 char *me_map; /**< the memory map of the data file */
765 MDB_txninfo *me_txns; /**< the memory map of the lock file */
766 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
767 MDB_txn *me_txn; /**< current write transaction */
768 size_t me_mapsize; /**< size of the data memory map */
769 off_t me_size; /**< current file size */
770 pgno_t me_maxpg; /**< me_mapsize / me_psize */
771 unsigned int me_psize; /**< size of a page, from #GetPageSize */
772 unsigned int me_db_toggle; /**< which DB table is current */
773 MDB_dbx *me_dbxs; /**< array of static DB info */
774 MDB_db *me_dbs[2]; /**< two arrays of MDB_db info */
775 MDB_oldpages *me_pghead; /**< list of old page records */
776 pthread_key_t me_txkey; /**< thread-key for readers */
777 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
778 /** IDL of pages that became unused in a write txn */
779 pgno_t me_free_pgs[MDB_IDL_UM_SIZE];
780 /** ID2L of pages that were written during a write txn */
781 ID2 me_dirty_list[MDB_IDL_UM_SIZE];
782 /** rwlock for the DB tables, if #LAZY_LOCKS is false */
783 LAZY_RWLOCK_DEF(me_dblock);
785 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
789 /** max number of pages to commit in one writev() call */
790 #define MDB_COMMIT_PAGES 64
792 static MDB_page *mdb_alloc_page(MDB_cursor *mc, int num);
793 static int mdb_touch(MDB_cursor *mc);
795 static int mdb_search_page_root(MDB_cursor *mc,
796 MDB_val *key, int modify);
797 static int mdb_search_page(MDB_cursor *mc,
798 MDB_val *key, int modify);
800 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
801 static int mdb_env_read_meta(MDB_env *env, int *which);
802 static int mdb_env_write_meta(MDB_txn *txn);
803 static int mdb_get_page(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
805 static MDB_node *mdb_search_node(MDB_cursor *mc, MDB_val *key, int *exactp);
806 static int mdb_add_node(MDB_cursor *mc, indx_t indx,
807 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags);
808 static void mdb_del_node(MDB_page *mp, indx_t indx, int ksize);
809 static int mdb_del0(MDB_cursor *mc, MDB_node *leaf);
810 static int mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
812 static int mdb_rebalance(MDB_cursor *mc);
813 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
814 static int mdb_move_node(MDB_cursor *csrc, MDB_cursor *cdst);
815 static int mdb_merge(MDB_cursor *csrc, MDB_cursor *cdst);
816 static int mdb_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
818 static MDB_page *mdb_new_page(MDB_cursor *mc, uint32_t flags, int num);
820 static void cursor_pop_page(MDB_cursor *mc);
821 static int cursor_push_page(MDB_cursor *mc, MDB_page *mp);
823 static int mdb_sibling(MDB_cursor *mc, int move_right);
824 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
825 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
826 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
828 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
829 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
831 static void mdb_xcursor_init0(MDB_cursor *mc);
832 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
833 static void mdb_xcursor_init2(MDB_cursor *mc);
834 static void mdb_xcursor_fini(MDB_cursor *mc);
836 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
837 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
839 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
842 static MDB_cmp_func memncmp, memnrcmp, intcmp, cintcmp;
846 static SECURITY_DESCRIPTOR mdb_null_sd;
847 static SECURITY_ATTRIBUTES mdb_all_sa;
848 static int mdb_sec_inited;
851 /** Return the library version info. */
853 mdb_version(int *major, int *minor, int *patch)
855 if (major) *major = MDB_VERSION_MAJOR;
856 if (minor) *minor = MDB_VERSION_MINOR;
857 if (patch) *patch = MDB_VERSION_PATCH;
858 return MDB_VERSION_STRING;
861 /** Table of descriptions for MDB @ref errors */
862 static char *const mdb_errstr[] = {
863 "MDB_KEYEXIST: Key/data pair already exists",
864 "MDB_NOTFOUND: No matching key/data pair found",
865 "MDB_PAGE_NOTFOUND: Requested page not found",
866 "MDB_CORRUPTED: Located page was wrong type",
867 "MDB_PANIC: Update of meta page failed",
868 "MDB_VERSION_MISMATCH: Database environment version mismatch"
872 mdb_strerror(int err)
875 return ("Successful return: 0");
877 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
878 return mdb_errstr[err - MDB_KEYEXIST];
880 return strerror(err);
884 /** Display a key in hexadecimal and return the address of the result.
885 * @param[in] key the key to display
886 * @param[in] buf the buffer to write into. Should always be #DKBUF.
887 * @return The key in hexadecimal form.
890 mdb_dkey(MDB_val *key, char *buf)
893 unsigned char *c = key->mv_data;
895 if (key->mv_size > MAXKEYSIZE)
897 /* may want to make this a dynamic check: if the key is mostly
898 * printable characters, print it as-is instead of converting to hex.
901 for (i=0; i<key->mv_size; i++)
902 ptr += sprintf(ptr, "%02x", *c++);
904 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
911 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
913 return txn->mt_dbxs[dbi].md_cmp(a, b);
916 /** Compare two data items according to a particular database.
917 * This returns a comparison as if the two items were data items of
918 * a sorted duplicates #MDB_DUPSORT database.
919 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
920 * @param[in] dbi A database handle returned by #mdb_open()
921 * @param[in] a The first item to compare
922 * @param[in] b The second item to compare
923 * @return < 0 if a < b, 0 if a == b, > 0 if a > b
926 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
928 if (txn->mt_dbxs[dbi].md_dcmp)
929 return txn->mt_dbxs[dbi].md_dcmp(a, b);
931 return EINVAL; /* too bad you can't distinguish this from a valid result */
934 /** Allocate pages for writing.
935 * If there are free pages available from older transactions, they
936 * will be re-used first. Otherwise a new page will be allocated.
937 * @param[in] mc cursor A cursor handle identifying the transaction and
938 * database for which we are allocating.
939 * @param[in] num the number of pages to allocate.
940 * @return Address of the allocated page(s). Requests for multiple pages
941 * will always be satisfied by a single contiguous chunk of memory.
944 mdb_alloc_page(MDB_cursor *mc, int num)
946 MDB_txn *txn = mc->mc_txn;
948 pgno_t pgno = P_INVALID;
951 if (txn->mt_txnid > 2) {
953 if (!txn->mt_env->me_pghead && mc->mc_dbi != FREE_DBI &&
954 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
955 /* See if there's anything in the free DB */
961 m2.mc_dbi = FREE_DBI;
964 mdb_search_page(&m2, NULL, 0);
965 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
966 kptr = (ULONG *)NODEKEY(leaf);
970 oldest = txn->mt_txnid - 1;
971 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
972 ULONG mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
973 if (mr && mr < oldest)
978 if (oldest > *kptr) {
979 /* It's usable, grab it.
985 mdb_read_data(txn, leaf, &data);
986 idl = (ULONG *)data.mv_data;
987 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
988 mop->mo_next = txn->mt_env->me_pghead;
989 mop->mo_txnid = *kptr;
990 txn->mt_env->me_pghead = mop;
991 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
996 DPRINTF("IDL read txn %lu root %lu num %lu",
997 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
998 for (i=0; i<idl[0]; i++) {
999 DPRINTF("IDL %lu", idl[i+1]);
1003 /* drop this IDL from the DB */
1004 m2.mc_ki[m2.mc_top] = 0;
1005 m2.mc_flags = C_INITIALIZED;
1006 mdb_cursor_del(&m2, 0);
1009 if (txn->mt_env->me_pghead) {
1010 MDB_oldpages *mop = txn->mt_env->me_pghead;
1012 /* FIXME: For now, always use fresh pages. We
1013 * really ought to search the free list for a
1018 /* peel pages off tail, so we only have to truncate the list */
1019 pgno = MDB_IDL_LAST(mop->mo_pages);
1020 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1022 if (mop->mo_pages[2] > mop->mo_pages[1])
1023 mop->mo_pages[0] = 0;
1027 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1028 txn->mt_env->me_pghead = mop->mo_next;
1035 if (pgno == P_INVALID) {
1036 /* DB size is maxed out */
1037 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg)
1040 if (txn->mt_env->me_dpages && num == 1) {
1041 np = txn->mt_env->me_dpages;
1042 txn->mt_env->me_dpages = np->mp_next;
1044 if ((np = malloc(txn->mt_env->me_psize * num )) == NULL)
1047 if (pgno == P_INVALID) {
1048 np->mp_pgno = txn->mt_next_pgno;
1049 txn->mt_next_pgno += num;
1053 mid.mid = np->mp_pgno;
1055 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1060 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1061 * @param[in] mc cursor pointing to the page to be touched
1062 * @return 0 on success, non-zero on failure.
1065 mdb_touch(MDB_cursor *mc)
1067 MDB_page *mp = mc->mc_pg[mc->mc_top];
1070 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1072 if ((np = mdb_alloc_page(mc, 1)) == NULL)
1074 DPRINTF("touched db %u page %lu -> %lu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1075 assert(mp->mp_pgno != np->mp_pgno);
1076 mdb_midl_append(mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1078 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1081 mp->mp_flags |= P_DIRTY;
1083 mc->mc_pg[mc->mc_top] = mp;
1084 /** If this page has a parent, update the parent to point to
1088 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1094 mdb_env_sync(MDB_env *env, int force)
1097 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1098 if (fdatasync(env->me_fd))
1105 mdb_txn_reset0(MDB_txn *txn);
1107 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1108 * @param[in] txn the transaction handle to initialize
1109 * @return 0 on success, non-zero on failure. This can only
1110 * fail for read-only transactions, and then only if the
1111 * reader table is full.
