2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2012 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
35 #include <sys/types.h>
37 #include <sys/param.h>
43 #ifdef HAVE_SYS_FILE_H
60 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
61 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
67 #include <semaphore.h>
72 #include <valgrind/memcheck.h>
73 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
74 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
75 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
76 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
77 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
79 #define VGMEMP_CREATE(h,r,z)
80 #define VGMEMP_ALLOC(h,a,s)
81 #define VGMEMP_FREE(h,a)
82 #define VGMEMP_DESTROY(h)
83 #define VGMEMP_DEFINED(a,s)
87 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
88 /* Solaris just defines one or the other */
89 # define LITTLE_ENDIAN 1234
90 # define BIG_ENDIAN 4321
91 # ifdef _LITTLE_ENDIAN
92 # define BYTE_ORDER LITTLE_ENDIAN
94 # define BYTE_ORDER BIG_ENDIAN
97 # define BYTE_ORDER __BYTE_ORDER
101 #ifndef LITTLE_ENDIAN
102 #define LITTLE_ENDIAN __LITTLE_ENDIAN
105 #define BIG_ENDIAN __BIG_ENDIAN
108 #if defined(__i386) || defined(__x86_64)
109 #define MISALIGNED_OK 1
115 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
116 # error "Unknown or unsupported endianness (BYTE_ORDER)"
117 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
118 # error "Two's complement, reasonably sized integer types, please"
121 /** @defgroup internal MDB Internals
124 /** @defgroup compat Windows Compatibility Macros
125 * A bunch of macros to minimize the amount of platform-specific ifdefs
126 * needed throughout the rest of the code. When the features this library
127 * needs are similar enough to POSIX to be hidden in a one-or-two line
128 * replacement, this macro approach is used.
132 #define pthread_t DWORD
133 #define pthread_mutex_t HANDLE
134 #define pthread_key_t DWORD
135 #define pthread_self() GetCurrentThreadId()
136 #define pthread_key_create(x,y) (*(x) = TlsAlloc())
137 #define pthread_key_delete(x) TlsFree(x)
138 #define pthread_getspecific(x) TlsGetValue(x)
139 #define pthread_setspecific(x,y) TlsSetValue(x,y)
140 #define pthread_mutex_unlock(x) ReleaseMutex(x)
141 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
142 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
143 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
144 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
145 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
146 #define getpid() GetCurrentProcessId()
147 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
148 #define ErrCode() GetLastError()
149 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
150 #define close(fd) CloseHandle(fd)
151 #define munmap(ptr,len) UnmapViewOfFile(ptr)
154 #define LOCK_MUTEX_R(env) sem_wait((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) sem_wait((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
158 #define MDB_FDATASYNC(fd) fsync(fd)
161 #define MDB_FDATASYNC(fd) fsync(fd)
163 /** Lock the reader mutex.
165 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
166 /** Unlock the reader mutex.
168 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
170 /** Lock the writer mutex.
171 * Only a single write transaction is allowed at a time. Other writers
172 * will block waiting for this mutex.
174 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
175 /** Unlock the writer mutex.
177 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
178 #endif /* __APPLE__ */
180 /** Get the error code for the last failed system function.
182 #define ErrCode() errno
184 /** An abstraction for a file handle.
185 * On POSIX systems file handles are small integers. On Windows
186 * they're opaque pointers.
190 /** A value for an invalid file handle.
191 * Mainly used to initialize file variables and signify that they are
194 #define INVALID_HANDLE_VALUE (-1)
196 /** Get the size of a memory page for the system.
197 * This is the basic size that the platform's memory manager uses, and is
198 * fundamental to the use of memory-mapped files.
200 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
203 #if defined(_WIN32) || defined(__APPLE__)
206 #define MNAME_LEN (sizeof(pthread_mutex_t))
212 /** A flag for opening a file and requesting synchronous data writes.
213 * This is only used when writing a meta page. It's not strictly needed;
214 * we could just do a normal write and then immediately perform a flush.
215 * But if this flag is available it saves us an extra system call.
217 * @note If O_DSYNC is undefined but exists in /usr/include,
218 * preferably set some compiler flag to get the definition.
219 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
222 # define MDB_DSYNC O_DSYNC
226 /** Function for flushing the data of a file. Define this to fsync
227 * if fdatasync() is not supported.
229 #ifndef MDB_FDATASYNC
230 # define MDB_FDATASYNC fdatasync
233 /** A page number in the database.
234 * Note that 64 bit page numbers are overkill, since pages themselves
235 * already represent 12-13 bits of addressable memory, and the OS will
236 * always limit applications to a maximum of 63 bits of address space.
238 * @note In the #MDB_node structure, we only store 48 bits of this value,
239 * which thus limits us to only 60 bits of addressable data.
241 typedef MDB_ID pgno_t;
243 /** A transaction ID.
244 * See struct MDB_txn.mt_txnid for details.
246 typedef MDB_ID txnid_t;
248 /** @defgroup debug Debug Macros
252 /** Enable debug output.
253 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
254 * read from and written to the database (used for free space management).
259 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
260 # define DPRINTF (void) /* Vararg macros may be unsupported */
262 static int mdb_debug;
263 static txnid_t mdb_debug_start;
265 /** Print a debug message with printf formatting. */
266 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
267 ((void) ((mdb_debug) && \
268 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
270 # define DPRINTF(fmt, ...) ((void) 0)
272 /** Print a debug string.
273 * The string is printed literally, with no format processing.
275 #define DPUTS(arg) DPRINTF("%s", arg)
278 /** A default memory page size.
279 * The actual size is platform-dependent, but we use this for
280 * boot-strapping. We probably should not be using this any more.
281 * The #GET_PAGESIZE() macro is used to get the actual size.
283 * Note that we don't currently support Huge pages. On Linux,
284 * regular data files cannot use Huge pages, and in general
285 * Huge pages aren't actually pageable. We rely on the OS
286 * demand-pager to read our data and page it out when memory
287 * pressure from other processes is high. So until OSs have
288 * actual paging support for Huge pages, they're not viable.
290 #define MDB_PAGESIZE 4096
292 /** The minimum number of keys required in a database page.
293 * Setting this to a larger value will place a smaller bound on the
294 * maximum size of a data item. Data items larger than this size will
295 * be pushed into overflow pages instead of being stored directly in
296 * the B-tree node. This value used to default to 4. With a page size
297 * of 4096 bytes that meant that any item larger than 1024 bytes would
298 * go into an overflow page. That also meant that on average 2-3KB of
299 * each overflow page was wasted space. The value cannot be lower than
300 * 2 because then there would no longer be a tree structure. With this
301 * value, items larger than 2KB will go into overflow pages, and on
302 * average only 1KB will be wasted.
304 #define MDB_MINKEYS 2
306 /** A stamp that identifies a file as an MDB file.
307 * There's nothing special about this value other than that it is easily
308 * recognizable, and it will reflect any byte order mismatches.
310 #define MDB_MAGIC 0xBEEFC0DE
312 /** The version number for a database's file format. */
313 #define MDB_VERSION 1
315 /** The maximum size of a key in the database.
316 * While data items have essentially unbounded size, we require that
317 * keys all fit onto a regular page. This limit could be raised a bit
318 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
320 #define MAXKEYSIZE 511
325 * This is used for printing a hex dump of a key's contents.
327 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
328 /** Display a key in hex.
330 * Invoke a function to display a key in hex.
332 #define DKEY(x) mdb_dkey(x, kbuf)
334 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
338 /** An invalid page number.
339 * Mainly used to denote an empty tree.
341 #define P_INVALID (~0UL)
343 /** Test if a flag \b f is set in a flag word \b w. */
344 #define F_ISSET(w, f) (((w) & (f)) == (f))
346 /** Used for offsets within a single page.
347 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
350 typedef uint16_t indx_t;
352 /** Default size of memory map.
353 * This is certainly too small for any actual applications. Apps should always set
354 * the size explicitly using #mdb_env_set_mapsize().
356 #define DEFAULT_MAPSIZE 1048576
358 /** @defgroup readers Reader Lock Table
359 * Readers don't acquire any locks for their data access. Instead, they
360 * simply record their transaction ID in the reader table. The reader
361 * mutex is needed just to find an empty slot in the reader table. The
362 * slot's address is saved in thread-specific data so that subsequent read
363 * transactions started by the same thread need no further locking to proceed.
365 * Since the database uses multi-version concurrency control, readers don't
366 * actually need any locking. This table is used to keep track of which
367 * readers are using data from which old transactions, so that we'll know
368 * when a particular old transaction is no longer in use. Old transactions
369 * that have discarded any data pages can then have those pages reclaimed
370 * for use by a later write transaction.
372 * The lock table is constructed such that reader slots are aligned with the
373 * processor's cache line size. Any slot is only ever used by one thread.
374 * This alignment guarantees that there will be no contention or cache
375 * thrashing as threads update their own slot info, and also eliminates
376 * any need for locking when accessing a slot.
378 * A writer thread will scan every slot in the table to determine the oldest
379 * outstanding reader transaction. Any freed pages older than this will be
380 * reclaimed by the writer. The writer doesn't use any locks when scanning
381 * this table. This means that there's no guarantee that the writer will
382 * see the most up-to-date reader info, but that's not required for correct
383 * operation - all we need is to know the upper bound on the oldest reader,
384 * we don't care at all about the newest reader. So the only consequence of
385 * reading stale information here is that old pages might hang around a
386 * while longer before being reclaimed. That's actually good anyway, because
387 * the longer we delay reclaiming old pages, the more likely it is that a
388 * string of contiguous pages can be found after coalescing old pages from
389 * many old transactions together.
391 * @todo We don't actually do such coalescing yet, we grab pages from one
392 * old transaction at a time.
395 /** Number of slots in the reader table.
396 * This value was chosen somewhat arbitrarily. 126 readers plus a
397 * couple mutexes fit exactly into 8KB on my development machine.
398 * Applications should set the table size using #mdb_env_set_maxreaders().
400 #define DEFAULT_READERS 126
402 /** The size of a CPU cache line in bytes. We want our lock structures
403 * aligned to this size to avoid false cache line sharing in the
405 * This value works for most CPUs. For Itanium this should be 128.
411 /** The information we store in a single slot of the reader table.
412 * In addition to a transaction ID, we also record the process and
413 * thread ID that owns a slot, so that we can detect stale information,
414 * e.g. threads or processes that went away without cleaning up.
415 * @note We currently don't check for stale records. We simply re-init
416 * the table when we know that we're the only process opening the
419 typedef struct MDB_rxbody {
420 /** The current Transaction ID when this transaction began.
421 * Multiple readers that start at the same time will probably have the
422 * same ID here. Again, it's not important to exclude them from
423 * anything; all we need to know is which version of the DB they
424 * started from so we can avoid overwriting any data used in that
425 * particular version.
428 /** The process ID of the process owning this reader txn. */
430 /** The thread ID of the thread owning this txn. */
434 /** The actual reader record, with cacheline padding. */
435 typedef struct MDB_reader {
438 /** shorthand for mrb_txnid */
439 #define mr_txnid mru.mrx.mrb_txnid
440 #define mr_pid mru.mrx.mrb_pid
441 #define mr_tid mru.mrx.mrb_tid
442 /** cache line alignment */
443 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
447 /** The header for the reader table.
448 * The table resides in a memory-mapped file. (This is a different file
449 * than is used for the main database.)
451 * For POSIX the actual mutexes reside in the shared memory of this
452 * mapped file. On Windows, mutexes are named objects allocated by the
453 * kernel; we store the mutex names in this mapped file so that other
454 * processes can grab them. This same approach is also used on
455 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
456 * process-shared POSIX mutexes. For these cases where a named object
457 * is used, the object name is derived from a 64 bit FNV hash of the
458 * environment pathname. As such, naming collisions are extremely
459 * unlikely. If a collision occurs, the results are unpredictable.
461 typedef struct MDB_txbody {
462 /** Stamp identifying this as an MDB file. It must be set
465 /** Version number of this lock file. Must be set to #MDB_VERSION. */
466 uint32_t mtb_version;
467 #if defined(_WIN32) || defined(__APPLE__)
468 char mtb_rmname[MNAME_LEN];
470 /** Mutex protecting access to this table.
471 * This is the reader lock that #LOCK_MUTEX_R acquires.
473 pthread_mutex_t mtb_mutex;
475 /** The ID of the last transaction committed to the database.
476 * This is recorded here only for convenience; the value can always
477 * be determined by reading the main database meta pages.
480 /** The number of slots that have been used in the reader table.
481 * This always records the maximum count, it is not decremented
482 * when readers release their slots.
484 unsigned mtb_numreaders;
487 /** The actual reader table definition. */
488 typedef struct MDB_txninfo {
491 #define mti_magic mt1.mtb.mtb_magic
492 #define mti_version mt1.mtb.mtb_version
493 #define mti_mutex mt1.mtb.mtb_mutex
494 #define mti_rmname mt1.mtb.mtb_rmname
495 #define mti_txnid mt1.mtb.mtb_txnid
496 #define mti_numreaders mt1.mtb.mtb_numreaders
497 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
500 #if defined(_WIN32) || defined(__APPLE__)
501 char mt2_wmname[MNAME_LEN];
502 #define mti_wmname mt2.mt2_wmname
504 pthread_mutex_t mt2_wmutex;
505 #define mti_wmutex mt2.mt2_wmutex
507 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
509 MDB_reader mti_readers[1];
513 /** Common header for all page types.
514 * Overflow records occupy a number of contiguous pages with no
515 * headers on any page after the first.
517 typedef struct MDB_page {
518 #define mp_pgno mp_p.p_pgno
519 #define mp_next mp_p.p_next
521 pgno_t p_pgno; /**< page number */
522 void * p_next; /**< for in-memory list of freed structs */
525 /** @defgroup mdb_page Page Flags
527 * Flags for the page headers.
530 #define P_BRANCH 0x01 /**< branch page */
531 #define P_LEAF 0x02 /**< leaf page */
532 #define P_OVERFLOW 0x04 /**< overflow page */
533 #define P_META 0x08 /**< meta page */
534 #define P_DIRTY 0x10 /**< dirty page */
535 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
536 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
538 uint16_t mp_flags; /**< @ref mdb_page */
539 #define mp_lower mp_pb.pb.pb_lower
540 #define mp_upper mp_pb.pb.pb_upper
541 #define mp_pages mp_pb.pb_pages
544 indx_t pb_lower; /**< lower bound of free space */
545 indx_t pb_upper; /**< upper bound of free space */
547 uint32_t pb_pages; /**< number of overflow pages */
549 indx_t mp_ptrs[1]; /**< dynamic size */
552 /** Size of the page header, excluding dynamic data at the end */
553 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
555 /** Address of first usable data byte in a page, after the header */
556 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
558 /** Number of nodes on a page */
559 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
561 /** The amount of space remaining in the page */
562 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
564 /** The percentage of space used in the page, in tenths of a percent. */
565 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
566 ((env)->me_psize - PAGEHDRSZ))
567 /** The minimum page fill factor, in tenths of a percent.
568 * Pages emptier than this are candidates for merging.
570 #define FILL_THRESHOLD 250
572 /** Test if a page is a leaf page */
573 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
574 /** Test if a page is a LEAF2 page */
575 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
576 /** Test if a page is a branch page */
577 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
578 /** Test if a page is an overflow page */
579 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
580 /** Test if a page is a sub page */
581 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
583 /** The number of overflow pages needed to store the given size. */
584 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
586 /** Header for a single key/data pair within a page.
587 * We guarantee 2-byte alignment for nodes.
589 typedef struct MDB_node {
590 /** lo and hi are used for data size on leaf nodes and for
591 * child pgno on branch nodes. On 64 bit platforms, flags
592 * is also used for pgno. (Branch nodes have no flags).
593 * They are in host byte order in case that lets some
594 * accesses be optimized into a 32-bit word access.
596 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
597 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
598 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
599 /** @defgroup mdb_node Node Flags
601 * Flags for node headers.
604 #define F_BIGDATA 0x01 /**< data put on overflow page */
605 #define F_SUBDATA 0x02 /**< data is a sub-database */
606 #define F_DUPDATA 0x04 /**< data has duplicates */
608 /** valid flags for #mdb_node_add() */
609 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
612 unsigned short mn_flags; /**< @ref mdb_node */
613 unsigned short mn_ksize; /**< key size */
614 char mn_data[1]; /**< key and data are appended here */
617 /** Size of the node header, excluding dynamic data at the end */
618 #define NODESIZE offsetof(MDB_node, mn_data)
620 /** Bit position of top word in page number, for shifting mn_flags */
621 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
623 /** Size of a node in a branch page with a given key.
624 * This is just the node header plus the key, there is no data.
626 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
628 /** Size of a node in a leaf page with a given key and data.
629 * This is node header plus key plus data size.