1114 mdb_txn_renew0(MDB_txn *txn)
1116 MDB_env *env = txn->mt_env;
1118 if (txn->mt_flags & MDB_TXN_RDONLY) {
1119 MDB_reader *r = pthread_getspecific(env->me_txkey);
1122 pid_t pid = getpid();
1123 pthread_t tid = pthread_self();
1126 for (i=0; i<env->me_txns->mti_numreaders; i++)
1127 if (env->me_txns->mti_readers[i].mr_pid == 0)
1129 if (i == env->me_maxreaders) {
1130 UNLOCK_MUTEX_R(env);
1133 env->me_txns->mti_readers[i].mr_pid = pid;
1134 env->me_txns->mti_readers[i].mr_tid = tid;
1135 if (i >= env->me_txns->mti_numreaders)
1136 env->me_txns->mti_numreaders = i+1;
1137 UNLOCK_MUTEX_R(env);
1138 r = &env->me_txns->mti_readers[i];
1139 pthread_setspecific(env->me_txkey, r);
1141 txn->mt_txnid = env->me_txns->mti_txnid;
1142 txn->mt_toggle = env->me_txns->mti_me_toggle;
1143 r->mr_txnid = txn->mt_txnid;
1144 txn->mt_u.reader = r;
1148 txn->mt_txnid = env->me_txns->mti_txnid+1;
1149 txn->mt_toggle = env->me_txns->mti_me_toggle;
1150 txn->mt_u.dirty_list = env->me_dirty_list;
1151 txn->mt_u.dirty_list[0].mid = 0;
1152 txn->mt_free_pgs = env->me_free_pgs;
1153 txn->mt_free_pgs[0] = 0;
1154 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1158 /* Copy the DB arrays */
1159 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
1160 txn->mt_numdbs = env->me_numdbs;
1161 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1162 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1163 if (txn->mt_numdbs > 2)
1164 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
1165 (txn->mt_numdbs - 2) * sizeof(MDB_db));
1166 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1172 mdb_txn_renew(MDB_txn *txn)
1179 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1180 DPUTS("environment had fatal error, must shutdown!");
1184 rc = mdb_txn_renew0(txn);
1185 if (rc == MDB_SUCCESS) {
1186 DPRINTF("renew txn %lu%c %p on mdbenv %p, root page %lu",
1187 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1188 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1194 mdb_txn_begin(MDB_env *env, unsigned int flags, MDB_txn **ret)
1199 if (env->me_flags & MDB_FATAL_ERROR) {
1200 DPUTS("environment had fatal error, must shutdown!");
1203 if ((txn = calloc(1, sizeof(MDB_txn) + env->me_maxdbs * sizeof(MDB_db))) == NULL) {
1204 DPRINTF("calloc: %s", strerror(ErrCode()));
1207 txn->mt_dbs = (MDB_db *)(txn+1);
1208 if (flags & MDB_RDONLY) {
1209 txn->mt_flags |= MDB_TXN_RDONLY;
1213 rc = mdb_txn_renew0(txn);
1218 DPRINTF("begin txn %lu%c %p on mdbenv %p, root page %lu",
1219 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1220 (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1226 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1227 * @param[in] txn the transaction handle to reset
1230 mdb_txn_reset0(MDB_txn *txn)
1232 MDB_env *env = txn->mt_env;
1234 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1235 txn->mt_u.reader->mr_txnid = 0;
1241 /* return all dirty pages to dpage list */
1242 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1243 dp = txn->mt_u.dirty_list[i].mptr;
1244 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1245 dp->mp_next = txn->mt_env->me_dpages;
1246 txn->mt_env->me_dpages = dp;
1248 /* large pages just get freed directly */
1253 while ((mop = txn->mt_env->me_pghead)) {
1254 txn->mt_env->me_pghead = mop->mo_next;
1259 for (i=2; i<env->me_numdbs; i++)
1260 env->me_dbxs[i].md_dirty = 0;
1261 /* The writer mutex was locked in mdb_txn_begin. */
1262 UNLOCK_MUTEX_W(env);
1267 mdb_txn_reset(MDB_txn *txn)
1272 DPRINTF("reset txn %lu%c %p on mdbenv %p, root page %lu",
1273 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1274 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1276 mdb_txn_reset0(txn);
1280 mdb_txn_abort(MDB_txn *txn)
1285 DPRINTF("abort txn %lu%c %p on mdbenv %p, root page %lu",
1286 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w', txn,
1287 (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1289 mdb_txn_reset0(txn);
1294 mdb_txn_commit(MDB_txn *txn)
1305 assert(txn != NULL);
1306 assert(txn->mt_env != NULL);
1310 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1315 if (txn != env->me_txn) {
1316 DPUTS("attempt to commit unknown transaction");
1321 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1322 DPUTS("error flag is set, can't commit");
1327 if (!txn->mt_u.dirty_list[0].mid)
1330 DPRINTF("committing txn %lu %p on mdbenv %p, root page %lu",
1331 txn->mt_txnid, txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1334 mc.mc_dbi = FREE_DBI;
1337 /* should only be one record now */
1338 if (env->me_pghead) {
1339 /* make sure first page of freeDB is touched and on freelist */
1340 mdb_search_page(&mc, NULL, 1);
1342 /* save to free list */
1343 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1347 /* make sure last page of freeDB is touched and on freelist */
1348 key.mv_size = MAXKEYSIZE+1;
1350 mdb_search_page(&mc, &key, 1);
1352 mdb_midl_sort(txn->mt_free_pgs);
1356 ULONG *idl = txn->mt_free_pgs;
1357 DPRINTF("IDL write txn %lu root %lu num %lu",
1358 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1359 for (i=0; i<idl[0]; i++) {
1360 DPRINTF("IDL %lu", idl[i+1]);
1364 /* write to last page of freeDB */
1365 key.mv_size = sizeof(pgno_t);
1366 key.mv_data = (char *)&txn->mt_txnid;
1367 data.mv_data = txn->mt_free_pgs;
1368 /* The free list can still grow during this call,
1369 * despite the pre-emptive touches above. So check
1370 * and make sure the entire thing got written.
1373 i = txn->mt_free_pgs[0];
1374 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1375 rc = mdb_cursor_put(&mc, &key, &data, 0);
1380 } while (i != txn->mt_free_pgs[0]);
1382 /* should only be one record now */
1383 if (env->me_pghead) {
1387 mop = env->me_pghead;
1388 key.mv_size = sizeof(pgno_t);
1389 key.mv_data = (char *)&mop->mo_txnid;
1390 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1391 data.mv_data = mop->mo_pages;
1392 mdb_cursor_put(&mc, &key, &data, 0);
1393 free(env->me_pghead);
1394 env->me_pghead = NULL;
1397 /* Update DB root pointers. Their pages have already been
1398 * touched so this is all in-place and cannot fail.
1402 data.mv_size = sizeof(MDB_db);
1404 mc.mc_dbi = MAIN_DBI;
1406 for (i = 2; i < txn->mt_numdbs; i++) {
1407 if (txn->mt_dbxs[i].md_dirty) {
1408 data.mv_data = &txn->mt_dbs[i];
1409 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1414 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
1420 /* Windows actually supports scatter/gather I/O, but only on
1421 * unbuffered file handles. Since we're relying on the OS page
1422 * cache for all our data, that's self-defeating. So we just
1423 * write pages one at a time. We use the ov structure to set
1424 * the write offset, to at least save the overhead of a Seek
1428 memset(&ov, 0, sizeof(ov));
1429 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1431 dp = txn->mt_u.dirty_list[i].mptr;
1432 DPRINTF("committing page %lu", dp->mp_pgno);
1433 size = dp->mp_pgno * env->me_psize;
1434 ov.Offset = size & 0xffffffff;
1435 ov.OffsetHigh = size >> 16;
1436 ov.OffsetHigh >>= 16;
1437 /* clear dirty flag */
1438 dp->mp_flags &= ~P_DIRTY;
1439 wsize = env->me_psize;
1440 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
1441 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
1444 DPRINTF("WriteFile: %d", n);
1451 struct iovec iov[MDB_COMMIT_PAGES];
1455 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1456 dp = txn->mt_u.dirty_list[i].mptr;
1457 if (dp->mp_pgno != next) {
1459 DPRINTF("committing %u dirty pages", n);
1460 rc = writev(env->me_fd, iov, n);
1464 DPUTS("short write, filesystem full?");
1466 DPRINTF("writev: %s", strerror(n));
1473 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
1476 DPRINTF("committing page %lu", dp->mp_pgno);
1477 iov[n].iov_len = env->me_psize;
1478 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
1479 iov[n].iov_base = dp;
1480 size += iov[n].iov_len;
1481 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
1482 /* clear dirty flag */
1483 dp->mp_flags &= ~P_DIRTY;
1484 if (++n >= MDB_COMMIT_PAGES) {
1494 DPRINTF("committing %u dirty pages", n);
1495 rc = writev(env->me_fd, iov, n);
1499 DPUTS("short write, filesystem full?");
1501 DPRINTF("writev: %s", strerror(n));
1508 /* Drop the dirty pages.
1510 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1511 dp = txn->mt_u.dirty_list[i].mptr;
1512 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1513 dp->mp_next = txn->mt_env->me_dpages;
1514 txn->mt_env->me_dpages = dp;
1518 txn->mt_u.dirty_list[i].mid = 0;
1520 txn->mt_u.dirty_list[0].mid = 0;
1522 if ((n = mdb_env_sync(env, 0)) != 0 ||
1523 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
1530 /* update the DB tables */
1532 int toggle = !env->me_db_toggle;
1535 ip = &env->me_dbs[toggle][2];
1536 jp = &txn->mt_dbs[2];
1537 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1538 for (i = 2; i < txn->mt_numdbs; i++) {
1539 if (ip->md_root != jp->md_root)
1544 for (i = 2; i < txn->mt_numdbs; i++) {
1545 if (txn->mt_dbxs[i].md_dirty)
1546 txn->mt_dbxs[i].md_dirty = 0;
1548 env->me_db_toggle = toggle;
1549 env->me_numdbs = txn->mt_numdbs;
1550 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1553 UNLOCK_MUTEX_W(env);
1559 /** Read the environment parameters of a DB environment before
1560 * mapping it into memory.
1561 * @param[in] env the environment handle
1562 * @param[out] meta address of where to store the meta information
1563 * @return 0 on success, non-zero on failure.
1566 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
1568 char page[PAGESIZE];
1573 /* We don't know the page size yet, so use a minimum value.
1577 if (!ReadFile(env->me_fd, page, PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
1579 if ((rc = read(env->me_fd, page, PAGESIZE)) == 0)
1584 else if (rc != PAGESIZE) {
1588 DPRINTF("read: %s", strerror(err));
1592 p = (MDB_page *)page;
1594 if (!F_ISSET(p->mp_flags, P_META)) {
1595 DPRINTF("page %lu not a meta page", p->mp_pgno);
1600 if (m->mm_magic != MDB_MAGIC) {
1601 DPUTS("meta has invalid magic");
1605 if (m->mm_version != MDB_VERSION) {
1606 DPRINTF("database is version %u, expected version %u",
1607 m->mm_version, MDB_VERSION);
1608 return MDB_VERSION_MISMATCH;
1611 memcpy(meta, m, sizeof(*m));
1615 /** Write the environment parameters of a freshly created DB environment.
1616 * @param[in] env the environment handle
1617 * @param[out] meta address of where to store the meta information
1618 * @return 0 on success, non-zero on failure.
1621 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
1628 DPUTS("writing new meta page");
1632 meta->mm_magic = MDB_MAGIC;
1633 meta->mm_version = MDB_VERSION;
1634 meta->mm_psize = psize;
1635 meta->mm_last_pg = 1;
1636 meta->mm_flags = env->me_flags & 0xffff;
1637 meta->mm_flags |= MDB_INTEGERKEY;
1638 meta->mm_dbs[0].md_root = P_INVALID;
1639 meta->mm_dbs[1].md_root = P_INVALID;
1641 p = calloc(2, psize);
1643 p->mp_flags = P_META;
1646 memcpy(m, meta, sizeof(*meta));
1648 q = (MDB_page *)((char *)p + psize);
1651 q->mp_flags = P_META;
1654 memcpy(m, meta, sizeof(*meta));
1659 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
1660 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
1663 rc = write(env->me_fd, p, psize * 2);
1664 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
1670 /** Update the environment info to commit a transaction.
1671 * @param[in] txn the transaction that's being committed
1672 * @return 0 on success, non-zero on failure.
1675 mdb_env_write_meta(MDB_txn *txn)
1678 MDB_meta meta, metab;
1680 int rc, len, toggle;
1686 assert(txn != NULL);
1687 assert(txn->mt_env != NULL);
1689 toggle = !txn->mt_toggle;
1690 DPRINTF("writing meta page %d for root page %lu",
1691 toggle, txn->mt_dbs[MAIN_DBI].md_root);
1695 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
1696 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
1698 ptr = (char *)&meta;
1699 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
1700 len = sizeof(MDB_meta) - off;
1703 meta.mm_dbs[0] = txn->mt_dbs[0];
1704 meta.mm_dbs[1] = txn->mt_dbs[1];
1705 meta.mm_last_pg = txn->mt_next_pgno - 1;
1706 meta.mm_txnid = txn->mt_txnid;
1709 off += env->me_psize;
1712 /* Write to the SYNC fd */
1715 memset(&ov, 0, sizeof(ov));
1717 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
1720 rc = pwrite(env->me_mfd, ptr, len, off);
1725 DPUTS("write failed, disk error?");
1726 /* On a failure, the pagecache still contains the new data.