631 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
633 /** Address of node \b i in page \b p */
634 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
636 /** Address of the key for the node */
637 #define NODEKEY(node) (void *)((node)->mn_data)
639 /** Address of the data for a node */
640 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
642 /** Get the page number pointed to by a branch node */
643 #define NODEPGNO(node) \
644 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
645 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
646 /** Set the page number in a branch node */
647 #define SETPGNO(node,pgno) do { \
648 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
649 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
651 /** Get the size of the data in a leaf node */
652 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
653 /** Set the size of the data for a leaf node */
654 #define SETDSZ(node,size) do { \
655 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
656 /** The size of a key in a node */
657 #define NODEKSZ(node) ((node)->mn_ksize)
659 /** Copy a page number from src to dst */
661 #define COPY_PGNO(dst,src) dst = src
663 #if SIZE_MAX > 4294967295UL
664 #define COPY_PGNO(dst,src) do { \
665 unsigned short *s, *d; \
666 s = (unsigned short *)&(src); \
667 d = (unsigned short *)&(dst); \
674 #define COPY_PGNO(dst,src) do { \
675 unsigned short *s, *d; \
676 s = (unsigned short *)&(src); \
677 d = (unsigned short *)&(dst); \
683 /** The address of a key in a LEAF2 page.
684 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
685 * There are no node headers, keys are stored contiguously.
687 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
689 /** Set the \b node's key into \b key, if requested. */
690 #define MDB_SET_KEY(node, key) { if ((key) != NULL) { \
691 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
693 /** Information about a single database in the environment. */
694 typedef struct MDB_db {
695 uint32_t md_pad; /**< also ksize for LEAF2 pages */
696 uint16_t md_flags; /**< @ref mdb_open */
697 uint16_t md_depth; /**< depth of this tree */
698 pgno_t md_branch_pages; /**< number of internal pages */
699 pgno_t md_leaf_pages; /**< number of leaf pages */
700 pgno_t md_overflow_pages; /**< number of overflow pages */
701 size_t md_entries; /**< number of data items */
702 pgno_t md_root; /**< the root page of this tree */
705 /** Handle for the DB used to track free pages. */
707 /** Handle for the default DB. */
710 /** Meta page content. */
711 typedef struct MDB_meta {
712 /** Stamp identifying this as an MDB file. It must be set
715 /** Version number of this lock file. Must be set to #MDB_VERSION. */
717 void *mm_address; /**< address for fixed mapping */
718 size_t mm_mapsize; /**< size of mmap region */
719 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
720 /** The size of pages used in this DB */
721 #define mm_psize mm_dbs[0].md_pad
722 /** Any persistent environment flags. @ref mdb_env */
723 #define mm_flags mm_dbs[0].md_flags
724 pgno_t mm_last_pg; /**< last used page in file */
725 txnid_t mm_txnid; /**< txnid that committed this page */
728 /** Buffer for a stack-allocated dirty page.
729 * The members define size and alignment, and silence type
730 * aliasing warnings. They are not used directly; that could
731 * mean incorrectly using several union members in parallel.
733 typedef union MDB_pagebuf {
734 char mb_raw[MDB_PAGESIZE];
737 char mm_pad[PAGEHDRSZ];
742 /** Auxiliary DB info.
743 * The information here is mostly static/read-only. There is
744 * only a single copy of this record in the environment.
746 typedef struct MDB_dbx {
747 MDB_val md_name; /**< name of the database */
748 MDB_cmp_func *md_cmp; /**< function for comparing keys */
749 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
750 MDB_rel_func *md_rel; /**< user relocate function */
751 void *md_relctx; /**< user-provided context for md_rel */
754 /** A database transaction.
755 * Every operation requires a transaction handle.
758 MDB_txn *mt_parent; /**< parent of a nested txn */
759 MDB_txn *mt_child; /**< nested txn under this txn */
760 pgno_t mt_next_pgno; /**< next unallocated page */
761 /** The ID of this transaction. IDs are integers incrementing from 1.
762 * Only committed write transactions increment the ID. If a transaction
763 * aborts, the ID may be re-used by the next writer.
766 MDB_env *mt_env; /**< the DB environment */
767 /** The list of pages that became unused during this transaction.
771 MDB_ID2L dirty_list; /**< modified pages */
772 MDB_reader *reader; /**< this thread's slot in the reader table */
774 /** Array of records for each DB known in the environment. */
776 /** Array of MDB_db records for each known DB */
778 /** @defgroup mt_dbflag Transaction DB Flags
782 #define DB_DIRTY 0x01 /**< DB was written in this txn */
783 #define DB_STALE 0x02 /**< DB record is older than txnID */
785 /** Array of cursors for each DB */
786 MDB_cursor **mt_cursors;
787 /** Array of flags for each DB */
788 unsigned char *mt_dbflags;
789 /** Number of DB records in use. This number only ever increments;
790 * we don't decrement it when individual DB handles are closed.
794 /** @defgroup mdb_txn Transaction Flags
798 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
799 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
801 unsigned int mt_flags; /**< @ref mdb_txn */
802 /** Tracks which of the two meta pages was used at the start
803 * of this transaction.
805 unsigned int mt_toggle;
808 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
809 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
810 * raise this on a 64 bit machine.
812 #define CURSOR_STACK 32
816 /** Cursors are used for all DB operations */
818 /** Next cursor on this DB in this txn */
820 /** Original cursor if this is a shadow */
822 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
823 struct MDB_xcursor *mc_xcursor;
824 /** The transaction that owns this cursor */
826 /** The database handle this cursor operates on */
828 /** The database record for this cursor */
830 /** The database auxiliary record for this cursor */
832 /** The @ref mt_dbflag for this database */
833 unsigned char *mc_dbflag;
834 unsigned short mc_snum; /**< number of pushed pages */
835 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
836 /** @defgroup mdb_cursor Cursor Flags
838 * Cursor state flags.
841 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
842 #define C_EOF 0x02 /**< No more data */
843 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
844 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
845 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
846 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
848 unsigned int mc_flags; /**< @ref mdb_cursor */
849 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
850 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
853 /** Context for sorted-dup records.
854 * We could have gone to a fully recursive design, with arbitrarily
855 * deep nesting of sub-databases. But for now we only handle these
856 * levels - main DB, optional sub-DB, sorted-duplicate DB.
858 typedef struct MDB_xcursor {
859 /** A sub-cursor for traversing the Dup DB */
860 MDB_cursor mx_cursor;
861 /** The database record for this Dup DB */
863 /** The auxiliary DB record for this Dup DB */
865 /** The @ref mt_dbflag for this Dup DB */
866 unsigned char mx_dbflag;
869 /** A set of pages freed by an earlier transaction. */
870 typedef struct MDB_oldpages {
871 /** Usually we only read one record from the FREEDB at a time, but
872 * in case we read more, this will chain them together.
874 struct MDB_oldpages *mo_next;
875 /** The ID of the transaction in which these pages were freed. */
877 /** An #MDB_IDL of the pages */
878 pgno_t mo_pages[1]; /* dynamic */
881 /** The database environment. */
883 HANDLE me_fd; /**< The main data file */
884 HANDLE me_lfd; /**< The lock file */
885 HANDLE me_mfd; /**< just for writing the meta pages */
886 /** Failed to update the meta page. Probably an I/O error. */
887 #define MDB_FATAL_ERROR 0x80000000U
888 uint32_t me_flags; /**< @ref mdb_env */
889 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
890 unsigned int me_maxreaders; /**< size of the reader table */
891 MDB_dbi me_numdbs; /**< number of DBs opened */
892 MDB_dbi me_maxdbs; /**< size of the DB table */
893 char *me_path; /**< path to the DB files */
894 char *me_map; /**< the memory map of the data file */
895 MDB_txninfo *me_txns; /**< the memory map of the lock file */
896 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
897 MDB_txn *me_txn; /**< current write transaction */
898 size_t me_mapsize; /**< size of the data memory map */
899 off_t me_size; /**< current file size */
900 pgno_t me_maxpg; /**< me_mapsize / me_psize */
901 txnid_t me_pgfirst; /**< ID of first old page record we used */
902 txnid_t me_pglast; /**< ID of last old page record we used */
903 MDB_dbx *me_dbxs; /**< array of static DB info */
904 uint16_t *me_dbflags; /**< array of DB flags */
905 MDB_oldpages *me_pghead; /**< list of old page records */
906 MDB_oldpages *me_pgfree; /**< list of page records to free */
907 pthread_key_t me_txkey; /**< thread-key for readers */
908 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
909 /** IDL of pages that became unused in a write txn */
911 /** ID2L of pages that were written during a write txn */
912 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
914 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
918 sem_t *me_rmutex; /* Apple doesn't support shared mutexes */
922 /** max number of pages to commit in one writev() call */
923 #define MDB_COMMIT_PAGES 64
924 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
925 #undef MDB_COMMIT_PAGES
926 #define MDB_COMMIT_PAGES IOV_MAX
929 static MDB_page *mdb_page_alloc(MDB_cursor *mc, int num);
930 static MDB_page *mdb_page_new(MDB_cursor *mc, uint32_t flags, int num);
931 static int mdb_page_touch(MDB_cursor *mc);
933 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
934 static int mdb_page_search_root(MDB_cursor *mc,
935 MDB_val *key, int modify);
936 #define MDB_PS_MODIFY 1
937 #define MDB_PS_ROOTONLY 2
938 static int mdb_page_search(MDB_cursor *mc,
939 MDB_val *key, int flags);
940 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
942 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
943 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
944 pgno_t newpgno, unsigned int nflags);
946 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
947 static int mdb_env_pick_meta(const MDB_env *env);
948 static int mdb_env_write_meta(MDB_txn *txn);
950 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
951 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
952 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
953 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
954 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
955 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
956 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
957 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
958 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
960 static int mdb_rebalance(MDB_cursor *mc);
961 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
963 static void mdb_cursor_pop(MDB_cursor *mc);
964 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
966 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
967 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
968 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
969 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
970 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
972 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
973 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
975 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
976 static void mdb_xcursor_init0(MDB_cursor *mc);
977 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
979 static int mdb_drop0(MDB_cursor *mc, int subs);
980 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
983 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
987 static SECURITY_DESCRIPTOR mdb_null_sd;
988 static SECURITY_ATTRIBUTES mdb_all_sa;
989 static int mdb_sec_inited;
992 /** Return the library version info. */
994 mdb_version(int *major, int *minor, int *patch)
996 if (major) *major = MDB_VERSION_MAJOR;
997 if (minor) *minor = MDB_VERSION_MINOR;
998 if (patch) *patch = MDB_VERSION_PATCH;
999 return MDB_VERSION_STRING;
1002 /** Table of descriptions for MDB @ref errors */
1003 static char *const mdb_errstr[] = {
1004 "MDB_KEYEXIST: Key/data pair already exists",
1005 "MDB_NOTFOUND: No matching key/data pair found",
1006 "MDB_PAGE_NOTFOUND: Requested page not found",
1007 "MDB_CORRUPTED: Located page was wrong type",
1008 "MDB_PANIC: Update of meta page failed",
1009 "MDB_VERSION_MISMATCH: Database environment version mismatch"
1013 mdb_strerror(int err)
1016 return ("Successful return: 0");
1018 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
1019 return mdb_errstr[err - MDB_KEYEXIST];
1021 return strerror(err);
1025 /** Display a key in hexadecimal and return the address of the result.
1026 * @param[in] key the key to display
1027 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1028 * @return The key in hexadecimal form.
1031 mdb_dkey(MDB_val *key, char *buf)
1034 unsigned char *c = key->mv_data;
1036 if (key->mv_size > MAXKEYSIZE)
1037 return "MAXKEYSIZE";
1038 /* may want to make this a dynamic check: if the key is mostly
1039 * printable characters, print it as-is instead of converting to hex.
1043 for (i=0; i<key->mv_size; i++)
1044 ptr += sprintf(ptr, "%02x", *c++);
1046 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1051 /** Display all the keys in the page. */
1053 mdb_page_keys(MDB_page *mp)
1056 unsigned int i, nkeys;
1060 nkeys = NUMKEYS(mp);
1061 fprintf(stderr, "numkeys %d\n", nkeys);
1062 for (i=0; i<nkeys; i++) {
1063 node = NODEPTR(mp, i);
1064 key.mv_size = node->mn_ksize;
1065 key.mv_data = node->mn_data;
1066 fprintf(stderr, "key %d: %s\n", i, DKEY(&key));
1071 mdb_cursor_chk(MDB_cursor *mc)
1077 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1078 for (i=0; i<mc->mc_top; i++) {
1080 node = NODEPTR(mp, mc->mc_ki[i]);
1081 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1084 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1090 /** Count all the pages in each DB and in the freelist
1091 * and make sure it matches the actual number of pages
1094 static void mdb_audit(MDB_txn *txn)
1098 MDB_ID freecount, count;
1103 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1104 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1105 freecount += *(MDB_ID *)data.mv_data;
1106 freecount += txn->mt_dbs[0].md_branch_pages + txn->mt_dbs[0].md_leaf_pages +
1107 txn->mt_dbs[0].md_overflow_pages;
1110 for (i = 0; i<txn->mt_numdbs; i++) {
1111 count += txn->mt_dbs[i].md_branch_pages +
1112 txn->mt_dbs[i].md_leaf_pages +
1113 txn->mt_dbs[i].md_overflow_pages;
1114 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1116 mdb_cursor_init(&mc, txn, i, &mx);
1117 mdb_page_search(&mc, NULL, 0);
1121 mp = mc.mc_pg[mc.mc_top];
1122 for (j=0; j<NUMKEYS(mp); j++) {
1123 MDB_node *leaf = NODEPTR(mp, j);
1124 if (leaf->mn_flags & F_SUBDATA) {
1126 memcpy(&db, NODEDATA(leaf), sizeof(db));
1127 count += db.md_branch_pages + db.md_leaf_pages +
1128 db.md_overflow_pages;
1132 while (mdb_cursor_sibling(&mc, 1) == 0);
1135 assert(freecount + count + 2 >= txn->mt_next_pgno - 1);
1140 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1142 return txn->mt_dbxs[dbi].md_cmp(a, b);
1146 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1148 if (txn->mt_dbxs[dbi].md_dcmp)
1149 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1151 return EINVAL; /* too bad you can't distinguish this from a valid result */
1154 /** Allocate a single page.
1155 * Re-use old malloc'd pages first, otherwise just malloc.
1158 mdb_page_malloc(MDB_cursor *mc) {
1160 size_t sz = mc->mc_txn->mt_env->me_psize;
1161 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1162 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1163 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1164 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1165 } else if ((ret = malloc(sz)) != NULL) {
1166 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1171 /** Allocate pages for writing.
1172 * If there are free pages available from older transactions, they
1173 * will be re-used first. Otherwise a new page will be allocated.
1174 * @param[in] mc cursor A cursor handle identifying the transaction and
1175 * database for which we are allocating.
1176 * @param[in] num the number of pages to allocate.
1177 * @return Address of the allocated page(s). Requests for multiple pages
1178 * will always be satisfied by a single contiguous chunk of memory.
1181 mdb_page_alloc(MDB_cursor *mc, int num)
1183 MDB_txn *txn = mc->mc_txn;
1185 pgno_t pgno = P_INVALID;
1188 /* The free list won't have any content at all until txn 2 has
1189 * committed. The pages freed by txn 2 will be unreferenced
1190 * after txn 3 commits, and so will be safe to re-use in txn 4.
1192 if (txn->mt_txnid > 3) {
1194 if (!txn->mt_env->me_pghead &&
1195 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1196 /* See if there's anything in the free DB */
1200 txnid_t *kptr, oldest, last;
1202 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1203 if (!txn->mt_env->me_pgfirst) {
1204 mdb_page_search(&m2, NULL, 0);
1205 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1206 kptr = (txnid_t *)NODEKEY(leaf);
1213 last = txn->mt_env->me_pglast + 1;
1215 key.mv_data = &last;
1216 key.mv_size = sizeof(last);
1217 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1220 last = *(txnid_t *)key.mv_data;
1225 oldest = txn->mt_txnid - 1;
1226 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
1227 txnid_t mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
1228 if (mr && mr < oldest)
1233 if (oldest > last) {
1234 /* It's usable, grab it.
1239 if (!txn->mt_env->me_pgfirst) {
1240 mdb_node_read(txn, leaf, &data);
1242 txn->mt_env->me_pglast = last;
1243 if (!txn->mt_env->me_pgfirst)
1244 txn->mt_env->me_pgfirst = last;
1245 idl = (MDB_ID *) data.mv_data;
1246 /* We might have a zero-length IDL due to freelist growth
1247 * during a prior commit
1249 if (!idl[0]) goto again;
1250 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1251 mop->mo_next = txn->mt_env->me_pghead;
1252 mop->mo_txnid = last;
1253 txn->mt_env->me_pghead = mop;
1254 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1259 DPRINTF("IDL read txn %zu root %zu num %zu",
1260 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1261 for (i=0; i<idl[0]; i++) {
1262 DPRINTF("IDL %zu", idl[i+1]);
1269 if (txn->mt_env->me_pghead) {
1270 MDB_oldpages *mop = txn->mt_env->me_pghead;
1272 /* FIXME: For now, always use fresh pages. We
1273 * really ought to search the free list for a
1278 /* peel pages off tail, so we only have to truncate the list */
1279 pgno = MDB_IDL_LAST(mop->mo_pages);
1280 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1282 if (mop->mo_pages[2] > mop->mo_pages[1])
1283 mop->mo_pages[0] = 0;
1287 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1288 txn->mt_env->me_pghead = mop->mo_next;
1289 if (mc->mc_dbi == FREE_DBI) {
1290 mop->mo_next = txn->mt_env->me_pgfree;
1291 txn->mt_env->me_pgfree = mop;
1300 if (pgno == P_INVALID) {
1301 /* DB size is maxed out */
1302 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1303 DPUTS("DB size maxed out");
1307 if (txn->mt_env->me_dpages && num == 1) {
1308 np = txn->mt_env->me_dpages;
1309 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1310 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1311 txn->mt_env->me_dpages = np->mp_next;
1313 size_t sz = txn->mt_env->me_psize * num;
1314 if ((np = malloc(sz)) == NULL)
1316 VGMEMP_ALLOC(txn->mt_env, np, sz);
1318 if (pgno == P_INVALID) {
1319 np->mp_pgno = txn->mt_next_pgno;
1320 txn->mt_next_pgno += num;
1324 mid.mid = np->mp_pgno;
1326 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1331 /** Copy a page: avoid copying unused portions of the page.