1727 * Write some old data back, to prevent it from being used.
1728 * Use the non-SYNC fd; we know it will fail anyway.
1730 meta.mm_last_pg = metab.mm_last_pg;
1731 meta.mm_txnid = metab.mm_txnid;
1733 WriteFile(env->me_fd, ptr, len, NULL, &ov);
1735 r2 = pwrite(env->me_fd, ptr, len, off);
1737 env->me_flags |= MDB_FATAL_ERROR;
1740 /* Memory ordering issues are irrelevant; since the entire writer
1741 * is wrapped by wmutex, all of these changes will become visible
1742 * after the wmutex is unlocked. Since the DB is multi-version,
1743 * readers will get consistent data regardless of how fresh or
1744 * how stale their view of these values is.
1746 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
1747 txn->mt_env->me_txns->mti_me_toggle = toggle;
1748 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
1749 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
1754 /** Check both meta pages to see which one is newer.
1755 * @param[in] env the environment handle
1756 * @param[out] which address of where to store the meta toggle ID
1757 * @return 0 on success, non-zero on failure.
1760 mdb_env_read_meta(MDB_env *env, int *which)
1764 assert(env != NULL);
1766 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1769 DPRINTF("Using meta page %d", toggle);
1776 mdb_env_create(MDB_env **env)
1780 e = calloc(1, sizeof(MDB_env));
1781 if (!e) return ENOMEM;
1783 e->me_maxreaders = DEFAULT_READERS;
1785 e->me_fd = INVALID_HANDLE_VALUE;
1786 e->me_lfd = INVALID_HANDLE_VALUE;
1787 e->me_mfd = INVALID_HANDLE_VALUE;
1793 mdb_env_set_mapsize(MDB_env *env, size_t size)
1797 env->me_mapsize = size;
1802 mdb_env_set_maxdbs(MDB_env *env, int dbs)
1806 env->me_maxdbs = dbs;
1811 mdb_env_set_maxreaders(MDB_env *env, int readers)
1815 env->me_maxreaders = readers;
1820 mdb_env_get_maxreaders(MDB_env *env, int *readers)
1822 if (!env || !readers)
1824 *readers = env->me_maxreaders;
1828 /** Further setup required for opening an MDB environment
1831 mdb_env_open2(MDB_env *env, unsigned int flags)
1833 int i, newenv = 0, toggle;
1837 env->me_flags = flags;
1839 memset(&meta, 0, sizeof(meta));
1841 if ((i = mdb_env_read_header(env, &meta)) != 0) {
1844 DPUTS("new mdbenv");
1848 if (!env->me_mapsize) {
1849 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
1855 LONG sizelo, sizehi;
1856 sizelo = env->me_mapsize & 0xffffffff;
1857 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
1859 /* Windows won't create mappings for zero length files.
1860 * Just allocate the maxsize right now.
1863 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
1864 if (!SetEndOfFile(env->me_fd))
1866 SetFilePointer(env->me_fd, 0, NULL, 0);
1868 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
1869 sizehi, sizelo, NULL);
1872 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
1880 if (meta.mm_address && (flags & MDB_FIXEDMAP))
1882 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
1884 if (env->me_map == MAP_FAILED)
1889 meta.mm_mapsize = env->me_mapsize;
1890 if (flags & MDB_FIXEDMAP)
1891 meta.mm_address = env->me_map;
1892 i = mdb_env_init_meta(env, &meta);
1893 if (i != MDB_SUCCESS) {
1894 munmap(env->me_map, env->me_mapsize);
1898 env->me_psize = meta.mm_psize;
1900 env->me_maxpg = env->me_mapsize / env->me_psize;
1902 p = (MDB_page *)env->me_map;
1903 env->me_metas[0] = METADATA(p);
1904 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
1906 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
1909 DPRINTF("opened database version %u, pagesize %u",
1910 env->me_metas[toggle]->mm_version, env->me_psize);
1911 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
1912 DPRINTF("entries: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
1913 DPRINTF("branch pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
1914 DPRINTF("leaf pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
1915 DPRINTF("overflow pages: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
1916 DPRINTF("root: %lu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
1922 /* Windows doesn't support destructor callbacks for thread-specific storage */
1924 mdb_env_reader_dest(void *ptr)
1926 MDB_reader *reader = ptr;
1928 reader->mr_txnid = 0;
1934 /* downgrade the exclusive lock on the region back to shared */
1936 mdb_env_share_locks(MDB_env *env)
1940 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
1942 env->me_txns->mti_me_toggle = toggle;
1943 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
1948 /* First acquire a shared lock. The Unlock will
1949 * then release the existing exclusive lock.
1951 memset(&ov, 0, sizeof(ov));
1952 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
1953 UnlockFile(env->me_lfd, 0, 0, 1, 0);
1957 struct flock lock_info;
1958 /* The shared lock replaces the existing lock */
1959 memset((void *)&lock_info, 0, sizeof(lock_info));
1960 lock_info.l_type = F_RDLCK;
1961 lock_info.l_whence = SEEK_SET;
1962 lock_info.l_start = 0;
1963 lock_info.l_len = 1;
1964 fcntl(env->me_lfd, F_SETLK, &lock_info);
1970 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
1978 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
1979 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
1980 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
1984 /* Try to get exclusive lock. If we succeed, then
1985 * nobody is using the lock region and we should initialize it.
1988 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
1992 memset(&ov, 0, sizeof(ov));
1993 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
1999 size = GetFileSize(env->me_lfd, NULL);
2001 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2005 /* Try to get exclusive lock. If we succeed, then
2006 * nobody is using the lock region and we should initialize it.
2009 struct flock lock_info;
2010 memset((void *)&lock_info, 0, sizeof(lock_info));
2011 lock_info.l_type = F_WRLCK;
2012 lock_info.l_whence = SEEK_SET;
2013 lock_info.l_start = 0;
2014 lock_info.l_len = 1;
2015 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2019 lock_info.l_type = F_RDLCK;
2020 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2027 size = lseek(env->me_lfd, 0, SEEK_END);
2029 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2030 if (size < rsize && *excl) {
2032 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2033 if (!SetEndOfFile(env->me_lfd)) {
2038 if (ftruncate(env->me_lfd, rsize) != 0) {
2045 size = rsize - sizeof(MDB_txninfo);
2046 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2051 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2057 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2059 if (!env->me_txns) {
2065 env->me_txns = mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2067 if (env->me_txns == MAP_FAILED) {
2075 if (!mdb_sec_inited) {
2076 InitializeSecurityDescriptor(&mdb_null_sd,
2077 SECURITY_DESCRIPTOR_REVISION);
2078 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2079 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2080 mdb_all_sa.bInheritHandle = FALSE;
2081 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2084 /* FIXME: only using up to 20 characters of the env path here,
2085 * probably not enough to assure uniqueness...
2087 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%.20s", lpath);
2088 ptr = env->me_txns->mti_rmname + sizeof("Global\\MDBr");
2089 while ((ptr = strchr(ptr, '\\')))
2091 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2092 if (!env->me_rmutex) {
2096 sprintf(env->me_txns->mti_rmname, "Global\\MDBw%.20s", lpath);
2097 ptr = env->me_txns->mti_rmname + sizeof("Global\\MDBw");
2098 while ((ptr = strchr(ptr, '\\')))
2100 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2101 if (!env->me_wmutex) {
2106 pthread_mutexattr_t mattr;
2108 pthread_mutexattr_init(&mattr);
2109 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2113 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2114 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2116 env->me_txns->mti_version = MDB_VERSION;
2117 env->me_txns->mti_magic = MDB_MAGIC;
2118 env->me_txns->mti_txnid = 0;
2119 env->me_txns->mti_numreaders = 0;
2120 env->me_txns->mti_me_toggle = 0;
2123 if (env->me_txns->mti_magic != MDB_MAGIC) {
2124 DPUTS("lock region has invalid magic");
2128 if (env->me_txns->mti_version != MDB_VERSION) {
2129 DPRINTF("lock region is version %u, expected version %u",
2130 env->me_txns->mti_version, MDB_VERSION);
2131 rc = MDB_VERSION_MISMATCH;
2135 if (rc != EACCES && rc != EAGAIN) {
2139 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2140 if (!env->me_rmutex) {
2144 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2145 if (!env->me_wmutex) {
2155 env->me_lfd = INVALID_HANDLE_VALUE;
2160 /** The name of the lock file in the DB environment */
2161 #define LOCKNAME "/lock.mdb"
2162 /** The name of the data file in the DB environment */
2163 #define DATANAME "/data.mdb"
2165 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2167 int oflags, rc, len, excl;
2168 char *lpath, *dpath;
2171 lpath = malloc(len + sizeof(LOCKNAME) + len + sizeof(DATANAME));
2174 dpath = lpath + len + sizeof(LOCKNAME);
2175 sprintf(lpath, "%s" LOCKNAME, path);
2176 sprintf(dpath, "%s" DATANAME, path);
2178 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
2183 if (F_ISSET(flags, MDB_RDONLY)) {
2184 oflags = GENERIC_READ;
2185 len = OPEN_EXISTING;
2187 oflags = GENERIC_READ|GENERIC_WRITE;
2190 mode = FILE_ATTRIBUTE_NORMAL;
2191 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2192 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2197 if (F_ISSET(flags, MDB_RDONLY))
2200 oflags = O_RDWR | O_CREAT;
2202 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
2208 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
2209 /* synchronous fd for meta writes */
2211 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2212 mode |= FILE_FLAG_WRITE_THROUGH;
2213 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2214 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2219 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2220 oflags |= MDB_DSYNC;
2221 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
2226 env->me_path = strdup(path);
2227 DPRINTF("opened dbenv %p", (void *) env);
2228 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
2229 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
2231 mdb_env_share_locks(env);
2232 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
2233 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
2234 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
2240 if (env->me_fd != INVALID_HANDLE_VALUE) {
2242 env->me_fd = INVALID_HANDLE_VALUE;
2244 if (env->me_lfd != INVALID_HANDLE_VALUE) {
2246 env->me_lfd = INVALID_HANDLE_VALUE;
2254 mdb_env_close(MDB_env *env)
2261 while (env->me_dpages) {
2262 dp = env->me_dpages;
2263 env->me_dpages = dp->mp_next;
2267 free(env->me_dbs[1]);
2268 free(env->me_dbs[0]);
2272 LAZY_RWLOCK_DESTROY(&env->me_dblock);
2273 pthread_key_delete(env->me_txkey);
2276 munmap(env->me_map, env->me_mapsize);
2281 pid_t pid = getpid();
2283 for (i=0; i<env->me_txns->mti_numreaders; i++)
2284 if (env->me_txns->mti_readers[i].mr_pid == pid)
2285 env->me_txns->mti_readers[i].mr_pid = 0;
2286 munmap(env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
2292 /* only for aligned ints */
2294 intcmp(const MDB_val *a, const MDB_val *b)
2296 if (a->mv_size == sizeof(long))
2298 unsigned long *la, *lb;
2303 unsigned int *ia, *ib;
2310 /* ints must always be the same size */
2312 cintcmp(const MDB_val *a, const MDB_val *b)
2314 #if __BYTE_ORDER == __LITTLE_ENDIAN
2315 unsigned short *u, *c;
2318 u = a->mv_data + a->mv_size;
2319 c = b->mv_data + a->mv_size;
2322 } while(!x && u > (unsigned short *)a->mv_data);
2325 return memcmp(a->mv_data, b->mv_data, a->mv_size);
2330 memncmp(const MDB_val *a, const MDB_val *b)
2336 len_diff = a->mv_size - b->mv_size;
2339 diff = memcmp(a->mv_data, b->mv_data, len);
2340 return diff ? diff : len_diff;
2344 memnrcmp(const MDB_val *a, const MDB_val *b)
2346 const unsigned char *p1, *p2, *p1_lim;
2349 if (b->mv_size == 0)
2350 return a->mv_size != 0;
2351 if (a->mv_size == 0)
2354 p1 = (const unsigned char *)a->mv_data + a->mv_size - 1;
2355 p2 = (const unsigned char *)b->mv_data + b->mv_size - 1;
2357 len_diff = a->mv_size - b->mv_size;
2359 p1_lim = p1 - a->mv_size;
2361 p1_lim = p1 - b->mv_size;
2363 while (p1 > p1_lim) {
2373 /* Search for key within a leaf page, using binary search.
2374 * Returns the smallest entry larger or equal to the key.
2375 * If exactp is non-null, stores whether the found entry was an exact match
2376 * in *exactp (1 or 0).
2377 * If kip is non-null, stores the index of the found entry in *kip.
2378 * If no entry larger or equal to the key is found, returns NULL.
2381 mdb_search_node(MDB_cursor *mc, MDB_val *key, int *exactp)
2383 unsigned int i = 0, nkeys;
2386 MDB_page *mp = mc->mc_pg[mc->mc_top];
2387 MDB_node *node = NULL;
2392 nkeys = NUMKEYS(mp);
2394 DPRINTF("searching %u keys in %s page %lu",
2395 nkeys, IS_LEAF(mp) ? "leaf" : "branch",
2400 low = IS_LEAF(mp) ? 0 : 1;
2402 cmp = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_cmp;
2404 nodekey.mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2405 node = NODEPTR(mp, 0); /* fake */
2407 while (low <= high) {
2408 i = (low + high) >> 1;
2411 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
2413 node = NODEPTR(mp, i);
2415 nodekey.mv_size = node->mn_ksize;
2416 nodekey.mv_data = NODEKEY(node);
2419 rc = cmp(key, &nodekey);
2423 DPRINTF("found leaf index %u [%s], rc = %i",
2424 i, DKEY(&nodekey), rc);
2426 DPRINTF("found branch index %u [%s -> %lu], rc = %i",
2427 i, DKEY(&nodekey), NODEPGNO(node), rc);
2438 if (rc > 0) { /* Found entry is less than the key. */
2439 i++; /* Skip to get the smallest entry larger than key. */
2441 node = NODEPTR(mp, i);
2444 *exactp = (rc == 0);
2445 /* store the key index */
2446 mc->mc_ki[mc->mc_top] = i;
2448 /* There is no entry larger or equal to the key. */
2451 /* nodeptr is fake for LEAF2 */
2456 cursor_pop_page(MDB_cursor *mc)
2461 top = mc->mc_pg[mc->mc_top];
2466 DPRINTF("popped page %lu off db %u cursor %p", top->mp_pgno,
2467 mc->mc_dbi, (void *) mc);
2472 cursor_push_page(MDB_cursor *mc, MDB_page *mp)
2474 DPRINTF("pushing page %lu on db %u cursor %p", mp->mp_pgno,
2475 mc->mc_dbi, (void *) mc);
2477 if (mc->mc_snum >= CURSOR_STACK)
2480 mc->mc_top = mc->mc_snum++;
2481 mc->mc_pg[mc->mc_top] = mp;
2482 mc->mc_ki[mc->mc_top] = 0;
2488 mdb_get_page(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
2492 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
2494 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
2495 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
2496 p = txn->mt_u.dirty_list[x].mptr;
2500 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
2501 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
2505 DPRINTF("page %lu not found", pgno);
2508 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
2512 mdb_search_page_root(MDB_cursor *mc, MDB_val *key, int modify)
2514 MDB_page *mp = mc->mc_pg[mc->mc_top];
2519 while (IS_BRANCH(mp)) {
2522 DPRINTF("branch page %lu has %u keys", mp->mp_pgno, NUMKEYS(mp));
2523 assert(NUMKEYS(mp) > 1);
2524 DPRINTF("found index 0 to page %lu", NODEPGNO(NODEPTR(mp, 0)));
2526 if (key == NULL) /* Initialize cursor to first page. */
2527 mc->mc_ki[mc->mc_top] = 0;
2528 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
2529 /* cursor to last page */
2530 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
2533 node = mdb_search_node(mc, key, &exact);
2535 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
2537 assert(mc->mc_ki[mc->mc_top] > 0);
2538 mc->mc_ki[mc->mc_top]--;
2543 DPRINTF("following index %u for key [%s]",
2544 mc->mc_ki[mc->mc_top], DKEY(key));
2545 assert(mc->mc_ki[mc->mc_top] < NUMKEYS(mp));
2546 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2548 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mp)))
2551 if ((rc = cursor_push_page(mc, mp)))
2555 if ((rc = mdb_touch(mc)) != 0)
2557 mp = mc->mc_pg[mc->mc_top];
2562 DPRINTF("internal error, index points to a %02X page!?",
2564 return MDB_CORRUPTED;
2567 DPRINTF("found leaf page %lu for key [%s]", mp->mp_pgno,
2568 key ? DKEY(key) : NULL);
2573 /* Search for the page a given key should be in.
2574 * Pushes parent pages on the cursor stack.
2575 * If key is NULL, search for the lowest page (used by mdb_cursor_first).
2576 * If modify is true, visited pages are updated with new page numbers.
2579 mdb_search_page(MDB_cursor *mc, MDB_val *key, int modify)
2584 /* Make sure the txn is still viable, then find the root from
2585 * the txn's db table.
2587 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
2588 DPUTS("transaction has failed, must abort");
2591 root = mc->mc_txn->mt_dbs[mc->mc_dbi].md_root;
2593 if (root == P_INVALID) { /* Tree is empty. */
2594 DPUTS("tree is empty");
2595 return MDB_NOTFOUND;
2598 if ((rc = mdb_get_page(mc->mc_txn, root, &mc->mc_pg[0])))
2604 DPRINTF("db %u root page %lu has flags 0x%X",
2605 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
2608 /* For sub-databases, update main root first */
2609 if (mc->mc_dbi > MAIN_DBI && !mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty) {
2611 mc2.mc_txn = mc->mc_txn;
2612 mc2.mc_dbi = MAIN_DBI;
2613 rc = mdb_search_page(&mc2, &mc->mc_txn->mt_dbxs[mc->mc_dbi].md_name, 1);
2616 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
2618 if (!F_ISSET(mc->mc_pg[0]->mp_flags, P_DIRTY)) {
2619 if ((rc = mdb_touch(mc)))
2621 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = mc->mc_pg[0]->mp_pgno;
2625 return mdb_search_page_root(mc, key, modify);
2629 mdb_read_data(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
2631 MDB_page *omp; /* overflow mpage */
2635 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
2636 data->mv_size = NODEDSZ(leaf);
2637 data->mv_data = NODEDATA(leaf);
2641 /* Read overflow data.
2643 data->mv_size = NODEDSZ(leaf);
2644 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
2645 if ((rc = mdb_get_page(txn, pgno, &omp))) {
2646 DPRINTF("read overflow page %lu failed", pgno);
2649 data->mv_data = METADATA(omp);
2655 mdb_get(MDB_txn *txn, MDB_dbi dbi,
2656 MDB_val *key, MDB_val *data)
2665 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
2667 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
2670 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
2677 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
2678 mc.mc_xcursor = &mx;
2679 mdb_xcursor_init0(&mc);
2681 mc.mc_xcursor = NULL;
2683 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
2687 mdb_sibling(MDB_cursor *mc, int move_right)
2694 if (mc->mc_snum < 2) {
2695 return MDB_NOTFOUND; /* root has no siblings */
2697 ptop = mc->mc_top-1;
2699 DPRINTF("parent page is page %lu, index %u",
2700 mc->mc_pg[ptop]->mp_pgno, mc->mc_ki[ptop]);
2702 cursor_pop_page(mc);
2703 if (move_right ? (mc->mc_ki[ptop] + 1u >= NUMKEYS(mc->mc_pg[ptop]))
2704 : (mc->mc_ki[ptop] == 0)) {
2705 DPRINTF("no more keys left, moving to %s sibling",
2706 move_right ? "right" : "left");
2707 if ((rc = mdb_sibling(mc, move_right)) != MDB_SUCCESS)
2714 DPRINTF("just moving to %s index key %u",
2715 move_right ? "right" : "left", mc->mc_ki[ptop]);
2717 assert(IS_BRANCH(mc->mc_pg[ptop]));
2719 indx = NODEPTR(mc->mc_pg[ptop], mc->mc_ki[ptop]);
2720 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(indx), &mp)))
2723 cursor_push_page(mc, mp);
2729 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
2735 if (mc->mc_flags & C_EOF) {
2736 return MDB_NOTFOUND;
2739 assert(mc->mc_flags & C_INITIALIZED);
2741 mp = mc->mc_pg[mc->mc_top];
2743 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) {
2744 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2745 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2746 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
2747 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
2748 if (op != MDB_NEXT || rc == MDB_SUCCESS)
2752 mc->mc_xcursor->mx_cursor.mc_flags = 0;
2753 if (op == MDB_NEXT_DUP)
2754 return MDB_NOTFOUND;
2758 DPRINTF("cursor_next: top page is %lu in cursor %p", mp->mp_pgno, (void *) mc);
2760 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
2761 DPUTS("=====> move to next sibling page");
2762 if (mdb_sibling(mc, 1) != MDB_SUCCESS) {
2763 mc->mc_flags |= C_EOF;
2764 return MDB_NOTFOUND;
2766 mp = mc->mc_pg[mc->mc_top];
2767 DPRINTF("next page is %lu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
2769 mc->mc_ki[mc->mc_top]++;
2771 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
2772 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
2775 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2776 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
2780 assert(IS_LEAF(mp));
2781 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2783 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2784 mdb_xcursor_init1(mc, leaf);
2787 if ((rc = mdb_read_data(mc->mc_txn, leaf, data) != MDB_SUCCESS))
2790 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2791 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
2792 if (rc != MDB_SUCCESS)
2797 MDB_SET_KEY(leaf, key);
2802 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
2808 assert(mc->mc_flags & C_INITIALIZED);
2810 mp = mc->mc_pg[mc->mc_top];
2812 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) {
2813 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2814 if (op == MDB_PREV || op == MDB_PREV_DUP) {
2815 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2816 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
2817 if (op != MDB_PREV || rc == MDB_SUCCESS)
2820 mc->mc_xcursor->mx_cursor.mc_flags = 0;
2821 if (op == MDB_PREV_DUP)
2822 return MDB_NOTFOUND;
2827 DPRINTF("cursor_prev: top page is %lu in cursor %p", mp->mp_pgno, (void *) mc);
2829 if (mc->mc_ki[mc->mc_top] == 0) {
2830 DPUTS("=====> move to prev sibling page");
2831 if (mdb_sibling(mc, 0) != MDB_SUCCESS) {
2832 mc->mc_flags &= ~C_INITIALIZED;
2833 return MDB_NOTFOUND;
2835 mp = mc->mc_pg[mc->mc_top];
2836 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
2837 DPRINTF("prev page is %lu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
2839 mc->mc_ki[mc->mc_top]--;
2841 mc->mc_flags &= ~C_EOF;
2843 DPRINTF("==> cursor points to page %lu with %u keys, key index %u",
2844 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
2847 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2848 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
2852 assert(IS_LEAF(mp));
2853 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
2855 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2856 mdb_xcursor_init1(mc, leaf);
2859 if ((rc = mdb_read_data(mc->mc_txn, leaf, data) != MDB_SUCCESS))
2862 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2863 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
2864 if (rc != MDB_SUCCESS)
2869 MDB_SET_KEY(leaf, key);
2874 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
2875 MDB_cursor_op op, int *exactp)
2883 assert(key->mv_size > 0);
2885 /* See if we're already on the right page */
2886 if (mc->mc_flags & C_INITIALIZED) {
2889 if (mc->mc_pg[mc->mc_top]->mp_flags & P_LEAF2) {
2890 nodekey.mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2891 nodekey.mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, nodekey.mv_size);
2893 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
2894 MDB_SET_KEY(leaf, &nodekey);
2896 rc = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_cmp(key, &nodekey);
2898 /* Probably happens rarely, but first node on the page
2899 * was the one we wanted.