1332 * @param[in] dst page to copy into
1333 * @param[in] src page to copy from
1336 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1338 dst->mp_flags = src->mp_flags | P_DIRTY;
1339 dst->mp_pages = src->mp_pages;
1341 if (IS_LEAF2(src)) {
1342 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1344 unsigned int i, nkeys = NUMKEYS(src);
1345 for (i=0; i<nkeys; i++)
1346 dst->mp_ptrs[i] = src->mp_ptrs[i];
1347 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1348 psize - src->mp_upper);
1352 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1353 * @param[in] mc cursor pointing to the page to be touched
1354 * @return 0 on success, non-zero on failure.
1357 mdb_page_touch(MDB_cursor *mc)
1359 MDB_page *mp = mc->mc_pg[mc->mc_top];
1362 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1364 if ((np = mdb_page_alloc(mc, 1)) == NULL)
1366 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1367 assert(mp->mp_pgno != np->mp_pgno);
1368 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1370 /* If page isn't full, just copy the used portion */
1371 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1374 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1376 np->mp_flags |= P_DIRTY;
1381 /* Adjust other cursors pointing to mp */
1382 if (mc->mc_flags & C_SUB) {
1383 MDB_cursor *m2, *m3;
1384 MDB_dbi dbi = mc->mc_dbi-1;
1386 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1387 if (m2 == mc) continue;
1388 m3 = &m2->mc_xcursor->mx_cursor;
1389 if (m3->mc_snum < mc->mc_snum) continue;
1390 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1391 m3->mc_pg[mc->mc_top] = mp;
1397 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1398 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1399 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1400 m2->mc_pg[mc->mc_top] = mp;
1404 mc->mc_pg[mc->mc_top] = mp;
1405 /** If this page has a parent, update the parent to point to
1409 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1411 mc->mc_db->md_root = mp->mp_pgno;
1412 } else if (mc->mc_txn->mt_parent) {
1415 /* If txn has a parent, make sure the page is in our
1418 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1419 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1420 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1421 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1422 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1423 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1424 mc->mc_pg[mc->mc_top] = mp;
1430 np = mdb_page_malloc(mc);
1431 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1432 mid.mid = np->mp_pgno;
1434 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1442 mdb_env_sync(MDB_env *env, int force)
1445 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1446 if (MDB_FDATASYNC(env->me_fd))
1452 /** Make shadow copies of all of parent txn's cursors */
1454 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1456 MDB_cursor *mc, *m2;
1457 unsigned int i, j, size;
1459 for (i=0;i<src->mt_numdbs; i++) {
1460 if (src->mt_cursors[i]) {
1461 size = sizeof(MDB_cursor);
1462 if (src->mt_cursors[i]->mc_xcursor)
1463 size += sizeof(MDB_xcursor);
1464 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1471 mc->mc_db = &dst->mt_dbs[i];
1472 mc->mc_dbx = m2->mc_dbx;
1473 mc->mc_dbflag = &dst->mt_dbflags[i];
1474 mc->mc_snum = m2->mc_snum;
1475 mc->mc_top = m2->mc_top;
1476 mc->mc_flags = m2->mc_flags | C_SHADOW;
1477 for (j=0; j<mc->mc_snum; j++) {
1478 mc->mc_pg[j] = m2->mc_pg[j];
1479 mc->mc_ki[j] = m2->mc_ki[j];
1481 if (m2->mc_xcursor) {
1482 MDB_xcursor *mx, *mx2;
1483 mx = (MDB_xcursor *)(mc+1);
1484 mc->mc_xcursor = mx;
1485 mx2 = m2->mc_xcursor;
1486 mx->mx_db = mx2->mx_db;
1487 mx->mx_dbx = mx2->mx_dbx;
1488 mx->mx_dbflag = mx2->mx_dbflag;
1489 mx->mx_cursor.mc_txn = dst;
1490 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1491 mx->mx_cursor.mc_db = &mx->mx_db;
1492 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1493 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1494 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1495 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1496 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1497 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1498 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1499 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1502 mc->mc_xcursor = NULL;
1504 mc->mc_next = dst->mt_cursors[i];
1505 dst->mt_cursors[i] = mc;
1512 /** Merge shadow cursors back into parent's */
1514 mdb_cursor_merge(MDB_txn *txn)
1517 for (i=0; i<txn->mt_numdbs; i++) {
1518 if (txn->mt_cursors[i]) {
1520 while ((mc = txn->mt_cursors[i])) {
1521 txn->mt_cursors[i] = mc->mc_next;
1522 if (mc->mc_flags & C_SHADOW) {
1523 MDB_cursor *m2 = mc->mc_orig;
1525 m2->mc_snum = mc->mc_snum;
1526 m2->mc_top = mc->mc_top;
1527 for (j=0; j<mc->mc_snum; j++) {
1528 m2->mc_pg[j] = mc->mc_pg[j];
1529 m2->mc_ki[j] = mc->mc_ki[j];
1532 if (mc->mc_flags & C_ALLOCD)
1540 mdb_txn_reset0(MDB_txn *txn);
1542 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1543 * @param[in] txn the transaction handle to initialize
1544 * @return 0 on success, non-zero on failure. This can only
1545 * fail for read-only transactions, and then only if the
1546 * reader table is full.
1549 mdb_txn_renew0(MDB_txn *txn)
1551 MDB_env *env = txn->mt_env;
1555 txn->mt_numdbs = env->me_numdbs;
1556 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1558 if (txn->mt_flags & MDB_TXN_RDONLY) {
1559 MDB_reader *r = pthread_getspecific(env->me_txkey);
1561 pid_t pid = getpid();
1562 pthread_t tid = pthread_self();
1565 for (i=0; i<env->me_txns->mti_numreaders; i++)
1566 if (env->me_txns->mti_readers[i].mr_pid == 0)
1568 if (i == env->me_maxreaders) {
1569 UNLOCK_MUTEX_R(env);
1572 env->me_txns->mti_readers[i].mr_pid = pid;
1573 env->me_txns->mti_readers[i].mr_tid = tid;
1574 if (i >= env->me_txns->mti_numreaders)
1575 env->me_txns->mti_numreaders = i+1;
1576 UNLOCK_MUTEX_R(env);
1577 r = &env->me_txns->mti_readers[i];
1578 pthread_setspecific(env->me_txkey, r);
1580 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1581 txn->mt_toggle = txn->mt_txnid & 1;
1582 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1583 txn->mt_u.reader = r;
1587 txn->mt_txnid = env->me_txns->mti_txnid;
1588 txn->mt_toggle = txn->mt_txnid & 1;
1589 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1592 if (txn->mt_txnid == mdb_debug_start)
1595 txn->mt_u.dirty_list = env->me_dirty_list;
1596 txn->mt_u.dirty_list[0].mid = 0;
1597 txn->mt_free_pgs = env->me_free_pgs;
1598 txn->mt_free_pgs[0] = 0;
1602 /* Copy the DB info and flags */
1603 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1604 for (i=2; i<txn->mt_numdbs; i++)
1605 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1606 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1607 memset(txn->mt_dbflags+2, DB_STALE, env->me_numdbs-2);
1613 mdb_txn_renew(MDB_txn *txn)
1620 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1621 DPUTS("environment had fatal error, must shutdown!");
1625 rc = mdb_txn_renew0(txn);
1626 if (rc == MDB_SUCCESS) {
1627 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1628 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1629 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1635 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1640 if (env->me_flags & MDB_FATAL_ERROR) {
1641 DPUTS("environment had fatal error, must shutdown!");
1645 /* parent already has an active child txn */
1646 if (parent->mt_child) {
1650 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1651 if (!(flags & MDB_RDONLY))
1652 size += env->me_maxdbs * sizeof(MDB_cursor *);
1654 if ((txn = calloc(1, size)) == NULL) {
1655 DPRINTF("calloc: %s", strerror(ErrCode()));
1658 txn->mt_dbs = (MDB_db *)(txn+1);
1659 if (flags & MDB_RDONLY) {
1660 txn->mt_flags |= MDB_TXN_RDONLY;
1661 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1663 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1664 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1669 txn->mt_free_pgs = mdb_midl_alloc();
1670 if (!txn->mt_free_pgs) {
1674 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1675 if (!txn->mt_u.dirty_list) {
1676 free(txn->mt_free_pgs);
1680 txn->mt_txnid = parent->mt_txnid;
1681 txn->mt_toggle = parent->mt_toggle;
1682 txn->mt_u.dirty_list[0].mid = 0;
1683 txn->mt_free_pgs[0] = 0;
1684 txn->mt_next_pgno = parent->mt_next_pgno;
1685 parent->mt_child = txn;
1686 txn->mt_parent = parent;
1687 txn->mt_numdbs = parent->mt_numdbs;
1688 txn->mt_dbxs = parent->mt_dbxs;
1689 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1690 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1691 mdb_cursor_shadow(parent, txn);
1694 rc = mdb_txn_renew0(txn);
1700 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1701 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1702 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1708 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1709 * @param[in] txn the transaction handle to reset
1712 mdb_txn_reset0(MDB_txn *txn)
1714 MDB_env *env = txn->mt_env;
1716 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1717 txn->mt_u.reader->mr_txnid = 0;
1723 /* close(free) all cursors */
1724 for (i=0; i<txn->mt_numdbs; i++) {
1725 if (txn->mt_cursors[i]) {
1727 while ((mc = txn->mt_cursors[i])) {
1728 txn->mt_cursors[i] = mc->mc_next;
1729 if (mc->mc_flags & C_ALLOCD)
1735 /* return all dirty pages to dpage list */
1736 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1737 dp = txn->mt_u.dirty_list[i].mptr;
1738 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1739 dp->mp_next = txn->mt_env->me_dpages;
1740 VGMEMP_FREE(txn->mt_env, dp);
1741 txn->mt_env->me_dpages = dp;
1743 /* large pages just get freed directly */
1744 VGMEMP_FREE(txn->mt_env, dp);
1749 if (txn->mt_parent) {
1750 txn->mt_parent->mt_child = NULL;
1751 free(txn->mt_free_pgs);
1752 free(txn->mt_u.dirty_list);
1755 if (mdb_midl_shrink(&txn->mt_free_pgs))
1756 env->me_free_pgs = txn->mt_free_pgs;
1759 while ((mop = txn->mt_env->me_pghead)) {
1760 txn->mt_env->me_pghead = mop->mo_next;
1763 txn->mt_env->me_pgfirst = 0;
1764 txn->mt_env->me_pglast = 0;
1767 /* The writer mutex was locked in mdb_txn_begin. */
1768 UNLOCK_MUTEX_W(env);
1773 mdb_txn_reset(MDB_txn *txn)
1778 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1779 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1780 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1782 mdb_txn_reset0(txn);
1786 mdb_txn_abort(MDB_txn *txn)
1791 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1792 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1793 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1796 mdb_txn_abort(txn->mt_child);
1798 mdb_txn_reset0(txn);
1803 mdb_txn_commit(MDB_txn *txn)
1811 pgno_t next, freecnt;
1814 assert(txn != NULL);
1815 assert(txn->mt_env != NULL);
1817 if (txn->mt_child) {
1818 mdb_txn_commit(txn->mt_child);
1819 txn->mt_child = NULL;
1824 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1825 if (txn->mt_numdbs > env->me_numdbs) {
1826 /* update the DB flags */
1828 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
1829 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
1836 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1837 DPUTS("error flag is set, can't commit");
1839 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1844 /* Merge (and close) our cursors with parent's */
1845 mdb_cursor_merge(txn);
1847 if (txn->mt_parent) {
1853 /* Update parent's DB table */
1854 ip = &txn->mt_parent->mt_dbs[2];
1855 jp = &txn->mt_dbs[2];
1856 for (i = 2; i < txn->mt_numdbs; i++) {
1857 if (ip->md_root != jp->md_root)
1861 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1863 /* Append our free list to parent's */
1864 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1866 mdb_midl_free(txn->mt_free_pgs);
1868 /* Merge our dirty list with parent's */
1869 dst = txn->mt_parent->mt_u.dirty_list;
1870 src = txn->mt_u.dirty_list;
1871 x = mdb_mid2l_search(dst, src[1].mid);
1872 for (y=1; y<=src[0].mid; y++) {
1873 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1877 dst[x].mptr = src[y].mptr;
1880 for (; y<=src[0].mid; y++) {
1881 if (++x >= MDB_IDL_UM_MAX) {
1888 free(txn->mt_u.dirty_list);
1889 txn->mt_parent->mt_child = NULL;
1894 if (txn != env->me_txn) {
1895 DPUTS("attempt to commit unknown transaction");
1900 if (!txn->mt_u.dirty_list[0].mid)
1903 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1904 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1906 /* Update DB root pointers. Their pages have already been
1907 * touched so this is all in-place and cannot fail.
1912 data.mv_size = sizeof(MDB_db);
1914 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
1915 for (i = 2; i < txn->mt_numdbs; i++) {
1916 if (txn->mt_dbflags[i] & DB_DIRTY) {
1917 data.mv_data = &txn->mt_dbs[i];
1918 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1923 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1925 /* should only be one record now */
1926 if (env->me_pghead) {
1927 /* make sure first page of freeDB is touched and on freelist */
1928 mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
1931 /* Delete IDLs we used from the free list */
1932 if (env->me_pgfirst) {
1937 key.mv_size = sizeof(cur);
1938 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
1941 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
1942 rc = mdb_cursor_del(&mc, 0);
1948 env->me_pgfirst = 0;
1952 /* save to free list */
1954 freecnt = txn->mt_free_pgs[0];
1955 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1958 /* make sure last page of freeDB is touched and on freelist */
1959 key.mv_size = MAXKEYSIZE+1;
1961 mdb_page_search(&mc, &key, MDB_PS_MODIFY);
1963 mdb_midl_sort(txn->mt_free_pgs);
1967 MDB_IDL idl = txn->mt_free_pgs;
1968 DPRINTF("IDL write txn %zu root %zu num %zu",
1969 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1970 for (i=0; i<idl[0]; i++) {
1971 DPRINTF("IDL %zu", idl[i+1]);
1975 /* write to last page of freeDB */
1976 key.mv_size = sizeof(pgno_t);
1977 key.mv_data = &txn->mt_txnid;
1978 data.mv_data = txn->mt_free_pgs;
1979 /* The free list can still grow during this call,
1980 * despite the pre-emptive touches above. So check
1981 * and make sure the entire thing got written.
1984 freecnt = txn->mt_free_pgs[0];
1985 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1986 rc = mdb_cursor_put(&mc, &key, &data, 0);
1991 } while (freecnt != txn->mt_free_pgs[0]);
1993 /* should only be one record now */
1995 if (env->me_pghead) {
2001 mop = env->me_pghead;
2003 key.mv_size = sizeof(id);
2005 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2006 data.mv_data = mop->mo_pages;
2007 orig = mop->mo_pages[0];
2008 /* These steps may grow the freelist again
2009 * due to freed overflow pages...
2011 mdb_cursor_put(&mc, &key, &data, 0);
2012 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2013 /* could have been used again here */
2014 if (mop->mo_pages[0] != orig) {
2015 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2016 data.mv_data = mop->mo_pages;
2018 mdb_cursor_put(&mc, &key, &data, 0);
2020 env->me_pghead = NULL;
2023 /* was completely used up */
2024 mdb_cursor_del(&mc, 0);
2028 env->me_pgfirst = 0;
2032 while (env->me_pgfree) {
2033 MDB_oldpages *mop = env->me_pgfree;
2034 env->me_pgfree = mop->mo_next;
2038 /* Check for growth of freelist again */
2039 if (freecnt != txn->mt_free_pgs[0])
2042 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2043 if (mdb_midl_shrink(&txn->mt_free_pgs))
2044 env->me_free_pgs = txn->mt_free_pgs;
2051 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2057 /* Windows actually supports scatter/gather I/O, but only on
2058 * unbuffered file handles. Since we're relying on the OS page
2059 * cache for all our data, that's self-defeating. So we just
2060 * write pages one at a time. We use the ov structure to set
2061 * the write offset, to at least save the overhead of a Seek
2065 memset(&ov, 0, sizeof(ov));
2066 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2068 dp = txn->mt_u.dirty_list[i].mptr;
2069 DPRINTF("committing page %zu", dp->mp_pgno);
2070 size = dp->mp_pgno * env->me_psize;
2071 ov.Offset = size & 0xffffffff;
2072 ov.OffsetHigh = size >> 16;
2073 ov.OffsetHigh >>= 16;
2074 /* clear dirty flag */
2075 dp->mp_flags &= ~P_DIRTY;
2076 wsize = env->me_psize;
2077 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2078 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2081 DPRINTF("WriteFile: %d", n);
2088 struct iovec iov[MDB_COMMIT_PAGES];
2092 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2093 dp = txn->mt_u.dirty_list[i].mptr;
2094 if (dp->mp_pgno != next) {
2096 rc = writev(env->me_fd, iov, n);
2100 DPUTS("short write, filesystem full?");
2102 DPRINTF("writev: %s", strerror(n));
2109 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2112 DPRINTF("committing page %zu", dp->mp_pgno);
2113 iov[n].iov_len = env->me_psize;
2114 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2115 iov[n].iov_base = (char *)dp;
2116 size += iov[n].iov_len;
2117 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2118 /* clear dirty flag */
2119 dp->mp_flags &= ~P_DIRTY;
2120 if (++n >= MDB_COMMIT_PAGES) {
2130 rc = writev(env->me_fd, iov, n);
2134 DPUTS("short write, filesystem full?");
2136 DPRINTF("writev: %s", strerror(n));
2143 /* Drop the dirty pages.