2901 mc->mc_ki[mc->mc_top] = 0;
2905 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
2910 if (NUMKEYS(mc->mc_pg[mc->mc_top]) > 1) {
2911 if (mc->mc_pg[mc->mc_top]->mp_flags & P_LEAF2) {
2912 nodekey.mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top],
2913 NUMKEYS(mc->mc_pg[mc->mc_top])-1, nodekey.mv_size);
2915 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
2916 MDB_SET_KEY(leaf, &nodekey);
2918 rc = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_cmp(key, &nodekey);
2920 /* last node was the one we wanted */
2921 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top])-1;
2925 /* This is definitely the right page, skip search_page */
2930 /* If any parents have right-sibs, search.
2931 * Otherwise, there's nothing further.
2933 for (i=0; i<mc->mc_top; i++)
2935 NUMKEYS(mc->mc_pg[i])-1)
2937 if (i == mc->mc_top) {
2938 /* There are no other pages */
2939 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
2940 return MDB_NOTFOUND;
2945 rc = mdb_search_page(mc, key, 0);
2946 if (rc != MDB_SUCCESS)
2949 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
2952 leaf = mdb_search_node(mc, key, exactp);
2953 if (exactp != NULL && !*exactp) {
2954 /* MDB_SET specified and not an exact match. */
2955 return MDB_NOTFOUND;
2959 DPUTS("===> inexact leaf not found, goto sibling");
2960 if ((rc = mdb_sibling(mc, 1)) != MDB_SUCCESS)
2961 return rc; /* no entries matched */
2962 mc->mc_ki[mc->mc_top] = 0;
2963 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
2964 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
2968 mc->mc_flags |= C_INITIALIZED;
2969 mc->mc_flags &= ~C_EOF;
2971 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
2972 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
2973 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
2977 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2978 mdb_xcursor_init1(mc, leaf);
2981 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
2982 if (op == MDB_SET || op == MDB_SET_RANGE) {
2983 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
2986 if (op == MDB_GET_BOTH) {
2992 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
2993 if (rc != MDB_SUCCESS)
2996 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
2998 if ((rc = mdb_read_data(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
3000 rc = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dcmp(data, &d2);
3002 if (op == MDB_GET_BOTH || rc > 0)
3003 return MDB_NOTFOUND;
3007 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3012 /* The key already matches in all other cases */
3013 if (op == MDB_SET_RANGE)
3014 MDB_SET_KEY(leaf, key);
3015 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3021 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3026 rc = mdb_search_page(mc, NULL, 0);
3027 if (rc != MDB_SUCCESS)
3029 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3031 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3032 mc->mc_flags |= C_INITIALIZED;
3033 mc->mc_flags &= ~C_EOF;
3035 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3036 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
3037 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
3042 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3043 mdb_xcursor_init1(mc, leaf);
3044 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3049 mc->mc_xcursor->mx_cursor.mc_flags = 0;
3050 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3054 MDB_SET_KEY(leaf, key);
3059 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3065 lkey.mv_size = MAXKEYSIZE+1;
3066 lkey.mv_data = NULL;
3068 rc = mdb_search_page(mc, &lkey, 0);
3069 if (rc != MDB_SUCCESS)
3071 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3073 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
3074 mc->mc_flags |= C_INITIALIZED;
3075 mc->mc_flags &= ~C_EOF;
3077 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
3079 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3080 key->mv_size = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
3081 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
3086 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3087 mdb_xcursor_init1(mc, leaf);
3088 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3092 if ((rc = mdb_read_data(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3097 MDB_SET_KEY(leaf, key);
3102 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3112 case MDB_GET_BOTH_RANGE:
3113 if (data == NULL || mc->mc_xcursor == NULL) {
3120 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3122 } else if (op == MDB_SET_RANGE)
3123 rc = mdb_cursor_set(mc, key, data, op, NULL);
3125 rc = mdb_cursor_set(mc, key, data, op, &exact);
3127 case MDB_GET_MULTIPLE:
3129 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPFIXED) ||
3130 !(mc->mc_flags & C_INITIALIZED)) {
3135 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
3136 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
3139 case MDB_NEXT_MULTIPLE:
3141 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPFIXED)) {
3145 if (!(mc->mc_flags & C_INITIALIZED))
3146 rc = mdb_cursor_first(mc, key, data);
3148 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
3149 if (rc == MDB_SUCCESS) {
3150 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
3153 mx = &mc->mc_xcursor->mx_cursor;
3154 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
3155 mx->mc_txn->mt_dbs[mx->mc_dbi].md_pad;
3156 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
3157 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
3165 case MDB_NEXT_NODUP:
3166 if (!(mc->mc_flags & C_INITIALIZED))
3167 rc = mdb_cursor_first(mc, key, data);
3169 rc = mdb_cursor_next(mc, key, data, op);
3173 case MDB_PREV_NODUP:
3174 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
3175 rc = mdb_cursor_last(mc, key, data);
3177 rc = mdb_cursor_prev(mc, key, data, op);
3180 rc = mdb_cursor_first(mc, key, data);
3184 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) ||
3185 !(mc->mc_flags & C_INITIALIZED) ||
3186 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3190 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3193 rc = mdb_cursor_last(mc, key, data);
3197 !(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT) ||
3198 !(mc->mc_flags & C_INITIALIZED) ||
3199 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3203 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3206 DPRINTF("unhandled/unimplemented cursor operation %u", op);
3215 mdb_cursor_touch(MDB_cursor *mc)
3219 if (mc->mc_dbi > MAIN_DBI && !mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty) {
3221 mc2.mc_txn = mc->mc_txn;
3222 mc2.mc_dbi = MAIN_DBI;
3223 rc = mdb_search_page(&mc2, &mc->mc_txn->mt_dbxs[mc->mc_dbi].md_name, 1);
3225 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
3227 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
3228 if (!F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) {
3232 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root =
3233 mc->mc_pg[mc->mc_top]->mp_pgno;
3237 mc->mc_top = mc->mc_snum-1;
3242 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3246 MDB_val xdata, *rdata, dkey;
3248 char dbuf[PAGESIZE];
3254 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
3257 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
3258 mc->mc_dbi, DKEY(key), key->mv_size, data->mv_size);
3262 if (flags == MDB_CURRENT) {
3263 if (!(mc->mc_flags & C_INITIALIZED))
3266 } else if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_root == P_INVALID) {
3268 /* new database, write a root leaf page */
3269 DPUTS("allocating new root leaf page");
3270 if ((np = mdb_new_page(mc, P_LEAF, 1)) == NULL) {
3274 cursor_push_page(mc, np);
3275 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = np->mp_pgno;
3276 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth++;
3277 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = 1;
3278 if ((mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
3280 np->mp_flags |= P_LEAF2;
3281 mc->mc_flags |= C_INITIALIZED;
3287 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
3288 if (flags == MDB_NOOVERWRITE && rc == 0) {
3289 DPRINTF("duplicate key [%s]", DKEY(key));
3291 return MDB_KEYEXIST;
3293 if (rc && rc != MDB_NOTFOUND)
3297 /* Cursor is positioned, now make sure all pages are writable */
3298 rc2 = mdb_cursor_touch(mc);
3299 if (rc2) return rc2;
3302 /* The key already exists */
3303 if (rc == MDB_SUCCESS) {
3304 /* there's only a key anyway, so this is a no-op */
3305 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3306 unsigned int ksize = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
3307 if (key->mv_size != ksize)
3309 if (flags == MDB_CURRENT) {
3310 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
3311 memcpy(ptr, key->mv_data, ksize);
3316 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3319 if (F_ISSET(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags, MDB_DUPSORT)) {
3320 /* Was a single item before, must convert now */
3321 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3322 dkey.mv_size = NODEDSZ(leaf);
3323 dkey.mv_data = dbuf;
3324 memcpy(dbuf, NODEDATA(leaf), dkey.mv_size);
3325 /* data matches, ignore it */
3326 if (!mdb_dcmp(mc->mc_txn, mc->mc_dbi, data, &dkey))
3327 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
3328 memset(&dummy, 0, sizeof(dummy));
3329 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPFIXED) {
3330 dummy.md_pad = data->mv_size;
3331 dummy.md_flags = MDB_DUPFIXED;
3332 if (mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_INTEGERDUP)
3333 dummy.md_flags |= MDB_INTEGERKEY;
3335 dummy.md_root = P_INVALID;
3336 if (dkey.mv_size == sizeof(MDB_db)) {
3337 memcpy(NODEDATA(leaf), &dummy, sizeof(dummy));
3340 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
3343 xdata.mv_size = sizeof(MDB_db);
3344 xdata.mv_data = &dummy;
3345 /* new sub-DB, must fully init xcursor */
3346 if (flags == MDB_CURRENT)
3352 /* same size, just replace it */
3353 if (!F_ISSET(leaf->mn_flags, F_BIGDATA) &&
3354 NODEDSZ(leaf) == data->mv_size) {
3355 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
3358 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
3360 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
3366 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
3367 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
3368 rc = mdb_split(mc, key, rdata, P_INVALID);
3370 /* There is room already in this leaf page. */
3371 rc = mdb_add_node(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, 0);
3374 if (rc != MDB_SUCCESS)
3375 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
3377 /* Remember if we just added a subdatabase */
3378 if (flags & F_SUBDATA) {
3379 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3380 leaf->mn_flags |= F_SUBDATA;
3383 /* Now store the actual data in the child DB. Note that we're
3384 * storing the user data in the keys field, so there are strict
3385 * size limits on dupdata. The actual data fields of the child
3386 * DB are all zero size.