2145 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2146 dp = txn->mt_u.dirty_list[i].mptr;
2147 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2148 dp->mp_next = txn->mt_env->me_dpages;
2149 VGMEMP_FREE(txn->mt_env, dp);
2150 txn->mt_env->me_dpages = dp;
2152 VGMEMP_FREE(txn->mt_env, dp);
2155 txn->mt_u.dirty_list[i].mid = 0;
2157 txn->mt_u.dirty_list[0].mid = 0;
2159 if ((n = mdb_env_sync(env, 0)) != 0 ||
2160 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2167 if (txn->mt_numdbs > env->me_numdbs) {
2168 /* update the DB flags */
2170 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2171 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2175 UNLOCK_MUTEX_W(env);
2181 /** Read the environment parameters of a DB environment before
2182 * mapping it into memory.
2183 * @param[in] env the environment handle
2184 * @param[out] meta address of where to store the meta information
2185 * @return 0 on success, non-zero on failure.
2188 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2195 /* We don't know the page size yet, so use a minimum value.
2199 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2201 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2206 else if (rc != MDB_PAGESIZE) {
2210 DPRINTF("read: %s", strerror(err));
2214 p = (MDB_page *)&pbuf;
2216 if (!F_ISSET(p->mp_flags, P_META)) {
2217 DPRINTF("page %zu not a meta page", p->mp_pgno);
2222 if (m->mm_magic != MDB_MAGIC) {
2223 DPUTS("meta has invalid magic");
2227 if (m->mm_version != MDB_VERSION) {
2228 DPRINTF("database is version %u, expected version %u",
2229 m->mm_version, MDB_VERSION);
2230 return MDB_VERSION_MISMATCH;
2233 memcpy(meta, m, sizeof(*m));
2237 /** Write the environment parameters of a freshly created DB environment.
2238 * @param[in] env the environment handle
2239 * @param[out] meta address of where to store the meta information
2240 * @return 0 on success, non-zero on failure.
2243 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2250 DPUTS("writing new meta page");
2252 GET_PAGESIZE(psize);
2254 meta->mm_magic = MDB_MAGIC;
2255 meta->mm_version = MDB_VERSION;
2256 meta->mm_psize = psize;
2257 meta->mm_last_pg = 1;
2258 meta->mm_flags = env->me_flags & 0xffff;
2259 meta->mm_flags |= MDB_INTEGERKEY;
2260 meta->mm_dbs[0].md_root = P_INVALID;
2261 meta->mm_dbs[1].md_root = P_INVALID;
2263 p = calloc(2, psize);
2265 p->mp_flags = P_META;
2268 memcpy(m, meta, sizeof(*meta));
2270 q = (MDB_page *)((char *)p + psize);
2273 q->mp_flags = P_META;
2276 memcpy(m, meta, sizeof(*meta));
2281 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2282 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2285 rc = write(env->me_fd, p, psize * 2);
2286 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2292 /** Update the environment info to commit a transaction.
2293 * @param[in] txn the transaction that's being committed
2294 * @return 0 on success, non-zero on failure.
2297 mdb_env_write_meta(MDB_txn *txn)
2300 MDB_meta meta, metab;
2302 int rc, len, toggle;
2308 assert(txn != NULL);
2309 assert(txn->mt_env != NULL);
2311 toggle = !txn->mt_toggle;
2312 DPRINTF("writing meta page %d for root page %zu",
2313 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2317 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2318 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2320 ptr = (char *)&meta;
2321 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2322 len = sizeof(MDB_meta) - off;
2325 meta.mm_dbs[0] = txn->mt_dbs[0];
2326 meta.mm_dbs[1] = txn->mt_dbs[1];
2327 meta.mm_last_pg = txn->mt_next_pgno - 1;
2328 meta.mm_txnid = txn->mt_txnid;
2331 off += env->me_psize;
2334 /* Write to the SYNC fd */
2337 memset(&ov, 0, sizeof(ov));
2339 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2342 rc = pwrite(env->me_mfd, ptr, len, off);
2347 DPUTS("write failed, disk error?");
2348 /* On a failure, the pagecache still contains the new data.
2349 * Write some old data back, to prevent it from being used.
2350 * Use the non-SYNC fd; we know it will fail anyway.
2352 meta.mm_last_pg = metab.mm_last_pg;
2353 meta.mm_txnid = metab.mm_txnid;
2355 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2357 r2 = pwrite(env->me_fd, ptr, len, off);
2359 env->me_flags |= MDB_FATAL_ERROR;
2362 /* Memory ordering issues are irrelevant; since the entire writer
2363 * is wrapped by wmutex, all of these changes will become visible
2364 * after the wmutex is unlocked. Since the DB is multi-version,
2365 * readers will get consistent data regardless of how fresh or
2366 * how stale their view of these values is.
2368 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2373 /** Check both meta pages to see which one is newer.
2374 * @param[in] env the environment handle
2375 * @return meta toggle (0 or 1).
2378 mdb_env_pick_meta(const MDB_env *env)
2380 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2384 mdb_env_create(MDB_env **env)
2388 e = calloc(1, sizeof(MDB_env));
2392 e->me_free_pgs = mdb_midl_alloc();
2393 if (!e->me_free_pgs) {
2397 e->me_maxreaders = DEFAULT_READERS;
2399 e->me_fd = INVALID_HANDLE_VALUE;
2400 e->me_lfd = INVALID_HANDLE_VALUE;
2401 e->me_mfd = INVALID_HANDLE_VALUE;
2402 VGMEMP_CREATE(e,0,0);
2408 mdb_env_set_mapsize(MDB_env *env, size_t size)
2412 env->me_mapsize = size;
2414 env->me_maxpg = env->me_mapsize / env->me_psize;
2419 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2423 env->me_maxdbs = dbs;
2428 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2430 if (env->me_map || readers < 1)
2432 env->me_maxreaders = readers;
2437 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2439 if (!env || !readers)
2441 *readers = env->me_maxreaders;
2445 /** Further setup required for opening an MDB environment
2448 mdb_env_open2(MDB_env *env, unsigned int flags)
2454 env->me_flags = flags;
2456 memset(&meta, 0, sizeof(meta));
2458 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2461 DPUTS("new mdbenv");
2465 if (!env->me_mapsize) {
2466 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2472 LONG sizelo, sizehi;
2473 sizelo = env->me_mapsize & 0xffffffff;
2474 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2476 /* Windows won't create mappings for zero length files.
2477 * Just allocate the maxsize right now.
2480 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2481 if (!SetEndOfFile(env->me_fd))
2483 SetFilePointer(env->me_fd, 0, NULL, 0);
2485 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
2486 sizehi, sizelo, NULL);
2489 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
2497 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2499 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
2501 if (env->me_map == MAP_FAILED) {
2508 meta.mm_mapsize = env->me_mapsize;
2509 if (flags & MDB_FIXEDMAP)
2510 meta.mm_address = env->me_map;
2511 i = mdb_env_init_meta(env, &meta);
2512 if (i != MDB_SUCCESS) {
2513 munmap(env->me_map, env->me_mapsize);
2517 env->me_psize = meta.mm_psize;
2519 env->me_maxpg = env->me_mapsize / env->me_psize;
2521 p = (MDB_page *)env->me_map;
2522 env->me_metas[0] = METADATA(p);
2523 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2527 int toggle = mdb_env_pick_meta(env);
2528 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2530 DPRINTF("opened database version %u, pagesize %u",
2531 env->me_metas[0]->mm_version, env->me_psize);
2532 DPRINTF("using meta page %d", toggle);
2533 DPRINTF("depth: %u", db->md_depth);
2534 DPRINTF("entries: %zu", db->md_entries);
2535 DPRINTF("branch pages: %zu", db->md_branch_pages);
2536 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2537 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2538 DPRINTF("root: %zu", db->md_root);
2546 /** Release a reader thread's slot in the reader lock table.
2547 * This function is called automatically when a thread exits.
2548 * @param[in] ptr This points to the slot in the reader lock table.
2551 mdb_env_reader_dest(void *ptr)
2553 MDB_reader *reader = ptr;
2555 reader->mr_txnid = 0;
2561 /** Junk for arranging thread-specific callbacks on Windows. This is
2562 * necessarily platform and compiler-specific. Windows supports up
2563 * to 1088 keys. Let's assume nobody opens more than 64 environments
2564 * in a single process, for now. They can override this if needed.
2566 #ifndef MAX_TLS_KEYS
2567 #define MAX_TLS_KEYS 64
2569 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2570 static int mdb_tls_nkeys;
2572 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2576 case DLL_PROCESS_ATTACH: break;
2577 case DLL_THREAD_ATTACH: break;
2578 case DLL_THREAD_DETACH:
2579 for (i=0; i<mdb_tls_nkeys; i++) {
2580 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2581 mdb_env_reader_dest(r);
2584 case DLL_PROCESS_DETACH: break;
2589 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2591 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2595 /* Force some symbol references.
2596 * _tls_used forces the linker to create the TLS directory if not already done
2597 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2599 #pragma comment(linker, "/INCLUDE:_tls_used")
2600 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2601 #pragma const_seg(".CRT$XLB")
2602 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2603 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2606 #pragma comment(linker, "/INCLUDE:__tls_used")
2607 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2608 #pragma data_seg(".CRT$XLB")
2609 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2611 #endif /* WIN 32/64 */
2612 #endif /* !__GNUC__ */
2615 /** Downgrade the exclusive lock on the region back to shared */
2617 mdb_env_share_locks(MDB_env *env)
2619 int toggle = mdb_env_pick_meta(env);
2621 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2626 /* First acquire a shared lock. The Unlock will
2627 * then release the existing exclusive lock.
2629 memset(&ov, 0, sizeof(ov));
2630 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2631 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2635 struct flock lock_info;
2636 /* The shared lock replaces the existing lock */
2637 memset((void *)&lock_info, 0, sizeof(lock_info));
2638 lock_info.l_type = F_RDLCK;
2639 lock_info.l_whence = SEEK_SET;
2640 lock_info.l_start = 0;
2641 lock_info.l_len = 1;
2642 fcntl(env->me_lfd, F_SETLK, &lock_info);
2646 #if defined(_WIN32) || defined(__APPLE__)
2648 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2650 * @(#) $Revision: 5.1 $
2651 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2652 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2654 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2658 * Please do not copyright this code. This code is in the public domain.
2660 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2661 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2662 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2663 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2664 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2665 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2666 * PERFORMANCE OF THIS SOFTWARE.
2669 * chongo <Landon Curt Noll> /\oo/\
2670 * http://www.isthe.com/chongo/
2672 * Share and Enjoy! :-)
2675 typedef unsigned long long mdb_hash_t;
2676 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2678 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2679 * @param[in] str string to hash
2680 * @param[in] hval initial value for hash
2681 * @return 64 bit hash
2683 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2684 * hval arg on the first call.
2687 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
2689 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
2690 unsigned char *end = s + val->mv_size;
2692 * FNV-1a hash each octet of the string
2695 /* xor the bottom with the current octet */
2696 hval ^= (mdb_hash_t)*s++;
2698 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2699 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2700 (hval << 7) + (hval << 8) + (hval << 40);
2702 /* return our new hash value */
2706 /** Hash the string and output the hash in hex.
2707 * @param[in] str string to hash
2708 * @param[out] hexbuf an array of 17 chars to hold the hash
2711 mdb_hash_hex(MDB_val *val, char *hexbuf)
2714 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
2715 for (i=0; i<8; i++) {
2716 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2722 /** Open and/or initialize the lock region for the environment.
2723 * @param[in] env The MDB environment.
2724 * @param[in] lpath The pathname of the file used for the lock region.
2725 * @param[in] mode The Unix permissions for the file, if we create it.
2726 * @param[out] excl Set to true if we got an exclusive lock on the region.
2727 * @return 0 on success, non-zero on failure.
2730 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2738 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2739 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2740 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2744 /* Try to get exclusive lock. If we succeed, then
2745 * nobody is using the lock region and we should initialize it.
2748 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2752 memset(&ov, 0, sizeof(ov));
2753 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2759 size = GetFileSize(env->me_lfd, NULL);
2765 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1)
2767 /* Lose record locks when exec*() */
2768 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
2769 fcntl(env->me_lfd, F_SETFD, fdflags);
2771 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
2772 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1)
2776 /* Try to get exclusive lock. If we succeed, then
2777 * nobody is using the lock region and we should initialize it.
2780 struct flock lock_info;
2781 memset((void *)&lock_info, 0, sizeof(lock_info));
2782 lock_info.l_type = F_WRLCK;
2783 lock_info.l_whence = SEEK_SET;
2784 lock_info.l_start = 0;
2785 lock_info.l_len = 1;
2786 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2790 lock_info.l_type = F_RDLCK;
2791 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2798 size = lseek(env->me_lfd, 0, SEEK_END);
2800 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2801 if (size < rsize && *excl) {
2803 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2804 if (!SetEndOfFile(env->me_lfd)) {
2809 if (ftruncate(env->me_lfd, rsize) != 0) {
2816 size = rsize - sizeof(MDB_txninfo);
2817 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2822 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2828 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2830 if (!env->me_txns) {
2835 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2837 if (m == MAP_FAILED) {
2838 env->me_txns = NULL;
2847 BY_HANDLE_FILE_INFORMATION stbuf;
2856 if (!mdb_sec_inited) {
2857 InitializeSecurityDescriptor(&mdb_null_sd,
2858 SECURITY_DESCRIPTOR_REVISION);
2859 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2860 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2861 mdb_all_sa.bInheritHandle = FALSE;
2862 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2865 GetFileInformationByHandle(env->me_lfd, &stbuf);
2866 idbuf.volume = stbuf.dwVolumeSerialNumber;
2867 idbuf.nhigh = stbuf.nFileIndexHigh;
2868 idbuf.nlow = stbuf.nFileIndexLow;
2869 val.mv_data = &idbuf;
2870 val.mv_size = sizeof(idbuf);
2871 mdb_hash_hex(&val, hexbuf);
2872 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
2873 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2874 if (!env->me_rmutex) {
2878 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
2879 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
2880 if (!env->me_wmutex) {
2894 fstat(env->me_lfd, &stbuf);
2895 idbuf.dev = stbuf.st_dev;
2896 idbuf.ino = stbuf.st_ino;
2897 val.mv_data = &idbuf;
2898 val.mv_size = sizeof(idbuf);
2899 mdb_hash_hex(&val, hexbuf);
2900 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
2901 if (sem_unlink(env->me_txns->mti_rmname)) {
2903 if (rc != ENOENT && rc != EINVAL)
2906 env->me_rmutex = sem_open(env->me_txns->mti_rmname, O_CREAT, mode, 1);
2907 if (!env->me_rmutex) {
2911 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
2912 if (sem_unlink(env->me_txns->mti_wmname)) {
2914 if (rc != ENOENT && rc != EINVAL)
2917 env->me_wmutex = sem_open(env->me_txns->mti_wmname, O_CREAT, mode, 1);
2918 if (!env->me_wmutex) {
2922 #else /* __APPLE__ */
2923 pthread_mutexattr_t mattr;
2925 pthread_mutexattr_init(&mattr);
2926 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2930 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2931 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2932 #endif /* __APPLE__ */
2934 env->me_txns->mti_version = MDB_VERSION;
2935 env->me_txns->mti_magic = MDB_MAGIC;
2936 env->me_txns->mti_txnid = 0;
2937 env->me_txns->mti_numreaders = 0;
2940 if (env->me_txns->mti_magic != MDB_MAGIC) {
2941 DPUTS("lock region has invalid magic");
2945 if (env->me_txns->mti_version != MDB_VERSION) {
2946 DPRINTF("lock region is version %u, expected version %u",
2947 env->me_txns->mti_version, MDB_VERSION);
2948 rc = MDB_VERSION_MISMATCH;
2952 if (rc != EACCES && rc != EAGAIN) {
2956 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2957 if (!env->me_rmutex) {
2961 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2962 if (!env->me_wmutex) {
2968 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
2969 if (!env->me_rmutex) {
2973 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
2974 if (!env->me_wmutex) {
2984 env->me_lfd = INVALID_HANDLE_VALUE;
2989 /** The name of the lock file in the DB environment */
2990 #define LOCKNAME "/lock.mdb"
2991 /** The name of the data file in the DB environment */
2992 #define DATANAME "/data.mdb"
2993 /** The suffix of the lock file when no subdir is used */
2994 #define LOCKSUFF "-lock"
2997 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2999 int oflags, rc, len, excl;
3000 char *lpath, *dpath;
3003 if (flags & MDB_NOSUBDIR) {
3004 rc = len + sizeof(LOCKSUFF) + len + 1;
3006 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3011 if (flags & MDB_NOSUBDIR) {
3012 dpath = lpath + len + sizeof(LOCKSUFF);
3013 sprintf(lpath, "%s" LOCKSUFF, path);
3014 strcpy(dpath, path);
3016 dpath = lpath + len + sizeof(LOCKNAME);
3017 sprintf(lpath, "%s" LOCKNAME, path);
3018 sprintf(dpath, "%s" DATANAME, path);
3021 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3026 if (F_ISSET(flags, MDB_RDONLY)) {
3027 oflags = GENERIC_READ;
3028 len = OPEN_EXISTING;
3030 oflags = GENERIC_READ|GENERIC_WRITE;
3033 mode = FILE_ATTRIBUTE_NORMAL;
3034 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3035 NULL, len, mode, NULL);
3037 if (F_ISSET(flags, MDB_RDONLY))
3040 oflags = O_RDWR | O_CREAT;
3042 env->me_fd = open(dpath, oflags, mode);
3044 if (env->me_fd == INVALID_HANDLE_VALUE) {
3049 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
3050 if (flags & (MDB_RDONLY|MDB_NOSYNC|MDB_NOMETASYNC)) {
3051 env->me_mfd = env->me_fd;
3053 /* synchronous fd for meta writes */
3055 env->me_mfd = CreateFile(dpath, oflags,
3056 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3057 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3059 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3061 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3066 env->me_path = strdup(path);
3067 DPRINTF("opened dbenv %p", (void *) env);
3068 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3070 /* Windows TLS callbacks need help finding their TLS info. */
3071 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3072 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3079 mdb_env_share_locks(env);
3081 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3082 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3083 if (!env->me_dbxs || !env->me_dbflags)
3089 if (env->me_fd != INVALID_HANDLE_VALUE) {
3091 env->me_fd = INVALID_HANDLE_VALUE;
3093 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3095 env->me_lfd = INVALID_HANDLE_VALUE;
3103 mdb_env_close(MDB_env *env)
3110 VGMEMP_DESTROY(env);
3111 while (env->me_dpages) {
3112 dp = env->me_dpages;
3113 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3114 env->me_dpages = dp->mp_next;
3118 free(env->me_dbflags);
3122 pthread_key_delete(env->me_txkey);
3124 /* Delete our key from the global list */
3126 for (i=0; i<mdb_tls_nkeys; i++)
3127 if (mdb_tls_keys[i] == env->me_txkey) {
3128 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3136 munmap(env->me_map, env->me_mapsize);
3138 if (env->me_mfd != env->me_fd)
3142 pid_t pid = getpid();
3144 for (i=0; i<env->me_txns->mti_numreaders; i++)
3145 if (env->me_txns->mti_readers[i].mr_pid == pid)
3146 env->me_txns->mti_readers[i].mr_pid = 0;
3147 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3150 mdb_midl_free(env->me_free_pgs);
3154 /** Compare two items pointing at aligned size_t's */
3156 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3158 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3159 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3162 /** Compare two items pointing at aligned int's */
3164 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3166 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3167 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3170 /** Compare two items pointing at ints of unknown alignment.