3389 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3391 if (flags == MDB_CURRENT)
3392 mdb_xcursor_init2(mc);
3394 mdb_xcursor_init1(mc, leaf);
3397 if (flags == MDB_NODUPDATA)
3398 flags = MDB_NOOVERWRITE;
3399 /* converted, write the original data first */
3401 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, flags);
3403 leaf->mn_flags |= F_DUPDATA;
3405 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, flags);
3406 mdb_xcursor_fini(mc);
3407 memcpy(NODEDATA(leaf),
3408 &mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi],
3411 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries++;
3418 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
3423 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
3426 if (!mc->mc_flags & C_INITIALIZED)
3429 rc = mdb_cursor_touch(mc);
3432 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3434 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3435 if (flags != MDB_NODUPDATA) {
3436 mdb_xcursor_init2(mc);
3437 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
3438 mdb_xcursor_fini(mc);
3439 /* If sub-DB still has entries, we're done */
3440 if (mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_root
3442 memcpy(NODEDATA(leaf),
3443 &mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi],
3445 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries--;
3448 /* otherwise fall thru and delete the sub-DB */
3451 /* add all the child DB's pages to the free list */
3452 rc = mdb_search_page(&mc->mc_xcursor->mx_cursor, NULL, 0);
3453 if (rc == MDB_SUCCESS) {
3458 mx = &mc->mc_xcursor->mx_cursor;
3459 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries -=
3460 mx->mc_txn->mt_dbs[mx->mc_dbi].md_entries;
3462 cursor_pop_page(mx);
3464 while (mx->mc_snum > 1) {
3465 for (i=0; i<NUMKEYS(mx->mc_pg[mx->mc_top]); i++) {
3467 ni = NODEPTR(mx->mc_pg[mx->mc_top], i);
3470 mdb_midl_append(mc->mc_txn->mt_free_pgs, pg);
3472 rc = mdb_sibling(mx, 1);
3477 mdb_midl_append(mc->mc_txn->mt_free_pgs,
3478 mx->mc_txn->mt_dbs[mx->mc_dbi].md_root);
3482 return mdb_del0(mc, leaf);
3485 /* Allocate a page and initialize it
3488 mdb_new_page(MDB_cursor *mc, uint32_t flags, int num)
3492 if ((np = mdb_alloc_page(mc, num)) == NULL)
3494 DPRINTF("allocated new mpage %lu, page size %u",
3495 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
3496 np->mp_flags = flags | P_DIRTY;
3497 np->mp_lower = PAGEHDRSZ;
3498 np->mp_upper = mc->mc_txn->mt_env->me_psize;
3501 mc->mc_txn->mt_dbs[mc->mc_dbi].md_branch_pages++;
3502 else if (IS_LEAF(np))
3503 mc->mc_txn->mt_dbs[mc->mc_dbi].md_leaf_pages++;
3504 else if (IS_OVERFLOW(np)) {
3505 mc->mc_txn->mt_dbs[mc->mc_dbi].md_overflow_pages += num;
3513 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
3517 sz = LEAFSIZE(key, data);
3518 if (data->mv_size >= env->me_psize / MDB_MINKEYS) {
3519 /* put on overflow page */
3520 sz -= data->mv_size - sizeof(pgno_t);
3524 return sz + sizeof(indx_t);
3528 mdb_branch_size(MDB_env *env, MDB_val *key)
3533 if (sz >= env->me_psize / MDB_MINKEYS) {
3534 /* put on overflow page */
3535 /* not implemented */
3536 /* sz -= key->size - sizeof(pgno_t); */
3539 return sz + sizeof(indx_t);
3543 mdb_add_node(MDB_cursor *mc, indx_t indx,
3544 MDB_val *key, MDB_val *data, pgno_t pgno, uint8_t flags)
3547 size_t node_size = NODESIZE;
3550 MDB_page *mp = mc->mc_pg[mc->mc_top];
3551 MDB_page *ofp = NULL; /* overflow page */
3554 assert(mp->mp_upper >= mp->mp_lower);
3556 DPRINTF("add to %s page %lu index %i, data size %zu key size %zu [%s]",
3557 IS_LEAF(mp) ? "leaf" : "branch",
3558 mp->mp_pgno, indx, data ? data->mv_size : 0,
3559 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
3562 /* Move higher keys up one slot. */
3563 int ksize = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad, dif;
3564 char *ptr = LEAF2KEY(mp, indx, ksize);
3565 dif = NUMKEYS(mp) - indx;
3567 memmove(ptr+ksize, ptr, dif*ksize);
3568 /* insert new key */
3569 memcpy(ptr, key->mv_data, ksize);
3571 /* Just using these for counting */
3572 mp->mp_lower += sizeof(indx_t);
3573 mp->mp_upper -= ksize - sizeof(indx_t);
3578 node_size += key->mv_size;
3582 if (F_ISSET(flags, F_BIGDATA)) {
3583 /* Data already on overflow page. */
3584 node_size += sizeof(pgno_t);
3585 } else if (data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
3586 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
3587 /* Put data on overflow page. */
3588 DPRINTF("data size is %zu, put on overflow page",
3590 node_size += sizeof(pgno_t);
3591 if ((ofp = mdb_new_page(mc, P_OVERFLOW, ovpages)) == NULL)
3593 DPRINTF("allocated overflow page %lu", ofp->mp_pgno);
3596 node_size += data->mv_size;
3599 node_size += node_size & 1;
3601 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
3602 DPRINTF("not enough room in page %lu, got %u ptrs",
3603 mp->mp_pgno, NUMKEYS(mp));
3604 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
3605 mp->mp_upper - mp->mp_lower);
3606 DPRINTF("node size = %zu", node_size);
3610 /* Move higher pointers up one slot. */
3611 for (i = NUMKEYS(mp); i > indx; i--)
3612 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
3614 /* Adjust free space offsets. */
3615 ofs = mp->mp_upper - node_size;
3616 assert(ofs >= mp->mp_lower + sizeof(indx_t));
3617 mp->mp_ptrs[indx] = ofs;
3619 mp->mp_lower += sizeof(indx_t);
3621 /* Write the node data. */
3622 node = NODEPTR(mp, indx);
3623 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
3624 node->mn_flags = flags;
3626 SETDSZ(node,data->mv_size);
3631 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
3636 if (F_ISSET(flags, F_BIGDATA))
3637 memcpy(node->mn_data + key->mv_size, data->mv_data,
3640 memcpy(node->mn_data + key->mv_size, data->mv_data,
3643 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
3645 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
3653 mdb_del_node(MDB_page *mp, indx_t indx, int ksize)
3656 indx_t i, j, numkeys, ptr;
3660 DPRINTF("delete node %u on %s page %lu", indx,
3661 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno);
3662 assert(indx < NUMKEYS(mp));
3665 int x = NUMKEYS(mp) - 1 - indx;
3666 base = LEAF2KEY(mp, indx, ksize);
3668 memmove(base, base + ksize, x * ksize);
3669 mp->mp_lower -= sizeof(indx_t);
3670 mp->mp_upper += ksize - sizeof(indx_t);
3674 node = NODEPTR(mp, indx);
3675 sz = NODESIZE + node->mn_ksize;
3677 if (F_ISSET(node->mn_flags, F_BIGDATA))
3678 sz += sizeof(pgno_t);
3680 sz += NODEDSZ(node);
3684 ptr = mp->mp_ptrs[indx];
3685 numkeys = NUMKEYS(mp);
3686 for (i = j = 0; i < numkeys; i++) {
3688 mp->mp_ptrs[j] = mp->mp_ptrs[i];
3689 if (mp->mp_ptrs[i] < ptr)
3690 mp->mp_ptrs[j] += sz;
3695 base = (char *)mp + mp->mp_upper;
3696 memmove(base + sz, base, ptr - mp->mp_upper);
3698 mp->mp_lower -= sizeof(indx_t);
3703 mdb_xcursor_init0(MDB_cursor *mc)
3705 MDB_xcursor *mx = mc->mc_xcursor;
3708 mx->mx_txn = *mc->mc_txn;
3709 mx->mx_txn.mt_dbxs = mx->mx_dbxs;
3710 mx->mx_txn.mt_dbs = mx->mx_dbs;
3711 mx->mx_dbxs[0] = mc->mc_txn->mt_dbxs[0];
3712 mx->mx_dbxs[1] = mc->mc_txn->mt_dbxs[1];
3713 if (mc->mc_dbi > 1) {
3714 mx->mx_dbxs[2] = mc->mc_txn->mt_dbxs[mc->mc_dbi];
3719 mx->mx_dbxs[dbn+1].md_parent = dbn;
3720 mx->mx_dbxs[dbn+1].md_cmp = mx->mx_dbxs[dbn].md_dcmp;
3721 mx->mx_dbxs[dbn+1].md_rel = mx->mx_dbxs[dbn].md_rel;
3722 mx->mx_dbxs[dbn+1].md_dirty = 0;
3723 mx->mx_txn.mt_numdbs = dbn+2;
3724 mx->mx_txn.mt_u = mc->mc_txn->mt_u;
3726 mx->mx_cursor.mc_xcursor = NULL;
3727 mx->mx_cursor.mc_txn = &mx->mx_txn;
3728 mx->mx_cursor.mc_dbi = dbn+1;
3732 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
3734 MDB_db *db = NODEDATA(node);
3735 MDB_xcursor *mx = mc->mc_xcursor;
3737 mx->mx_dbs[0] = mc->mc_txn->mt_dbs[0];
3738 mx->mx_dbs[1] = mc->mc_txn->mt_dbs[1];
3739 if (mc->mc_dbi > 1) {
3740 mx->mx_dbs[2] = mc->mc_txn->mt_dbs[mc->mc_dbi];
3741 mx->mx_dbxs[2].md_dirty = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty;
3746 DPRINTF("Sub-db %u for db %u root page %lu", dbn, mc->mc_dbi, db->md_root);
3747 mx->mx_dbs[dbn] = *db;
3748 if (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY))
3749 mx->mx_dbxs[dbn].md_dirty = 1;
3750 mx->mx_dbxs[dbn].md_name.mv_data = NODEKEY(node);
3751 mx->mx_dbxs[dbn].md_name.mv_size = node->mn_ksize;
3752 mx->mx_txn.mt_next_pgno = mc->mc_txn->mt_next_pgno;
3753 mx->mx_cursor.mc_snum = 0;
3754 mx->mx_cursor.mc_flags = 0;
3758 mdb_xcursor_init2(MDB_cursor *mc)
3760 MDB_xcursor *mx = mc->mc_xcursor;
3762 mx->mx_dbs[0] = mc->mc_txn->mt_dbs[0];
3763 mx->mx_dbs[1] = mc->mc_txn->mt_dbs[1];
3764 if (mc->mc_dbi > 1) {
3765 mx->mx_dbs[2] = mc->mc_txn->mt_dbs[mc->mc_dbi];
3766 mx->mx_dbxs[2].md_dirty = mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty;
3771 DPRINTF("Sub-db %u for db %u root page %lu", dbn, mc->mc_dbi,
3772 mx->mx_dbs[dbn].md_root);
3773 mx->mx_txn.mt_next_pgno = mc->mc_txn->mt_next_pgno;
3777 mdb_xcursor_fini(MDB_cursor *mc)
3779 MDB_xcursor *mx = mc->mc_xcursor;
3780 mc->mc_txn->mt_next_pgno = mx->mx_txn.mt_next_pgno;
3781 mc->mc_txn->mt_dbs[0] = mx->mx_dbs[0];
3782 mc->mc_txn->mt_dbs[1] = mx->mx_dbs[1];
3783 if (mc->mc_dbi > 1) {
3784 mc->mc_txn->mt_dbs[mc->mc_dbi] = mx->mx_dbs[2];
3785 mc->mc_txn->mt_dbxs[mc->mc_dbi].md_dirty = mx->mx_dbxs[2].md_dirty;
3790 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
3793 size_t size = sizeof(MDB_cursor);
3795 if (txn == NULL || ret == NULL || !dbi || dbi >= txn->mt_numdbs)
3798 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
3799 size += sizeof(MDB_xcursor);
3801 if ((mc = calloc(1, size)) != NULL) {
3804 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
3805 MDB_xcursor *mx = (MDB_xcursor *)(mc + 1);
3806 mc->mc_xcursor = mx;
3807 mdb_xcursor_init0(mc);
3818 /* Return the count of duplicate data items for the current key */
3820 mdb_cursor_count(MDB_cursor *mc, unsigned long *countp)
3824 if (mc == NULL || countp == NULL)
3827 if (!(mc->mc_txn->mt_dbs[mc->mc_dbi].md_flags & MDB_DUPSORT))
3830 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3831 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3834 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
3837 *countp = mc->mc_xcursor->mx_txn.mt_dbs[mc->mc_xcursor->mx_cursor.mc_dbi].md_entries;
3843 mdb_cursor_close(MDB_cursor *mc)
3851 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
3853 indx_t ptr, i, numkeys;
3860 node = NODEPTR(mp, indx);
3861 ptr = mp->mp_ptrs[indx];
3862 DPRINTF("update key %u (ofs %u) [%.*s] to [%s] on page %lu",
3864 (int)node->mn_ksize, (char *)NODEKEY(node),
3868 delta = key->mv_size - node->mn_ksize;
3870 if (delta > 0 && SIZELEFT(mp) < delta) {
3871 DPRINTF("OUCH! Not enough room, delta = %d", delta);
3875 numkeys = NUMKEYS(mp);
3876 for (i = 0; i < numkeys; i++) {
3877 if (mp->mp_ptrs[i] <= ptr)
3878 mp->mp_ptrs[i] -= delta;
3881 base = (char *)mp + mp->mp_upper;
3882 len = ptr - mp->mp_upper + NODESIZE;
3883 memmove(base - delta, base, len);
3884 mp->mp_upper -= delta;
3886 node = NODEPTR(mp, indx);
3887 node->mn_ksize = key->mv_size;
3890 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
3895 /* Move a node from csrc to cdst.
3898 mdb_move_node(MDB_cursor *csrc, MDB_cursor *cdst)
3905 /* Mark src and dst as dirty. */
3906 if ((rc = mdb_touch(csrc)) ||
3907 (rc = mdb_touch(cdst)))
3910 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3911 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
3912 key.mv_size = csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_pad;
3913 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3915 data.mv_data = NULL;
3917 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
3918 unsigned int snum = csrc->mc_snum;
3919 /* must find the lowest key below src */
3920 mdb_search_page_root(csrc, NULL, 0);
3921 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
3922 csrc->mc_snum = snum--;
3923 csrc->mc_top = snum;
3925 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
3927 key.mv_size = NODEKSZ(srcnode);
3928 key.mv_data = NODEKEY(srcnode);
3929 data.mv_size = NODEDSZ(srcnode);
3930 data.mv_data = NODEDATA(srcnode);
3932 DPRINTF("moving %s node %u [%s] on page %lu to node %u on page %lu",
3933 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
3934 csrc->mc_ki[csrc->mc_top],
3936 csrc->mc_pg[csrc->mc_top]->mp_pgno,
3937 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
3939 /* Add the node to the destination page.
3941 rc = mdb_add_node(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, NODEPGNO(srcnode),
3943 if (rc != MDB_SUCCESS)
3946 /* Delete the node from the source page.
3948 mdb_del_node(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3950 /* Update the parent separators.
3952 if (csrc->mc_ki[csrc->mc_top] == 0) {
3953 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
3954 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3955 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
3957 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
3958 key.mv_size = NODEKSZ(srcnode);
3959 key.mv_data = NODEKEY(srcnode);
3961 DPRINTF("update separator for source page %lu to [%s]",
3962 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
3963 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
3964 &key)) != MDB_SUCCESS)
3967 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
3969 nullkey.mv_size = 0;
3970 assert(mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey) == MDB_SUCCESS);
3974 if (cdst->mc_ki[cdst->mc_top] == 0) {
3975 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
3976 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
3977 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
3979 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
3980 key.mv_size = NODEKSZ(srcnode);
3981 key.mv_data = NODEKEY(srcnode);
3983 DPRINTF("update separator for destination page %lu to [%s]",
3984 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
3985 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
3986 &key)) != MDB_SUCCESS)
3989 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
3991 nullkey.mv_size = 0;
3992 assert(mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey) == MDB_SUCCESS);
4000 mdb_merge(MDB_cursor *csrc, MDB_cursor *cdst)
4007 DPRINTF("merging page %lu into %lu", csrc->mc_pg[csrc->mc_top]->mp_pgno, cdst->mc_pg[cdst->mc_top]->mp_pgno);
4009 assert(csrc->mc_snum > 1); /* can't merge root page */
4010 assert(cdst->mc_snum > 1);
4012 /* Mark dst as dirty. */
4013 if ((rc = mdb_touch(cdst)))
4016 /* Move all nodes from src to dst.
4018 j = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
4019 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
4020 key.mv_size = csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_pad;
4021 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
4022 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
4023 rc = mdb_add_node(cdst, j, &key, NULL, 0, 0);
4024 if (rc != MDB_SUCCESS)
4026 key.mv_data = (char *)key.mv_data + key.mv_size;
4029 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
4030 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
4032 key.mv_size = srcnode->mn_ksize;
4033 key.mv_data = NODEKEY(srcnode);
4034 data.mv_size = NODEDSZ(srcnode);
4035 data.mv_data = NODEDATA(srcnode);
4036 rc = mdb_add_node(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
4037 if (rc != MDB_SUCCESS)
4042 DPRINTF("dst page %lu now has %u keys (%.1f%% filled)",
4043 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);
4045 /* Unlink the src page from parent and add to free list.
4047 mdb_del_node(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
4048 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
4050 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
4054 mdb_midl_append(csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
4055 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
4056 csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_leaf_pages--;
4058 csrc->mc_txn->mt_dbs[csrc->mc_dbi].md_branch_pages--;
4059 cursor_pop_page(csrc);
4061 return mdb_rebalance(csrc);
4065 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
4069 cdst->mc_txn = csrc->mc_txn;
4070 cdst->mc_dbi = csrc->mc_dbi;
4071 cdst->mc_snum = csrc->mc_snum;
4072 cdst->mc_top = csrc->mc_top;
4073 cdst->mc_flags = csrc->mc_flags;
4075 for (i=0; i<csrc->mc_snum; i++) {
4076 cdst->mc_pg[i] = csrc->mc_pg[i];
4077 cdst->mc_ki[i] = csrc->mc_ki[i];
4082 mdb_rebalance(MDB_cursor *mc)
4090 DPRINTF("rebalancing %s page %lu (has %u keys, %.1f%% full)",
4091 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
4092 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);
4094 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
4095 DPRINTF("no need to rebalance page %lu, above fill threshold",
4096 mc->mc_pg[mc->mc_top]->mp_pgno);
4100 if (mc->mc_snum < 2) {
4101 if (NUMKEYS(mc->mc_pg[mc->mc_top]) == 0) {
4102 DPUTS("tree is completely empty");
4103 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = P_INVALID;
4104 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth = 0;
4105 mc->mc_txn->mt_dbs[mc->mc_dbi].md_leaf_pages = 0;
4106 mdb_midl_append(mc->mc_txn->mt_free_pgs, mc->mc_pg[mc->mc_top]->mp_pgno);
4107 } else if (IS_BRANCH(mc->mc_pg[mc->mc_top]) && NUMKEYS(mc->mc_pg[mc->mc_top]) == 1) {
4108 DPUTS("collapsing root page!");
4109 mdb_midl_append(mc->mc_txn->mt_free_pgs, mc->mc_pg[mc->mc_top]->mp_pgno);
4110 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = NODEPGNO(NODEPTR(mc->mc_pg[mc->mc_top], 0));
4111 if ((rc = mdb_get_page(mc->mc_txn, mc->mc_txn->mt_dbs[mc->mc_dbi].md_root, &root)))
4113 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth--;
4114 mc->mc_txn->mt_dbs[mc->mc_dbi].md_branch_pages--;
4116 DPUTS("root page doesn't need rebalancing");
4120 /* The parent (branch page) must have at least 2 pointers,
4121 * otherwise the tree is invalid.
4123 ptop = mc->mc_top-1;
4124 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
4126 /* Leaf page fill factor is below the threshold.
4127 * Try to move keys from left or right neighbor, or
4128 * merge with a neighbor page.
4133 mdb_cursor_copy(mc, &mn);
4134 mn.mc_xcursor = NULL;
4136 if (mc->mc_ki[ptop] == 0) {
4137 /* We're the leftmost leaf in our parent.
4139 DPUTS("reading right neighbor");
4141 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
4142 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
4144 mn.mc_ki[mn.mc_top] = 0;
4145 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
4147 /* There is at least one neighbor to the left.
4149 DPUTS("reading left neighbor");
4151 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
4152 if ((rc = mdb_get_page(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
4154 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
4155 mc->mc_ki[mc->mc_top] = 0;
4158 DPRINTF("found neighbor page %lu (%u keys, %.1f%% full)",
4159 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);
4161 /* If the neighbor page is above threshold and has at least two
4162 * keys, move one key from it.
4164 * Otherwise we should try to merge them.
4166 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
4167 return mdb_move_node(&mn, mc);
4168 else { /* FIXME: if (has_enough_room()) */
4169 if (mc->mc_ki[ptop] == 0)
4170 return mdb_merge(&mn, mc);
4172 return mdb_merge(mc, &mn);
4177 mdb_del0(MDB_cursor *mc, MDB_node *leaf)
4181 /* add overflow pages to free list */
4182 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4186 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4187 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4188 for (i=0; i<ovpages; i++) {
4189 DPRINTF("freed ov page %lu", pg);
4190 mdb_midl_append(mc->mc_txn->mt_free_pgs, pg);
4194 mdb_del_node(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad);
4195 mc->mc_txn->mt_dbs[mc->mc_dbi].md_entries--;
4196 rc = mdb_rebalance(mc);
4197 if (rc != MDB_SUCCESS)
4198 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4204 mdb_del(MDB_txn *txn, MDB_dbi dbi,
4205 MDB_val *key, MDB_val *data)
4210 MDB_val rdata, *xdata;
4214 assert(key != NULL);
4216 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
4218 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4221 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
4225 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4232 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4233 mc.mc_xcursor = &mx;
4234 mdb_xcursor_init0(&mc);
4236 mc.mc_xcursor = NULL;
4248 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
4250 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
4254 /* Split page <mc->top>, and insert <key,(data|newpgno)> in either left or
4255 * right sibling, at index <mc->ki> (as if unsplit). Updates mc->top and
4256 * mc->ki with the actual values after split, ie if mc->top and mc->ki
4257 * refer to a node in the new right sibling page.