3171 * Nodes and keys are guaranteed to be 2-byte aligned.
3174 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3176 #if BYTE_ORDER == LITTLE_ENDIAN
3177 unsigned short *u, *c;
3180 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3181 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3184 } while(!x && u > (unsigned short *)a->mv_data);
3187 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3191 /** Compare two items lexically */
3193 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3200 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3206 diff = memcmp(a->mv_data, b->mv_data, len);
3207 return diff ? diff : len_diff<0 ? -1 : len_diff;
3210 /** Compare two items in reverse byte order */
3212 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3214 const unsigned char *p1, *p2, *p1_lim;
3218 p1_lim = (const unsigned char *)a->mv_data;
3219 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3220 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3222 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3228 while (p1 > p1_lim) {
3229 diff = *--p1 - *--p2;
3233 return len_diff<0 ? -1 : len_diff;
3236 /** Search for key within a page, using binary search.
3237 * Returns the smallest entry larger or equal to the key.
3238 * If exactp is non-null, stores whether the found entry was an exact match
3239 * in *exactp (1 or 0).
3240 * Updates the cursor index with the index of the found entry.
3241 * If no entry larger or equal to the key is found, returns NULL.
3244 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3246 unsigned int i = 0, nkeys;
3249 MDB_page *mp = mc->mc_pg[mc->mc_top];
3250 MDB_node *node = NULL;
3255 nkeys = NUMKEYS(mp);
3260 COPY_PGNO(pgno, mp->mp_pgno);
3261 DPRINTF("searching %u keys in %s %spage %zu",
3262 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3269 low = IS_LEAF(mp) ? 0 : 1;
3271 cmp = mc->mc_dbx->md_cmp;
3273 /* Branch pages have no data, so if using integer keys,
3274 * alignment is guaranteed. Use faster mdb_cmp_int.
3276 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3277 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3284 nodekey.mv_size = mc->mc_db->md_pad;
3285 node = NODEPTR(mp, 0); /* fake */
3286 while (low <= high) {
3287 i = (low + high) >> 1;
3288 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3289 rc = cmp(key, &nodekey);
3290 DPRINTF("found leaf index %u [%s], rc = %i",
3291 i, DKEY(&nodekey), rc);
3300 while (low <= high) {
3301 i = (low + high) >> 1;
3303 node = NODEPTR(mp, i);
3304 nodekey.mv_size = NODEKSZ(node);
3305 nodekey.mv_data = NODEKEY(node);
3307 rc = cmp(key, &nodekey);
3310 DPRINTF("found leaf index %u [%s], rc = %i",
3311 i, DKEY(&nodekey), rc);
3313 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3314 i, DKEY(&nodekey), NODEPGNO(node), rc);
3325 if (rc > 0) { /* Found entry is less than the key. */
3326 i++; /* Skip to get the smallest entry larger than key. */
3328 node = NODEPTR(mp, i);
3331 *exactp = (rc == 0);
3332 /* store the key index */
3333 mc->mc_ki[mc->mc_top] = i;
3335 /* There is no entry larger or equal to the key. */
3338 /* nodeptr is fake for LEAF2 */
3344 mdb_cursor_adjust(MDB_cursor *mc, func)
3348 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3349 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3356 /** Pop a page off the top of the cursor's stack. */
3358 mdb_cursor_pop(MDB_cursor *mc)
3363 top = mc->mc_pg[mc->mc_top];
3368 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3369 mc->mc_dbi, (void *) mc);
3373 /** Push a page onto the top of the cursor's stack. */
3375 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3377 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3378 mc->mc_dbi, (void *) mc);
3380 if (mc->mc_snum >= CURSOR_STACK) {
3381 assert(mc->mc_snum < CURSOR_STACK);
3385 mc->mc_top = mc->mc_snum++;
3386 mc->mc_pg[mc->mc_top] = mp;
3387 mc->mc_ki[mc->mc_top] = 0;
3392 /** Find the address of the page corresponding to a given page number.
3393 * @param[in] txn the transaction for this access.
3394 * @param[in] pgno the page number for the page to retrieve.
3395 * @param[out] ret address of a pointer where the page's address will be stored.
3396 * @return 0 on success, non-zero on failure.
3399 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3403 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3405 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3406 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3407 p = txn->mt_u.dirty_list[x].mptr;
3411 if (pgno < txn->mt_next_pgno)
3412 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3416 DPRINTF("page %zu not found", pgno);
3419 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3422 /** Search for the page a given key should be in.
3423 * Pushes parent pages on the cursor stack. This function continues a
3424 * search on a cursor that has already been initialized. (Usually by
3425 * #mdb_page_search() but also by #mdb_node_move().)
3426 * @param[in,out] mc the cursor for this operation.
3427 * @param[in] key the key to search for. If NULL, search for the lowest
3428 * page. (This is used by #mdb_cursor_first().)
3429 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3430 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3431 * @return 0 on success, non-zero on failure.
3434 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3436 MDB_page *mp = mc->mc_pg[mc->mc_top];
3441 while (IS_BRANCH(mp)) {
3445 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3446 assert(NUMKEYS(mp) > 1);
3447 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3449 if (key == NULL) /* Initialize cursor to first page. */
3451 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3452 /* cursor to last page */
3456 node = mdb_node_search(mc, key, &exact);
3458 i = NUMKEYS(mp) - 1;
3460 i = mc->mc_ki[mc->mc_top];
3469 DPRINTF("following index %u for key [%s]",
3471 assert(i < NUMKEYS(mp));
3472 node = NODEPTR(mp, i);
3474 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3477 mc->mc_ki[mc->mc_top] = i;
3478 if ((rc = mdb_cursor_push(mc, mp)))
3482 if ((rc = mdb_page_touch(mc)) != 0)
3484 mp = mc->mc_pg[mc->mc_top];
3489 DPRINTF("internal error, index points to a %02X page!?",
3491 return MDB_CORRUPTED;
3494 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3495 key ? DKEY(key) : NULL);
3500 /** Search for the page a given key should be in.
3501 * Pushes parent pages on the cursor stack. This function just sets up
3502 * the search; it finds the root page for \b mc's database and sets this
3503 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3504 * called to complete the search.
3505 * @param[in,out] mc the cursor for this operation.
3506 * @param[in] key the key to search for. If NULL, search for the lowest
3507 * page. (This is used by #mdb_cursor_first().)
3508 * @param[in] modify If true, visited pages are updated with new page numbers.
3509 * @return 0 on success, non-zero on failure.
3512 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
3517 /* Make sure the txn is still viable, then find the root from
3518 * the txn's db table.
3520 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3521 DPUTS("transaction has failed, must abort");
3524 /* Make sure we're using an up-to-date root */
3525 if (mc->mc_dbi > MAIN_DBI) {
3526 if ((*mc->mc_dbflag & DB_STALE) ||
3527 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
3529 unsigned char dbflag = 0;
3530 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3531 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
3534 if (*mc->mc_dbflag & DB_STALE) {
3537 MDB_node *leaf = mdb_node_search(&mc2,
3538 &mc->mc_dbx->md_name, &exact);
3540 return MDB_NOTFOUND;
3541 mdb_node_read(mc->mc_txn, leaf, &data);
3542 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3544 if (flags & MDB_PS_MODIFY)
3546 *mc->mc_dbflag = dbflag;
3549 root = mc->mc_db->md_root;
3551 if (root == P_INVALID) { /* Tree is empty. */
3552 DPUTS("tree is empty");
3553 return MDB_NOTFOUND;
3558 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
3559 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3565 DPRINTF("db %u root page %zu has flags 0x%X",
3566 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3568 if (flags & MDB_PS_MODIFY) {
3569 if ((rc = mdb_page_touch(mc)))
3573 if (flags & MDB_PS_ROOTONLY)
3576 return mdb_page_search_root(mc, key, flags);
3579 /** Return the data associated with a given node.
3580 * @param[in] txn The transaction for this operation.
3581 * @param[in] leaf The node being read.
3582 * @param[out] data Updated to point to the node's data.
3583 * @return 0 on success, non-zero on failure.
3586 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3588 MDB_page *omp; /* overflow page */
3592 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3593 data->mv_size = NODEDSZ(leaf);
3594 data->mv_data = NODEDATA(leaf);
3598 /* Read overflow data.
3600 data->mv_size = NODEDSZ(leaf);
3601 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3602 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3603 DPRINTF("read overflow page %zu failed", pgno);
3606 data->mv_data = METADATA(omp);
3612 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3613 MDB_val *key, MDB_val *data)
3622 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3624 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3627 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3631 mdb_cursor_init(&mc, txn, dbi, &mx);
3632 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3635 /** Find a sibling for a page.
3636 * Replaces the page at the top of the cursor's stack with the
3637 * specified sibling, if one exists.
3638 * @param[in] mc The cursor for this operation.
3639 * @param[in] move_right Non-zero if the right sibling is requested,
3640 * otherwise the left sibling.
3641 * @return 0 on success, non-zero on failure.
3644 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3650 if (mc->mc_snum < 2) {
3651 return MDB_NOTFOUND; /* root has no siblings */
3655 DPRINTF("parent page is page %zu, index %u",
3656 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3658 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3659 : (mc->mc_ki[mc->mc_top] == 0)) {
3660 DPRINTF("no more keys left, moving to %s sibling",
3661 move_right ? "right" : "left");
3662 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3666 mc->mc_ki[mc->mc_top]++;
3668 mc->mc_ki[mc->mc_top]--;
3669 DPRINTF("just moving to %s index key %u",
3670 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3672 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3674 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3675 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3678 mdb_cursor_push(mc, mp);
3683 /** Move the cursor to the next data item. */
3685 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3691 if (mc->mc_flags & C_EOF) {
3692 return MDB_NOTFOUND;
3695 assert(mc->mc_flags & C_INITIALIZED);
3697 mp = mc->mc_pg[mc->mc_top];
3699 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3700 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3701 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3702 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3703 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3704 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3708 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3709 if (op == MDB_NEXT_DUP)
3710 return MDB_NOTFOUND;
3714 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3716 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3717 DPUTS("=====> move to next sibling page");
3718 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3719 mc->mc_flags |= C_EOF;
3720 mc->mc_flags &= ~C_INITIALIZED;
3721 return MDB_NOTFOUND;
3723 mp = mc->mc_pg[mc->mc_top];
3724 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3726 mc->mc_ki[mc->mc_top]++;
3728 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3729 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3732 key->mv_size = mc->mc_db->md_pad;
3733 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3737 assert(IS_LEAF(mp));
3738 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3740 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3741 mdb_xcursor_init1(mc, leaf);
3744 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3747 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3748 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3749 if (rc != MDB_SUCCESS)
3754 MDB_SET_KEY(leaf, key);
3758 /** Move the cursor to the previous data item. */
3760 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3766 assert(mc->mc_flags & C_INITIALIZED);
3768 mp = mc->mc_pg[mc->mc_top];
3770 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3771 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3772 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3773 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3774 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3775 if (op != MDB_PREV || rc == MDB_SUCCESS)
3778 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3779 if (op == MDB_PREV_DUP)
3780 return MDB_NOTFOUND;
3785 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3787 if (mc->mc_ki[mc->mc_top] == 0) {
3788 DPUTS("=====> move to prev sibling page");
3789 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3790 mc->mc_flags &= ~C_INITIALIZED;
3791 return MDB_NOTFOUND;
3793 mp = mc->mc_pg[mc->mc_top];
3794 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3795 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3797 mc->mc_ki[mc->mc_top]--;
3799 mc->mc_flags &= ~C_EOF;
3801 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3802 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3805 key->mv_size = mc->mc_db->md_pad;
3806 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3810 assert(IS_LEAF(mp));
3811 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3813 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3814 mdb_xcursor_init1(mc, leaf);
3817 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3820 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3821 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3822 if (rc != MDB_SUCCESS)
3827 MDB_SET_KEY(leaf, key);
3831 /** Set the cursor on a specific data item. */
3833 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3834 MDB_cursor_op op, int *exactp)
3843 assert(key->mv_size > 0);
3845 /* See if we're already on the right page */
3846 if (mc->mc_flags & C_INITIALIZED) {
3849 mp = mc->mc_pg[mc->mc_top];
3851 mc->mc_ki[mc->mc_top] = 0;
3852 return MDB_NOTFOUND;
3854 if (mp->mp_flags & P_LEAF2) {
3855 nodekey.mv_size = mc->mc_db->md_pad;
3856 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
3858 leaf = NODEPTR(mp, 0);
3859 MDB_SET_KEY(leaf, &nodekey);
3861 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3863 /* Probably happens rarely, but first node on the page
3864 * was the one we wanted.
3866 mc->mc_ki[mc->mc_top] = 0;
3873 unsigned int nkeys = NUMKEYS(mp);
3875 if (mp->mp_flags & P_LEAF2) {
3876 nodekey.mv_data = LEAF2KEY(mp,
3877 nkeys-1, nodekey.mv_size);
3879 leaf = NODEPTR(mp, nkeys-1);
3880 MDB_SET_KEY(leaf, &nodekey);
3882 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3884 /* last node was the one we wanted */
3885 mc->mc_ki[mc->mc_top] = nkeys-1;
3891 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
3892 /* This is definitely the right page, skip search_page */
3893 if (mp->mp_flags & P_LEAF2) {
3894 nodekey.mv_data = LEAF2KEY(mp,
3895 mc->mc_ki[mc->mc_top], nodekey.mv_size);
3897 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3898 MDB_SET_KEY(leaf, &nodekey);
3900 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3902 /* current node was the one we wanted */
3912 /* If any parents have right-sibs, search.
3913 * Otherwise, there's nothing further.
3915 for (i=0; i<mc->mc_top; i++)
3917 NUMKEYS(mc->mc_pg[i])-1)
3919 if (i == mc->mc_top) {
3920 /* There are no other pages */
3921 mc->mc_ki[mc->mc_top] = nkeys;
3922 return MDB_NOTFOUND;
3926 /* There are no other pages */
3927 mc->mc_ki[mc->mc_top] = 0;
3928 return MDB_NOTFOUND;
3932 rc = mdb_page_search(mc, key, 0);
3933 if (rc != MDB_SUCCESS)
3936 mp = mc->mc_pg[mc->mc_top];
3937 assert(IS_LEAF(mp));
3940 leaf = mdb_node_search(mc, key, exactp);
3941 if (exactp != NULL && !*exactp) {
3942 /* MDB_SET specified and not an exact match. */
3943 return MDB_NOTFOUND;
3947 DPUTS("===> inexact leaf not found, goto sibling");
3948 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
3949 return rc; /* no entries matched */
3950 mp = mc->mc_pg[mc->mc_top];
3951 assert(IS_LEAF(mp));
3952 leaf = NODEPTR(mp, 0);
3956 mc->mc_flags |= C_INITIALIZED;
3957 mc->mc_flags &= ~C_EOF;
3960 key->mv_size = mc->mc_db->md_pad;
3961 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3965 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3966 mdb_xcursor_init1(mc, leaf);
3969 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3970 if (op == MDB_SET || op == MDB_SET_RANGE) {
3971 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3974 if (op == MDB_GET_BOTH) {
3980 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
3981 if (rc != MDB_SUCCESS)
3984 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
3986 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
3988 rc = mc->mc_dbx->md_dcmp(data, &d2);
3990 if (op == MDB_GET_BOTH || rc > 0)
3991 return MDB_NOTFOUND;
3996 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3997 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4002 /* The key already matches in all other cases */
4003 if (op == MDB_SET_RANGE)
4004 MDB_SET_KEY(leaf, key);
4005 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4010 /** Move the cursor to the first item in the database. */
4012 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4017 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4018 rc = mdb_page_search(mc, NULL, 0);
4019 if (rc != MDB_SUCCESS)
4022 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4024 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4025 mc->mc_flags |= C_INITIALIZED;
4026 mc->mc_flags &= ~C_EOF;
4028 mc->mc_ki[mc->mc_top] = 0;
4030 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4031 key->mv_size = mc->mc_db->md_pad;
4032 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4037 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4038 mdb_xcursor_init1(mc, leaf);
4039 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4044 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4045 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4049 MDB_SET_KEY(leaf, key);
4053 /** Move the cursor to the last item in the database. */
4055 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4060 if (!(mc->mc_flags & C_EOF)) {
4062 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4065 lkey.mv_size = MAXKEYSIZE+1;
4066 lkey.mv_data = NULL;
4067 rc = mdb_page_search(mc, &lkey, 0);
4068 if (rc != MDB_SUCCESS)
4071 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4073 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4074 mc->mc_flags |= C_INITIALIZED|C_EOF;
4076 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4078 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4079 key->mv_size = mc->mc_db->md_pad;
4080 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4085 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4086 mdb_xcursor_init1(mc, leaf);
4087 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4092 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4093 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4098 MDB_SET_KEY(leaf, key);
4103 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4113 case MDB_GET_BOTH_RANGE:
4114 if (data == NULL || mc->mc_xcursor == NULL) {
4121 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4123 } else if (op == MDB_SET_RANGE)
4124 rc = mdb_cursor_set(mc, key, data, op, NULL);
4126 rc = mdb_cursor_set(mc, key, data, op, &exact);
4128 case MDB_GET_MULTIPLE:
4130 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4131 !(mc->mc_flags & C_INITIALIZED)) {
4136 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4137 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4140 case MDB_NEXT_MULTIPLE:
4142 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4146 if (!(mc->mc_flags & C_INITIALIZED))
4147 rc = mdb_cursor_first(mc, key, data);
4149 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4150 if (rc == MDB_SUCCESS) {
4151 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4154 mx = &mc->mc_xcursor->mx_cursor;
4155 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4157 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4158 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4166 case MDB_NEXT_NODUP:
4167 if (!(mc->mc_flags & C_INITIALIZED))
4168 rc = mdb_cursor_first(mc, key, data);
4170 rc = mdb_cursor_next(mc, key, data, op);
4174 case MDB_PREV_NODUP:
4175 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4176 rc = mdb_cursor_last(mc, key, data);
4177 mc->mc_flags &= ~C_EOF;
4179 rc = mdb_cursor_prev(mc, key, data, op);
4182 rc = mdb_cursor_first(mc, key, data);
4186 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4187 !(mc->mc_flags & C_INITIALIZED) ||
4188 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4192 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4195 rc = mdb_cursor_last(mc, key, data);
4199 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4200 !(mc->mc_flags & C_INITIALIZED) ||
4201 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4205 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4208 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4216 /** Touch all the pages in the cursor stack.
4217 * Makes sure all the pages are writable, before attempting a write operation.
4218 * @param[in] mc The cursor to operate on.
4221 mdb_cursor_touch(MDB_cursor *mc)
4225 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4227 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4228 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4231 *mc->mc_dbflag = DB_DIRTY;
4233 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4234 rc = mdb_page_touch(mc);
4238 mc->mc_top = mc->mc_snum-1;
4243 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4246 MDB_node *leaf = NULL;
4247 MDB_val xdata, *rdata, dkey;
4250 int do_sub = 0, insert = 0;
4251 unsigned int mcount = 0;
4255 char dbuf[MAXKEYSIZE+1];
4256 unsigned int nflags;
4259 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4262 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4263 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4267 if (flags == MDB_CURRENT) {
4268 if (!(mc->mc_flags & C_INITIALIZED))
4271 } else if (mc->mc_db->md_root == P_INVALID) {
4273 /* new database, write a root leaf page */
4274 DPUTS("allocating new root leaf page");
4275 if ((np = mdb_page_new(mc, P_LEAF, 1)) == NULL) {
4279 mdb_cursor_push(mc, np);
4280 mc->mc_db->md_root = np->mp_pgno;
4281 mc->mc_db->md_depth++;
4282 *mc->mc_dbflag = DB_DIRTY;
4283 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4285 np->mp_flags |= P_LEAF2;
4286 mc->mc_flags |= C_INITIALIZED;
4292 if (flags & MDB_APPEND) {
4294 rc = mdb_cursor_last(mc, &k2, &d2);
4296 rc = mc->mc_dbx->md_cmp(key, &k2);
4299 mc->mc_ki[mc->mc_top]++;
4305 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4307 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4308 DPRINTF("duplicate key [%s]", DKEY(key));
4310 return MDB_KEYEXIST;
4312 if (rc && rc != MDB_NOTFOUND)
4316 /* Cursor is positioned, now make sure all pages are writable */
4317 rc2 = mdb_cursor_touch(mc);
4322 /* The key already exists */
4323 if (rc == MDB_SUCCESS) {
4324 /* there's only a key anyway, so this is a no-op */
4325 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4326 unsigned int ksize = mc->mc_db->md_pad;
4327 if (key->mv_size != ksize)
4329 if (flags == MDB_CURRENT) {
4330 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4331 memcpy(ptr, key->mv_data, ksize);
4336 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4339 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4340 /* Was a single item before, must convert now */
4342 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4343 /* Just overwrite the current item */
4344 if (flags == MDB_CURRENT)
4347 dkey.mv_size = NODEDSZ(leaf);
4348 dkey.mv_data = NODEDATA(leaf);
4349 #if UINT_MAX < SIZE_MAX
4350 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4351 #ifdef MISALIGNED_OK
4352 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4354 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4357 /* if data matches, ignore it */
4358 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4359 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4361 /* create a fake page for the dup items */
4362 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4363 dkey.mv_data = dbuf;
4364 fp = (MDB_page *)&pbuf;
4365 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4366 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4367 fp->mp_lower = PAGEHDRSZ;
4368 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4369 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4370 fp->mp_flags |= P_LEAF2;
4371 fp->mp_pad = data->mv_size;
4372 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4374 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4375 (dkey.mv_size & 1) + (data->mv_size & 1);
4377 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4380 xdata.mv_size = fp->mp_upper;
4385 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4386 /* See if we need to convert from fake page to subDB */
4388 unsigned int offset;
4391 fp = NODEDATA(leaf);
4392 if (flags == MDB_CURRENT) {
4394 fp->mp_flags |= P_DIRTY;
4395 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4396 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4400 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4401 offset = fp->mp_pad;
4402 if (SIZELEFT(fp) >= offset)
4404 offset *= 4; /* space for 4 more */
4406 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4408 offset += offset & 1;
4409 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4410 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4412 /* yes, convert it */
4414 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4415 dummy.md_pad = fp->mp_pad;
4416 dummy.md_flags = MDB_DUPFIXED;
4417 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4418 dummy.md_flags |= MDB_INTEGERKEY;
4421 dummy.md_branch_pages = 0;
4422 dummy.md_leaf_pages = 1;
4423 dummy.md_overflow_pages = 0;
4424 dummy.md_entries = NUMKEYS(fp);
4426 xdata.mv_size = sizeof(MDB_db);
4427 xdata.mv_data = &dummy;
4428 mp = mdb_page_alloc(mc, 1);
4431 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4432 flags |= F_DUPDATA|F_SUBDATA;
4433 dummy.md_root = mp->mp_pgno;
4435 /* no, just grow it */
4437 xdata.mv_size = NODEDSZ(leaf) + offset;
4438 xdata.mv_data = &pbuf;
4439 mp = (MDB_page *)&pbuf;
4440 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4443 mp->mp_flags = fp->mp_flags | P_DIRTY;
4444 mp->mp_pad = fp->mp_pad;
4445 mp->mp_lower = fp->mp_lower;
4446 mp->mp_upper = fp->mp_upper + offset;
4448 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4450 nsize = NODEDSZ(leaf) - fp->mp_upper;
4451 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4452 for (i=0; i<NUMKEYS(fp); i++)
4453 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4455 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4459 /* data is on sub-DB, just store it */
4460 flags |= F_DUPDATA|F_SUBDATA;
4464 /* overflow page overwrites need special handling */
4465 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4468 int ovpages, dpages;
4470 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4471 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4472 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4473 mdb_page_get(mc->mc_txn, pg, &omp);
4474 /* Is the ov page writable and large enough? */
4475 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4476 /* yes, overwrite it. Note in this case we don't
4477 * bother to try shrinking the node if the new data
4478 * is smaller than the overflow threshold.
4480 if (F_ISSET(flags, MDB_RESERVE))
4481 data->mv_data = METADATA(omp);
4483 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4486 /* no, free ovpages */
4488 mc->mc_db->md_overflow_pages -= ovpages;
4489 for (i=0; i<ovpages; i++) {
4490 DPRINTF("freed ov page %zu", pg);
4491 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4495 } else if (NODEDSZ(leaf) == data->mv_size) {
4496 /* same size, just replace it. Note that we could
4497 * also reuse this node if the new data is smaller,
4498 * but instead we opt to shrink the node in that case.
4500 if (F_ISSET(flags, MDB_RESERVE))
4501 data->mv_data = NODEDATA(leaf);
4503 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4506 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4507 mc->mc_db->md_entries--;
4509 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4516 nflags = flags & NODE_ADD_FLAGS;
4517 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4518 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4519 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4520 nflags &= ~MDB_APPEND;
4522 nflags |= MDB_SPLIT_REPLACE;
4523 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4525 /* There is room already in this leaf page. */
4526 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4527 if (rc == 0 && !do_sub && insert) {
4528 /* Adjust other cursors pointing to mp */
4529 MDB_cursor *m2, *m3;
4530 MDB_dbi dbi = mc->mc_dbi;
4531 unsigned i = mc->mc_top;
4532 MDB_page *mp = mc->mc_pg[i];
4534 if (mc->mc_flags & C_SUB)
4537 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4538 if (mc->mc_flags & C_SUB)
4539 m3 = &m2->mc_xcursor->mx_cursor;
4542 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4543 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4550 if (rc != MDB_SUCCESS)
4551 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4553 /* Now store the actual data in the child DB. Note that we're
4554 * storing the user data in the keys field, so there are strict
4555 * size limits on dupdata. The actual data fields of the child
4556 * DB are all zero size.
4563 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4564 if (flags & MDB_CURRENT) {
4565 xflags = MDB_CURRENT;
4567 mdb_xcursor_init1(mc, leaf);
4568 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4570 /* converted, write the original data first */
4572 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4576 /* Adjust other cursors pointing to mp */
4578 unsigned i = mc->mc_top;
4579 MDB_page *mp = mc->mc_pg[i];
4581 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4582 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4583 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4584 mdb_xcursor_init1(m2, leaf);
4589 if (flags & MDB_APPENDDUP)
4590 xflags |= MDB_APPEND;
4591 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4592 if (flags & F_SUBDATA) {
4593 void *db = NODEDATA(leaf);
4594 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4597 /* sub-writes might have failed so check rc again.
4598 * Don't increment count if we just replaced an existing item.
4600 if (!rc && !(flags & MDB_CURRENT))
4601 mc->mc_db->md_entries++;
4602 if (flags & MDB_MULTIPLE) {
4604 if (mcount < data[1].mv_size) {
4605 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4606 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4616 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4621 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4624 if (!mc->mc_flags & C_INITIALIZED)
4627 rc = mdb_cursor_touch(mc);
4631 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4633 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4634 if (flags != MDB_NODUPDATA) {
4635 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4636 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4638 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4639 /* If sub-DB still has entries, we're done */
4640 if (mc->mc_xcursor->mx_db.md_entries) {
4641 if (leaf->mn_flags & F_SUBDATA) {
4642 /* update subDB info */
4643 void *db = NODEDATA(leaf);
4644 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4646 /* shrink fake page */
4647 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4649 mc->mc_db->md_entries--;
4652 /* otherwise fall thru and delete the sub-DB */
4655 if (leaf->mn_flags & F_SUBDATA) {
4656 /* add all the child DB's pages to the free list */
4657 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4658 if (rc == MDB_SUCCESS) {
4659 mc->mc_db->md_entries -=
4660 mc->mc_xcursor->mx_db.md_entries;
4665 return mdb_cursor_del0(mc, leaf);
4668 /** Allocate and initialize new pages for a database.
4669 * @param[in] mc a cursor on the database being added to.
4670 * @param[in] flags flags defining what type of page is being allocated.
4671 * @param[in] num the number of pages to allocate. This is usually 1,
4672 * unless allocating overflow pages for a large record.
4673 * @return Address of a page, or NULL on failure.
4676 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num)
4680 if ((np = mdb_page_alloc(mc, num)) == NULL)
4682 DPRINTF("allocated new mpage %zu, page size %u",
4683 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4684 np->mp_flags = flags | P_DIRTY;
4685 np->mp_lower = PAGEHDRSZ;
4686 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4689 mc->mc_db->md_branch_pages++;
4690 else if (IS_LEAF(np))
4691 mc->mc_db->md_leaf_pages++;
4692 else if (IS_OVERFLOW(np)) {
4693 mc->mc_db->md_overflow_pages += num;
4700 /** Calculate the size of a leaf node.
4701 * The size depends on the environment's page size; if a data item
4702 * is too large it will be put onto an overflow page and the node
4703 * size will only include the key and not the data. Sizes are always
4704 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4705 * of the #MDB_node headers.
4706 * @param[in] env The environment handle.
4707 * @param[in] key The key for the node.
4708 * @param[in] data The data for the node.
4709 * @return The number of bytes needed to store the node.
4712 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4716 sz = LEAFSIZE(key, data);
4717 if (sz >= env->me_psize / MDB_MINKEYS) {
4718 /* put on overflow page */
4719 sz -= data->mv_size - sizeof(pgno_t);
4723 return sz + sizeof(indx_t);
4726 /** Calculate the size of a branch node.
4727 * The size should depend on the environment's page size but since
4728 * we currently don't support spilling large keys onto overflow
4729 * pages, it's simply the size of the #MDB_node header plus the
4730 * size of the key. Sizes are always rounded up to an even number
4731 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4732 * @param[in] env The environment handle.
4733 * @param[in] key The key for the node.
4734 * @return The number of bytes needed to store the node.
4737 mdb_branch_size(MDB_env *env, MDB_val *key)
4742 if (sz >= env->me_psize / MDB_MINKEYS) {
4743 /* put on overflow page */
4744 /* not implemented */
4745 /* sz -= key->size - sizeof(pgno_t); */
4748 return sz + sizeof(indx_t);
4751 /** Add a node to the page pointed to by the cursor.
4752 * @param[in] mc The cursor for this operation.
4753 * @param[in] indx The index on the page where the new node should be added.
4754 * @param[in] key The key for the new node.
4755 * @param[in] data The data for the new node, if any.
4756 * @param[in] pgno The page number, if adding a branch node.
4757 * @param[in] flags Flags for the node.
4758 * @return 0 on success, non-zero on failure. Possible errors are:
4760 * <li>ENOMEM - failed to allocate overflow pages for the node.
4761 * <li>ENOSPC - there is insufficient room in the page. This error
4762 * should never happen since all callers already calculate the
4763 * page's free space before calling this function.