4260 mdb_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno)
4263 int rc = MDB_SUCCESS, ins_new = 0;
4266 unsigned int i, j, split_indx, nkeys, pmax;
4268 MDB_val sepkey, rkey, rdata;
4270 MDB_page *mp, *rp, *pp;
4275 mp = mc->mc_pg[mc->mc_top];
4276 newindx = mc->mc_ki[mc->mc_top];
4278 DPRINTF("-----> splitting %s page %lu and adding [%s] at index %i",
4279 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
4280 DKEY(newkey), mc->mc_ki[mc->mc_top]);
4282 if (mc->mc_snum < 2) {
4283 if ((pp = mdb_new_page(mc, P_BRANCH, 1)) == NULL)
4285 /* shift current top to make room for new parent */
4286 mc->mc_pg[1] = mc->mc_pg[0];
4287 mc->mc_ki[1] = mc->mc_ki[0];
4290 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = pp->mp_pgno;
4291 DPRINTF("root split! new root = %lu", pp->mp_pgno);
4292 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth++;
4294 /* Add left (implicit) pointer. */
4295 if ((rc = mdb_add_node(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
4296 /* undo the pre-push */
4297 mc->mc_pg[0] = mc->mc_pg[1];
4298 mc->mc_ki[0] = mc->mc_ki[1];
4299 mc->mc_txn->mt_dbs[mc->mc_dbi].md_root = mp->mp_pgno;
4300 mc->mc_txn->mt_dbs[mc->mc_dbi].md_depth--;
4307 ptop = mc->mc_top-1;
4308 DPRINTF("parent branch page is %lu", mc->mc_pg[ptop]->mp_pgno);
4311 /* Create a right sibling. */
4312 if ((rp = mdb_new_page(mc, mp->mp_flags, 1)) == NULL)
4314 mdb_cursor_copy(mc, &mn);
4315 mn.mc_pg[mn.mc_top] = rp;
4316 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
4317 DPRINTF("new right sibling: page %lu", rp->mp_pgno);
4319 nkeys = NUMKEYS(mp);
4320 split_indx = nkeys / 2 + 1;
4325 unsigned int lsize, rsize, ksize;
4326 /* Move half of the keys to the right sibling */
4328 x = mc->mc_ki[mc->mc_top] - split_indx;
4329 ksize = mc->mc_txn->mt_dbs[mc->mc_dbi].md_pad;
4330 split = LEAF2KEY(mp, split_indx, ksize);
4331 rsize = (nkeys - split_indx) * ksize;
4332 lsize = (nkeys - split_indx) * sizeof(indx_t);
4333 mp->mp_lower -= lsize;
4334 rp->mp_lower += lsize;
4335 mp->mp_upper += rsize - lsize;
4336 rp->mp_upper -= rsize - lsize;
4337 sepkey.mv_size = ksize;
4338 if (newindx == split_indx) {
4339 sepkey.mv_data = newkey->mv_data;
4341 sepkey.mv_data = split;
4344 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
4345 memcpy(rp->mp_ptrs, split, rsize);
4346 sepkey.mv_data = rp->mp_ptrs;
4347 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
4348 memcpy(ins, newkey->mv_data, ksize);
4349 mp->mp_lower += sizeof(indx_t);
4350 mp->mp_upper -= ksize - sizeof(indx_t);
4353 memcpy(rp->mp_ptrs, split, x * ksize);
4354 ins = LEAF2KEY(rp, x, ksize);
4355 memcpy(ins, newkey->mv_data, ksize);
4356 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
4357 rp->mp_lower += sizeof(indx_t);
4358 rp->mp_upper -= ksize - sizeof(indx_t);
4359 mc->mc_ki[mc->mc_top] = x;
4360 mc->mc_pg[mc->mc_top] = rp;
4365 /* For leaf pages, check the split point based on what
4366 * fits where, since otherwise add_node can fail.
4369 unsigned int psize, nsize;
4370 /* Maximum free space in an empty page */
4371 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
4372 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
4373 if (newindx < split_indx) {
4375 for (i=0; i<split_indx; i++) {
4376 node = NODEPTR(mp, i);
4377 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
4378 if (F_ISSET(node->mn_flags, F_BIGDATA))
4379 psize += sizeof(pgno_t);
4381 psize += NODEDSZ(node);
4390 for (i=nkeys-1; i>=split_indx; i--) {
4391 node = NODEPTR(mp, i);
4392 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
4393 if (F_ISSET(node->mn_flags, F_BIGDATA))
4394 psize += sizeof(pgno_t);
4396 psize += NODEDSZ(node);
4406 /* First find the separating key between the split pages.
4408 if (newindx == split_indx) {
4409 sepkey.mv_size = newkey->mv_size;
4410 sepkey.mv_data = newkey->mv_data;
4412 node = NODEPTR(mp, split_indx);
4413 sepkey.mv_size = node->mn_ksize;
4414 sepkey.mv_data = NODEKEY(node);
4418 DPRINTF("separator is [%s]", DKEY(&sepkey));
4420 /* Copy separator key to the parent.
4422 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
4425 rc = mdb_split(&mn, &sepkey, NULL, rp->mp_pgno);
4427 /* Right page might now have changed parent.
4428 * Check if left page also changed parent.
4430 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
4431 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
4432 mc->mc_pg[ptop] = mn.mc_pg[ptop];
4433 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
4437 rc = mdb_add_node(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
4443 if (rc != MDB_SUCCESS) {
4447 /* Move half of the keys to the right sibling. */
4449 /* grab a page to hold a temporary copy */
4450 if (mc->mc_txn->mt_env->me_dpages) {
4451 copy = mc->mc_txn->mt_env->me_dpages;
4452 mc->mc_txn->mt_env->me_dpages = copy->mp_next;
4454 if ((copy = malloc(mc->mc_txn->mt_env->me_psize)) == NULL)
4458 copy->mp_pgno = mp->mp_pgno;
4459 copy->mp_flags = mp->mp_flags;
4460 copy->mp_lower = PAGEHDRSZ;
4461 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
4462 mc->mc_pg[mc->mc_top] = copy;
4463 for (i = j = 0; i <= nkeys; j++) {
4464 if (i == split_indx) {
4465 /* Insert in right sibling. */
4466 /* Reset insert index for right sibling. */
4467 j = (i == newindx && ins_new);
4468 mc->mc_pg[mc->mc_top] = rp;
4471 if (i == newindx && !ins_new) {
4472 /* Insert the original entry that caused the split. */
4473 rkey.mv_data = newkey->mv_data;
4474 rkey.mv_size = newkey->mv_size;
4476 rdata.mv_data = newdata->mv_data;
4477 rdata.mv_size = newdata->mv_size;
4484 /* Update page and index for the new key. */
4485 mc->mc_ki[mc->mc_top] = j;
4486 } else if (i == nkeys) {
4489 node = NODEPTR(mp, i);
4490 rkey.mv_data = NODEKEY(node);
4491 rkey.mv_size = node->mn_ksize;
4493 rdata.mv_data = NODEDATA(node);
4494 rdata.mv_size = NODEDSZ(node);
4496 pgno = NODEPGNO(node);
4497 flags = node->mn_flags;
4502 if (!IS_LEAF(mp) && j == 0) {
4503 /* First branch index doesn't need key data. */
4507 rc = mdb_add_node(mc, j, &rkey, &rdata, pgno, flags);
4510 /* reset back to original page */
4511 if (newindx < split_indx)
4512 mc->mc_pg[mc->mc_top] = mp;
4514 nkeys = NUMKEYS(copy);
4515 for (i=0; i<nkeys; i++)
4516 mp->mp_ptrs[i] = copy->mp_ptrs[i];
4517 mp->mp_lower = copy->mp_lower;
4518 mp->mp_upper = copy->mp_upper;
4519 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
4520 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
4522 /* return tmp page to freelist */
4523 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
4524 mc->mc_txn->mt_env->me_dpages = copy;
4529 mdb_put(MDB_txn *txn, MDB_dbi dbi,
4530 MDB_val *key, MDB_val *data, unsigned int flags)
4535 assert(key != NULL);
4536 assert(data != NULL);
4538 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4541 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
4545 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4549 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA)) != flags)
4556 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4557 mc.mc_xcursor = &mx;
4558 mdb_xcursor_init0(&mc);
4560 mc.mc_xcursor = NULL;
4562 return mdb_cursor_put(&mc, key, data, flags);
4566 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
4568 /** Only a subset of the @ref mdb_env flags can be changed
4569 * at runtime. Changing other flags requires closing the environment
4570 * and re-opening it with the new flags.
4572 #define CHANGEABLE (MDB_NOSYNC)
4573 if ((flag & CHANGEABLE) != flag)
4576 env->me_flags |= flag;
4578 env->me_flags &= ~flag;
4583 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
4588 *arg = env->me_flags;
4593 mdb_env_get_path(MDB_env *env, const char **arg)
4598 *arg = env->me_path;
4603 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
4605 arg->ms_psize = env->me_psize;
4606 arg->ms_depth = db->md_depth;
4607 arg->ms_branch_pages = db->md_branch_pages;
4608 arg->ms_leaf_pages = db->md_leaf_pages;
4609 arg->ms_overflow_pages = db->md_overflow_pages;
4610 arg->ms_entries = db->md_entries;
4615 mdb_env_stat(MDB_env *env, MDB_stat *arg)
4619 if (env == NULL || arg == NULL)
4622 mdb_env_read_meta(env, &toggle);
4624 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
4628 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
4630 if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEKEY)
4631 txn->mt_dbxs[dbi].md_cmp = memnrcmp;
4632 else if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERKEY)
4633 txn->mt_dbxs[dbi].md_cmp = cintcmp;
4635 txn->mt_dbxs[dbi].md_cmp = memncmp;
4637 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4638 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP) {
4639 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED)
4640 txn->mt_dbxs[dbi].md_dcmp = intcmp;
4642 txn->mt_dbxs[dbi].md_dcmp = cintcmp;
4643 } else if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEDUP) {
4644 txn->mt_dbxs[dbi].md_dcmp = memnrcmp;
4646 txn->mt_dbxs[dbi].md_dcmp = memncmp;
4649 txn->mt_dbxs[dbi].md_dcmp = NULL;
4653 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
4660 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
4661 mdb_default_cmp(txn, FREE_DBI);
4667 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
4668 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
4669 mdb_default_cmp(txn, MAIN_DBI);
4673 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
4674 mdb_default_cmp(txn, MAIN_DBI);
4677 /* Is the DB already open? */
4679 for (i=2; i<txn->mt_numdbs; i++) {
4680 if (len == txn->mt_dbxs[i].md_name.mv_size &&
4681 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
4687 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
4690 /* Find the DB info */
4692 key.mv_data = (void *)name;
4693 rc = mdb_get(txn, MAIN_DBI, &key, &data);
4695 /* Create if requested */
4696 if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
4699 data.mv_size = sizeof(MDB_db);
4700 data.mv_data = &dummy;
4701 memset(&dummy, 0, sizeof(dummy));
4702 dummy.md_root = P_INVALID;
4703 dummy.md_flags = flags & 0xffff;
4705 mc.mc_dbi = MAIN_DBI;
4707 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
4711 /* OK, got info, add to table */
4712 if (rc == MDB_SUCCESS) {
4713 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
4714 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
4715 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
4716 txn->mt_dbxs[txn->mt_numdbs].md_parent = MAIN_DBI;
4717 txn->mt_dbxs[txn->mt_numdbs].md_dirty = dirty;
4718 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
4719 *dbi = txn->mt_numdbs;
4720 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
4721 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
4722 mdb_default_cmp(txn, txn->mt_numdbs);
4729 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
4731 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
4734 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
4737 void mdb_close(MDB_txn *txn, MDB_dbi dbi)
4740 if (dbi <= MAIN_DBI || dbi >= txn->mt_numdbs)
4742 ptr = txn->mt_dbxs[dbi].md_name.mv_data;
4743 txn->mt_dbxs[dbi].md_name.mv_data = NULL;
4744 txn->mt_dbxs[dbi].md_name.mv_size = 0;
4748 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
4750 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4753 txn->mt_dbxs[dbi].md_cmp = cmp;
4757 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
4759 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4762 txn->mt_dbxs[dbi].md_dcmp = cmp;
4766 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
4768 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4771 txn->mt_dbxs[dbi].md_rel = rel;