4767 mdb_node_add(MDB_cursor *mc, indx_t indx,
4768 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4771 size_t node_size = NODESIZE;
4774 MDB_page *mp = mc->mc_pg[mc->mc_top];
4775 MDB_page *ofp = NULL; /* overflow page */
4778 assert(mp->mp_upper >= mp->mp_lower);
4780 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4781 IS_LEAF(mp) ? "leaf" : "branch",
4782 IS_SUBP(mp) ? "sub-" : "",
4783 mp->mp_pgno, indx, data ? data->mv_size : 0,
4784 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4787 /* Move higher keys up one slot. */
4788 int ksize = mc->mc_db->md_pad, dif;
4789 char *ptr = LEAF2KEY(mp, indx, ksize);
4790 dif = NUMKEYS(mp) - indx;
4792 memmove(ptr+ksize, ptr, dif*ksize);
4793 /* insert new key */
4794 memcpy(ptr, key->mv_data, ksize);
4796 /* Just using these for counting */
4797 mp->mp_lower += sizeof(indx_t);
4798 mp->mp_upper -= ksize - sizeof(indx_t);
4803 node_size += key->mv_size;
4807 if (F_ISSET(flags, F_BIGDATA)) {
4808 /* Data already on overflow page. */
4809 node_size += sizeof(pgno_t);
4810 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4811 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4812 /* Put data on overflow page. */
4813 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4814 data->mv_size, node_size+data->mv_size);
4815 node_size += sizeof(pgno_t);
4816 if ((ofp = mdb_page_new(mc, P_OVERFLOW, ovpages)) == NULL)
4818 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4821 node_size += data->mv_size;
4824 node_size += node_size & 1;
4826 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4827 DPRINTF("not enough room in page %zu, got %u ptrs",
4828 mp->mp_pgno, NUMKEYS(mp));
4829 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4830 mp->mp_upper - mp->mp_lower);
4831 DPRINTF("node size = %zu", node_size);
4835 /* Move higher pointers up one slot. */
4836 for (i = NUMKEYS(mp); i > indx; i--)
4837 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
4839 /* Adjust free space offsets. */
4840 ofs = mp->mp_upper - node_size;
4841 assert(ofs >= mp->mp_lower + sizeof(indx_t));
4842 mp->mp_ptrs[indx] = ofs;
4844 mp->mp_lower += sizeof(indx_t);
4846 /* Write the node data. */
4847 node = NODEPTR(mp, indx);
4848 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
4849 node->mn_flags = flags;
4851 SETDSZ(node,data->mv_size);
4856 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4861 if (F_ISSET(flags, F_BIGDATA))
4862 memcpy(node->mn_data + key->mv_size, data->mv_data,
4864 else if (F_ISSET(flags, MDB_RESERVE))
4865 data->mv_data = node->mn_data + key->mv_size;
4867 memcpy(node->mn_data + key->mv_size, data->mv_data,
4870 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
4872 if (F_ISSET(flags, MDB_RESERVE))
4873 data->mv_data = METADATA(ofp);
4875 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
4882 /** Delete the specified node from a page.
4883 * @param[in] mp The page to operate on.
4884 * @param[in] indx The index of the node to delete.
4885 * @param[in] ksize The size of a node. Only used if the page is
4886 * part of a #MDB_DUPFIXED database.
4889 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
4892 indx_t i, j, numkeys, ptr;
4899 COPY_PGNO(pgno, mp->mp_pgno);
4900 DPRINTF("delete node %u on %s page %zu", indx,
4901 IS_LEAF(mp) ? "leaf" : "branch", pgno);
4904 assert(indx < NUMKEYS(mp));
4907 int x = NUMKEYS(mp) - 1 - indx;
4908 base = LEAF2KEY(mp, indx, ksize);
4910 memmove(base, base + ksize, x * ksize);
4911 mp->mp_lower -= sizeof(indx_t);
4912 mp->mp_upper += ksize - sizeof(indx_t);
4916 node = NODEPTR(mp, indx);
4917 sz = NODESIZE + node->mn_ksize;
4919 if (F_ISSET(node->mn_flags, F_BIGDATA))
4920 sz += sizeof(pgno_t);
4922 sz += NODEDSZ(node);
4926 ptr = mp->mp_ptrs[indx];
4927 numkeys = NUMKEYS(mp);
4928 for (i = j = 0; i < numkeys; i++) {
4930 mp->mp_ptrs[j] = mp->mp_ptrs[i];
4931 if (mp->mp_ptrs[i] < ptr)
4932 mp->mp_ptrs[j] += sz;
4937 base = (char *)mp + mp->mp_upper;
4938 memmove(base + sz, base, ptr - mp->mp_upper);
4940 mp->mp_lower -= sizeof(indx_t);
4944 /** Compact the main page after deleting a node on a subpage.
4945 * @param[in] mp The main page to operate on.
4946 * @param[in] indx The index of the subpage on the main page.
4949 mdb_node_shrink(MDB_page *mp, indx_t indx)
4956 indx_t i, numkeys, ptr;
4958 node = NODEPTR(mp, indx);
4959 sp = (MDB_page *)NODEDATA(node);
4960 osize = NODEDSZ(node);
4962 delta = sp->mp_upper - sp->mp_lower;
4963 SETDSZ(node, osize - delta);
4964 xp = (MDB_page *)((char *)sp + delta);
4966 /* shift subpage upward */
4968 nsize = NUMKEYS(sp) * sp->mp_pad;
4969 memmove(METADATA(xp), METADATA(sp), nsize);
4972 nsize = osize - sp->mp_upper;
4973 numkeys = NUMKEYS(sp);
4974 for (i=numkeys-1; i>=0; i--)
4975 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
4977 xp->mp_upper = sp->mp_lower;
4978 xp->mp_lower = sp->mp_lower;
4979 xp->mp_flags = sp->mp_flags;
4980 xp->mp_pad = sp->mp_pad;
4981 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
4983 /* shift lower nodes upward */
4984 ptr = mp->mp_ptrs[indx];
4985 numkeys = NUMKEYS(mp);
4986 for (i = 0; i < numkeys; i++) {
4987 if (mp->mp_ptrs[i] <= ptr)
4988 mp->mp_ptrs[i] += delta;
4991 base = (char *)mp + mp->mp_upper;
4992 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
4993 mp->mp_upper += delta;
4996 /** Initial setup of a sorted-dups cursor.
4997 * Sorted duplicates are implemented as a sub-database for the given key.
4998 * The duplicate data items are actually keys of the sub-database.
4999 * Operations on the duplicate data items are performed using a sub-cursor
5000 * initialized when the sub-database is first accessed. This function does
5001 * the preliminary setup of the sub-cursor, filling in the fields that
5002 * depend only on the parent DB.
5003 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5006 mdb_xcursor_init0(MDB_cursor *mc)
5008 MDB_xcursor *mx = mc->mc_xcursor;
5010 mx->mx_cursor.mc_xcursor = NULL;
5011 mx->mx_cursor.mc_txn = mc->mc_txn;
5012 mx->mx_cursor.mc_db = &mx->mx_db;
5013 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5014 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5015 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5016 mx->mx_cursor.mc_snum = 0;
5017 mx->mx_cursor.mc_top = 0;
5018 mx->mx_cursor.mc_flags = C_SUB;
5019 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5020 mx->mx_dbx.md_dcmp = NULL;
5021 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5024 /** Final setup of a sorted-dups cursor.
5025 * Sets up the fields that depend on the data from the main cursor.
5026 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5027 * @param[in] node The data containing the #MDB_db record for the
5028 * sorted-dup database.
5031 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5033 MDB_xcursor *mx = mc->mc_xcursor;
5035 if (node->mn_flags & F_SUBDATA) {
5036 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5037 mx->mx_cursor.mc_pg[0] = 0;
5038 mx->mx_cursor.mc_snum = 0;
5039 mx->mx_cursor.mc_flags = C_SUB;
5041 MDB_page *fp = NODEDATA(node);
5042 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5043 mx->mx_db.md_flags = 0;
5044 mx->mx_db.md_depth = 1;
5045 mx->mx_db.md_branch_pages = 0;
5046 mx->mx_db.md_leaf_pages = 1;
5047 mx->mx_db.md_overflow_pages = 0;
5048 mx->mx_db.md_entries = NUMKEYS(fp);
5049 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5050 mx->mx_cursor.mc_snum = 1;
5051 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5052 mx->mx_cursor.mc_top = 0;
5053 mx->mx_cursor.mc_pg[0] = fp;
5054 mx->mx_cursor.mc_ki[0] = 0;
5055 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5056 mx->mx_db.md_flags = MDB_DUPFIXED;
5057 mx->mx_db.md_pad = fp->mp_pad;
5058 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5059 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5062 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5064 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5066 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5067 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5068 #if UINT_MAX < SIZE_MAX
5069 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5070 #ifdef MISALIGNED_OK
5071 mx->mx_dbx.md_cmp = mdb_cmp_long;
5073 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5078 /** Initialize a cursor for a given transaction and database. */
5080 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5085 mc->mc_db = &txn->mt_dbs[dbi];
5086 mc->mc_dbx = &txn->mt_dbxs[dbi];
5087 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5092 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5094 mc->mc_xcursor = mx;
5095 mdb_xcursor_init0(mc);
5097 mc->mc_xcursor = NULL;
5099 if (*mc->mc_dbflag & DB_STALE) {
5100 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5105 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5108 MDB_xcursor *mx = NULL;
5109 size_t size = sizeof(MDB_cursor);
5111 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5114 /* Allow read access to the freelist */
5115 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5118 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5119 size += sizeof(MDB_xcursor);
5121 if ((mc = malloc(size)) != NULL) {
5122 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5123 mx = (MDB_xcursor *)(mc + 1);
5125 mdb_cursor_init(mc, txn, dbi, mx);
5126 if (txn->mt_cursors) {
5127 mc->mc_next = txn->mt_cursors[dbi];
5128 txn->mt_cursors[dbi] = mc;
5130 mc->mc_flags |= C_ALLOCD;
5140 /* Return the count of duplicate data items for the current key */
5142 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5146 if (mc == NULL || countp == NULL)
5149 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5152 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5153 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5156 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5159 *countp = mc->mc_xcursor->mx_db.md_entries;
5165 mdb_cursor_close(MDB_cursor *mc)
5168 /* remove from txn, if tracked */
5169 if (mc->mc_txn->mt_cursors) {
5170 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5171 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5173 *prev = mc->mc_next;
5175 if (mc->mc_flags & C_ALLOCD)
5181 mdb_cursor_txn(MDB_cursor *mc)
5183 if (!mc) return NULL;
5188 mdb_cursor_dbi(MDB_cursor *mc)
5194 /** Replace the key for a node with a new key.
5195 * @param[in] mp The page containing the node to operate on.
5196 * @param[in] indx The index of the node to operate on.
5197 * @param[in] key The new key to use.
5198 * @return 0 on success, non-zero on failure.
5201 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5207 indx_t ptr, i, numkeys;
5210 node = NODEPTR(mp, indx);
5211 ptr = mp->mp_ptrs[indx];
5215 char kbuf2[(MAXKEYSIZE*2+1)];
5216 k2.mv_data = NODEKEY(node);
5217 k2.mv_size = node->mn_ksize;
5218 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5220 mdb_dkey(&k2, kbuf2),
5226 delta0 = delta = key->mv_size - node->mn_ksize;
5228 /* Must be 2-byte aligned. If new key is
5229 * shorter by 1, the shift will be skipped.
5231 delta += (delta & 1);
5233 if (delta > 0 && SIZELEFT(mp) < delta) {
5234 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5238 numkeys = NUMKEYS(mp);
5239 for (i = 0; i < numkeys; i++) {
5240 if (mp->mp_ptrs[i] <= ptr)
5241 mp->mp_ptrs[i] -= delta;
5244 base = (char *)mp + mp->mp_upper;
5245 len = ptr - mp->mp_upper + NODESIZE;
5246 memmove(base - delta, base, len);
5247 mp->mp_upper -= delta;
5249 node = NODEPTR(mp, indx);
5252 /* But even if no shift was needed, update ksize */
5254 node->mn_ksize = key->mv_size;
5257 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5262 /** Move a node from csrc to cdst.
5265 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5271 unsigned short flags;
5275 /* Mark src and dst as dirty. */
5276 if ((rc = mdb_page_touch(csrc)) ||
5277 (rc = mdb_page_touch(cdst)))
5280 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5281 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5282 key.mv_size = csrc->mc_db->md_pad;
5283 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5285 data.mv_data = NULL;
5289 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5290 assert(!((long)srcnode&1));
5291 srcpg = NODEPGNO(srcnode);
5292 flags = srcnode->mn_flags;
5293 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5294 unsigned int snum = csrc->mc_snum;
5296 /* must find the lowest key below src */
5297 mdb_page_search_root(csrc, NULL, 0);
5298 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5299 key.mv_size = csrc->mc_db->md_pad;
5300 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5302 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5303 key.mv_size = NODEKSZ(s2);
5304 key.mv_data = NODEKEY(s2);
5306 csrc->mc_snum = snum--;
5307 csrc->mc_top = snum;
5309 key.mv_size = NODEKSZ(srcnode);
5310 key.mv_data = NODEKEY(srcnode);
5312 data.mv_size = NODEDSZ(srcnode);
5313 data.mv_data = NODEDATA(srcnode);
5315 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5316 unsigned int snum = cdst->mc_snum;
5319 /* must find the lowest key below dst */
5320 mdb_page_search_root(cdst, NULL, 0);
5321 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5322 bkey.mv_size = cdst->mc_db->md_pad;
5323 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5325 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5326 bkey.mv_size = NODEKSZ(s2);
5327 bkey.mv_data = NODEKEY(s2);
5329 cdst->mc_snum = snum--;
5330 cdst->mc_top = snum;
5331 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5334 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5335 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5336 csrc->mc_ki[csrc->mc_top],
5338 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5339 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5341 /* Add the node to the destination page.
5343 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5344 if (rc != MDB_SUCCESS)
5347 /* Delete the node from the source page.
5349 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5352 /* Adjust other cursors pointing to mp */
5353 MDB_cursor *m2, *m3;
5354 MDB_dbi dbi = csrc->mc_dbi;
5355 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5357 if (csrc->mc_flags & C_SUB)
5360 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5361 if (m2 == csrc) continue;
5362 if (csrc->mc_flags & C_SUB)
5363 m3 = &m2->mc_xcursor->mx_cursor;
5366 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5367 csrc->mc_ki[csrc->mc_top]) {
5368 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5369 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5374 /* Update the parent separators.
5376 if (csrc->mc_ki[csrc->mc_top] == 0) {
5377 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5378 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5379 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5381 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5382 key.mv_size = NODEKSZ(srcnode);
5383 key.mv_data = NODEKEY(srcnode);
5385 DPRINTF("update separator for source page %zu to [%s]",
5386 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5387 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5388 &key)) != MDB_SUCCESS)
5391 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5393 nullkey.mv_size = 0;
5394 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5395 assert(rc == MDB_SUCCESS);
5399 if (cdst->mc_ki[cdst->mc_top] == 0) {
5400 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5401 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5402 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5404 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5405 key.mv_size = NODEKSZ(srcnode);
5406 key.mv_data = NODEKEY(srcnode);
5408 DPRINTF("update separator for destination page %zu to [%s]",
5409 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5410 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5411 &key)) != MDB_SUCCESS)
5414 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5416 nullkey.mv_size = 0;
5417 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5418 assert(rc == MDB_SUCCESS);
5425 /** Merge one page into another.
5426 * The nodes from the page pointed to by \b csrc will
5427 * be copied to the page pointed to by \b cdst and then
5428 * the \b csrc page will be freed.
5429 * @param[in] csrc Cursor pointing to the source page.
5430 * @param[in] cdst Cursor pointing to the destination page.
5433 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5441 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5442 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5444 assert(csrc->mc_snum > 1); /* can't merge root page */
5445 assert(cdst->mc_snum > 1);
5447 /* Mark dst as dirty. */
5448 if ((rc = mdb_page_touch(cdst)))
5451 /* Move all nodes from src to dst.
5453 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5454 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5455 key.mv_size = csrc->mc_db->md_pad;
5456 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5457 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5458 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5459 if (rc != MDB_SUCCESS)
5461 key.mv_data = (char *)key.mv_data + key.mv_size;
5464 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5465 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5466 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5467 unsigned int snum = csrc->mc_snum;
5469 /* must find the lowest key below src */
5470 mdb_page_search_root(csrc, NULL, 0);
5471 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5472 key.mv_size = csrc->mc_db->md_pad;
5473 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5475 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5476 key.mv_size = NODEKSZ(s2);
5477 key.mv_data = NODEKEY(s2);
5479 csrc->mc_snum = snum--;
5480 csrc->mc_top = snum;
5482 key.mv_size = srcnode->mn_ksize;
5483 key.mv_data = NODEKEY(srcnode);
5486 data.mv_size = NODEDSZ(srcnode);
5487 data.mv_data = NODEDATA(srcnode);
5488 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5489 if (rc != MDB_SUCCESS)
5494 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5495 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);
5497 /* Unlink the src page from parent and add to free list.
5499 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5500 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5502 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5506 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5507 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5508 csrc->mc_db->md_leaf_pages--;
5510 csrc->mc_db->md_branch_pages--;
5512 /* Adjust other cursors pointing to mp */
5513 MDB_cursor *m2, *m3;
5514 MDB_dbi dbi = csrc->mc_dbi;
5515 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5517 if (csrc->mc_flags & C_SUB)
5520 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5521 if (csrc->mc_flags & C_SUB)
5522 m3 = &m2->mc_xcursor->mx_cursor;
5525 if (m3 == csrc) continue;
5526 if (m3->mc_snum < csrc->mc_snum) continue;
5527 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5528 m3->mc_pg[csrc->mc_top] = mp;
5529 m3->mc_ki[csrc->mc_top] += nkeys;
5533 mdb_cursor_pop(csrc);
5535 return mdb_rebalance(csrc);
5538 /** Copy the contents of a cursor.
5539 * @param[in] csrc The cursor to copy from.
5540 * @param[out] cdst The cursor to copy to.
5543 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5547 cdst->mc_txn = csrc->mc_txn;
5548 cdst->mc_dbi = csrc->mc_dbi;
5549 cdst->mc_db = csrc->mc_db;
5550 cdst->mc_dbx = csrc->mc_dbx;
5551 cdst->mc_snum = csrc->mc_snum;
5552 cdst->mc_top = csrc->mc_top;
5553 cdst->mc_flags = csrc->mc_flags;
5555 for (i=0; i<csrc->mc_snum; i++) {
5556 cdst->mc_pg[i] = csrc->mc_pg[i];
5557 cdst->mc_ki[i] = csrc->mc_ki[i];
5561 /** Rebalance the tree after a delete operation.
5562 * @param[in] mc Cursor pointing to the page where rebalancing
5564 * @return 0 on success, non-zero on failure.
5567 mdb_rebalance(MDB_cursor *mc)
5577 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5578 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5579 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5580 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5584 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5587 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5588 DPRINTF("no need to rebalance page %zu, above fill threshold",
5594 if (mc->mc_snum < 2) {
5595 MDB_page *mp = mc->mc_pg[0];
5596 if (NUMKEYS(mp) == 0) {
5597 DPUTS("tree is completely empty");
5598 mc->mc_db->md_root = P_INVALID;
5599 mc->mc_db->md_depth = 0;
5600 mc->mc_db->md_leaf_pages = 0;
5601 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5605 /* Adjust other cursors pointing to mp */
5606 MDB_cursor *m2, *m3;
5607 MDB_dbi dbi = mc->mc_dbi;
5609 if (mc->mc_flags & C_SUB)
5612 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5613 if (m2 == mc) continue;
5614 if (mc->mc_flags & C_SUB)
5615 m3 = &m2->mc_xcursor->mx_cursor;
5618 if (m3->mc_snum < mc->mc_snum) continue;
5619 if (m3->mc_pg[0] == mp) {
5625 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5626 DPUTS("collapsing root page!");
5627 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5628 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5629 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5632 mc->mc_db->md_depth--;
5633 mc->mc_db->md_branch_pages--;
5635 /* Adjust other cursors pointing to mp */
5636 MDB_cursor *m2, *m3;
5637 MDB_dbi dbi = mc->mc_dbi;
5639 if (mc->mc_flags & C_SUB)
5642 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5643 if (m2 == mc) continue;
5644 if (mc->mc_flags & C_SUB)
5645 m3 = &m2->mc_xcursor->mx_cursor;
5648 if (m3->mc_snum < mc->mc_snum) continue;
5649 if (m3->mc_pg[0] == mp) {
5650 m3->mc_pg[0] = mc->mc_pg[0];
5655 DPUTS("root page doesn't need rebalancing");
5659 /* The parent (branch page) must have at least 2 pointers,
5660 * otherwise the tree is invalid.
5662 ptop = mc->mc_top-1;
5663 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5665 /* Leaf page fill factor is below the threshold.
5666 * Try to move keys from left or right neighbor, or
5667 * merge with a neighbor page.
5672 mdb_cursor_copy(mc, &mn);
5673 mn.mc_xcursor = NULL;
5675 if (mc->mc_ki[ptop] == 0) {
5676 /* We're the leftmost leaf in our parent.
5678 DPUTS("reading right neighbor");
5680 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5681 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5683 mn.mc_ki[mn.mc_top] = 0;
5684 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5686 /* There is at least one neighbor to the left.
5688 DPUTS("reading left neighbor");
5690 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5691 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5693 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5694 mc->mc_ki[mc->mc_top] = 0;
5697 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5698 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);
5700 /* If the neighbor page is above threshold and has at least two
5701 * keys, move one key from it.
5703 * Otherwise we should try to merge them.
5705 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5706 return mdb_node_move(&mn, mc);
5707 else { /* FIXME: if (has_enough_room()) */
5708 mc->mc_flags &= ~C_INITIALIZED;
5709 if (mc->mc_ki[ptop] == 0)
5710 return mdb_page_merge(&mn, mc);
5712 return mdb_page_merge(mc, &mn);
5716 /** Complete a delete operation started by #mdb_cursor_del(). */
5718 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5722 /* add overflow pages to free list */
5723 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5727 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5728 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5729 mc->mc_db->md_overflow_pages -= ovpages;
5730 for (i=0; i<ovpages; i++) {
5731 DPRINTF("freed ov page %zu", pg);
5732 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5736 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5737 mc->mc_db->md_entries--;
5738 rc = mdb_rebalance(mc);
5739 if (rc != MDB_SUCCESS)
5740 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5746 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5747 MDB_val *key, MDB_val *data)
5752 MDB_val rdata, *xdata;
5756 assert(key != NULL);
5758 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5760 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5763 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5767 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5771 mdb_cursor_init(&mc, txn, dbi, &mx);
5782 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5784 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5788 /** Split a page and insert a new node.
5789 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5790 * The cursor will be updated to point to the actual page and index where
5791 * the node got inserted after the split.
5792 * @param[in] newkey The key for the newly inserted node.
5793 * @param[in] newdata The data for the newly inserted node.
5794 * @param[in] newpgno The page number, if the new node is a branch node.
5795 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
5796 * @return 0 on success, non-zero on failure.
5799 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5800 unsigned int nflags)
5803 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
5806 unsigned int i, j, split_indx, nkeys, pmax;
5808 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5810 MDB_page *mp, *rp, *pp;
5815 mp = mc->mc_pg[mc->mc_top];
5816 newindx = mc->mc_ki[mc->mc_top];
5818 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
5819 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
5820 DKEY(newkey), mc->mc_ki[mc->mc_top]);
5822 /* Create a right sibling. */
5823 if ((rp = mdb_page_new(mc, mp->mp_flags, 1)) == NULL)
5825 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
5827 if (mc->mc_snum < 2) {
5828 if ((pp = mdb_page_new(mc, P_BRANCH, 1)) == NULL)
5830 /* shift current top to make room for new parent */
5831 mc->mc_pg[1] = mc->mc_pg[0];
5832 mc->mc_ki[1] = mc->mc_ki[0];
5835 mc->mc_db->md_root = pp->mp_pgno;
5836 DPRINTF("root split! new root = %zu", pp->mp_pgno);
5837 mc->mc_db->md_depth++;
5840 /* Add left (implicit) pointer. */
5841 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
5842 /* undo the pre-push */
5843 mc->mc_pg[0] = mc->mc_pg[1];
5844 mc->mc_ki[0] = mc->mc_ki[1];
5845 mc->mc_db->md_root = mp->mp_pgno;
5846 mc->mc_db->md_depth--;
5853 ptop = mc->mc_top-1;
5854 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
5857 mc->mc_flags |= C_SPLITTING;
5858 mdb_cursor_copy(mc, &mn);
5859 mn.mc_pg[mn.mc_top] = rp;
5860 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
5862 if (nflags & MDB_APPEND) {
5863 mn.mc_ki[mn.mc_top] = 0;
5865 split_indx = newindx;
5870 nkeys = NUMKEYS(mp);
5871 split_indx = (nkeys + 1) / 2;
5872 if (newindx < split_indx)
5878 unsigned int lsize, rsize, ksize;
5879 /* Move half of the keys to the right sibling */
5881 x = mc->mc_ki[mc->mc_top] - split_indx;
5882 ksize = mc->mc_db->md_pad;
5883 split = LEAF2KEY(mp, split_indx, ksize);
5884 rsize = (nkeys - split_indx) * ksize;
5885 lsize = (nkeys - split_indx) * sizeof(indx_t);
5886 mp->mp_lower -= lsize;
5887 rp->mp_lower += lsize;
5888 mp->mp_upper += rsize - lsize;
5889 rp->mp_upper -= rsize - lsize;
5890 sepkey.mv_size = ksize;
5891 if (newindx == split_indx) {
5892 sepkey.mv_data = newkey->mv_data;
5894 sepkey.mv_data = split;
5897 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
5898 memcpy(rp->mp_ptrs, split, rsize);
5899 sepkey.mv_data = rp->mp_ptrs;
5900 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
5901 memcpy(ins, newkey->mv_data, ksize);
5902 mp->mp_lower += sizeof(indx_t);
5903 mp->mp_upper -= ksize - sizeof(indx_t);
5906 memcpy(rp->mp_ptrs, split, x * ksize);
5907 ins = LEAF2KEY(rp, x, ksize);
5908 memcpy(ins, newkey->mv_data, ksize);
5909 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
5910 rp->mp_lower += sizeof(indx_t);
5911 rp->mp_upper -= ksize - sizeof(indx_t);
5912 mc->mc_ki[mc->mc_top] = x;
5913 mc->mc_pg[mc->mc_top] = rp;
5918 /* For leaf pages, check the split point based on what
5919 * fits where, since otherwise mdb_node_add can fail.
5921 * This check is only needed when the data items are
5922 * relatively large, such that being off by one will
5923 * make the difference between success or failure.
5924 * When the size of the data items is much smaller than
5925 * one-half of a page, this check is irrelevant.
5928 unsigned int psize, nsize;
5929 /* Maximum free space in an empty page */
5930 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
5931 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
5932 if ((nkeys < 20) || (nsize > pmax/4)) {
5933 if (newindx <= split_indx) {
5936 for (i=0; i<split_indx; i++) {
5937 node = NODEPTR(mp, i);
5938 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5939 if (F_ISSET(node->mn_flags, F_BIGDATA))
5940 psize += sizeof(pgno_t);
5942 psize += NODEDSZ(node);
5945 if (i == split_indx - 1 && newindx == split_indx)
5954 for (i=nkeys-1; i>=split_indx; i--) {
5955 node = NODEPTR(mp, i);
5956 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5957 if (F_ISSET(node->mn_flags, F_BIGDATA))
5958 psize += sizeof(pgno_t);
5960 psize += NODEDSZ(node);
5971 /* First find the separating key between the split pages.
5972 * The case where newindx == split_indx is ambiguous; the
5973 * new item could go to the new page or stay on the original
5974 * page. If newpos == 1 it goes to the new page.
5976 if (newindx == split_indx && newpos) {
5977 sepkey.mv_size = newkey->mv_size;
5978 sepkey.mv_data = newkey->mv_data;
5980 node = NODEPTR(mp, split_indx);
5981 sepkey.mv_size = node->mn_ksize;
5982 sepkey.mv_data = NODEKEY(node);
5986 DPRINTF("separator is [%s]", DKEY(&sepkey));
5988 /* Copy separator key to the parent.
5990 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
5994 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
5997 if (mn.mc_snum == mc->mc_snum) {
5998 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
5999 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6000 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6001 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6006 /* Right page might now have changed parent.
6007 * Check if left page also changed parent.
6009 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6010 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6011 for (i=0; i<ptop; i++) {
6012 mc->mc_pg[i] = mn.mc_pg[i];
6013 mc->mc_ki[i] = mn.mc_ki[i];
6015 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6016 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6020 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6023 mc->mc_flags ^= C_SPLITTING;
6024 if (rc != MDB_SUCCESS) {
6027 if (nflags & MDB_APPEND) {
6028 mc->mc_pg[mc->mc_top] = rp;
6029 mc->mc_ki[mc->mc_top] = 0;
6030 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6033 for (i=0; i<mc->mc_top; i++)
6034 mc->mc_ki[i] = mn.mc_ki[i];
6041 /* Move half of the keys to the right sibling. */
6043 /* grab a page to hold a temporary copy */
6044 copy = mdb_page_malloc(mc);
6048 copy->mp_pgno = mp->mp_pgno;
6049 copy->mp_flags = mp->mp_flags;
6050 copy->mp_lower = PAGEHDRSZ;
6051 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6052 mc->mc_pg[mc->mc_top] = copy;
6053 for (i = j = 0; i <= nkeys; j++) {
6054 if (i == split_indx) {
6055 /* Insert in right sibling. */
6056 /* Reset insert index for right sibling. */
6057 if (i != newindx || (newpos ^ ins_new)) {
6059 mc->mc_pg[mc->mc_top] = rp;
6063 if (i == newindx && !ins_new) {
6064 /* Insert the original entry that caused the split. */
6065 rkey.mv_data = newkey->mv_data;
6066 rkey.mv_size = newkey->mv_size;
6075 /* Update index for the new key. */
6076 mc->mc_ki[mc->mc_top] = j;
6077 } else if (i == nkeys) {
6080 node = NODEPTR(mp, i);
6081 rkey.mv_data = NODEKEY(node);
6082 rkey.mv_size = node->mn_ksize;
6084 xdata.mv_data = NODEDATA(node);
6085 xdata.mv_size = NODEDSZ(node);
6088 pgno = NODEPGNO(node);
6089 flags = node->mn_flags;
6094 if (!IS_LEAF(mp) && j == 0) {
6095 /* First branch index doesn't need key data. */
6099 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6103 nkeys = NUMKEYS(copy);
6104 for (i=0; i<nkeys; i++)
6105 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6106 mp->mp_lower = copy->mp_lower;
6107 mp->mp_upper = copy->mp_upper;
6108 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6109 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6111 /* reset back to original page */
6112 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6113 mc->mc_pg[mc->mc_top] = mp;
6114 if (nflags & MDB_RESERVE) {
6115 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6116 if (!(node->mn_flags & F_BIGDATA))
6117 newdata->mv_data = NODEDATA(node);
6123 /* return tmp page to freelist */
6124 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
6125 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
6126 mc->mc_txn->mt_env->me_dpages = copy;
6129 /* Adjust other cursors pointing to mp */
6130 MDB_cursor *m2, *m3;
6131 MDB_dbi dbi = mc->mc_dbi;
6132 int fixup = NUMKEYS(mp);
6134 if (mc->mc_flags & C_SUB)
6137 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6138 if (m2 == mc) continue;
6139 if (mc->mc_flags & C_SUB)
6140 m3 = &m2->mc_xcursor->mx_cursor;
6143 if (!(m3->mc_flags & C_INITIALIZED))
6145 if (m3->mc_flags & C_SPLITTING)
6150 for (k=m3->mc_top; k>=0; k--) {
6151 m3->mc_ki[k+1] = m3->mc_ki[k];
6152 m3->mc_pg[k+1] = m3->mc_pg[k];
6154 if (m3->mc_ki[0] >= split_indx) {
6159 m3->mc_pg[0] = mc->mc_pg[0];
6163 if (m3->mc_pg[mc->mc_top] == mp) {
6164 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6165 m3->mc_ki[mc->mc_top]++;
6166 if (m3->mc_ki[mc->mc_top] >= fixup) {
6167 m3->mc_pg[mc->mc_top] = rp;
6168 m3->mc_ki[mc->mc_top] -= fixup;
6169 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6171 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6172 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6181 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6182 MDB_val *key, MDB_val *data, unsigned int flags)
6187 assert(key != NULL);
6188 assert(data != NULL);
6190 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6193 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6197 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6201 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6204 mdb_cursor_init(&mc, txn, dbi, &mx);
6205 return mdb_cursor_put(&mc, key, data, flags);
6208 /** Only a subset of the @ref mdb_env flags can be changed
6209 * at runtime. Changing other flags requires closing the environment
6210 * and re-opening it with the new flags.
6212 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC)
6214 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6216 if ((flag & CHANGEABLE) != flag)
6219 env->me_flags |= flag;
6221 env->me_flags &= ~flag;
6226 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6231 *arg = env->me_flags;
6236 mdb_env_get_path(MDB_env *env, const char **arg)
6241 *arg = env->me_path;
6245 /** Common code for #mdb_stat() and #mdb_env_stat().
6246 * @param[in] env the environment to operate in.
6247 * @param[in] db the #MDB_db record containing the stats to return.
6248 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6249 * @return 0, this function always succeeds.
6252 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6254 arg->ms_psize = env->me_psize;
6255 arg->ms_depth = db->md_depth;
6256 arg->ms_branch_pages = db->md_branch_pages;
6257 arg->ms_leaf_pages = db->md_leaf_pages;
6258 arg->ms_overflow_pages = db->md_overflow_pages;
6259 arg->ms_entries = db->md_entries;
6264 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6268 if (env == NULL || arg == NULL)
6271 toggle = mdb_env_pick_meta(env);
6273 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6276 /** Set the default comparison functions for a database.
6277 * Called immediately after a database is opened to set the defaults.
6278 * The user can then override them with #mdb_set_compare() or
6279 * #mdb_set_dupsort().
6280 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6281 * @param[in] dbi A database handle returned by #mdb_open()
6284 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6286 uint16_t f = txn->mt_dbs[dbi].md_flags;
6288 txn->mt_dbxs[dbi].md_cmp =
6289 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6290 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6292 txn->mt_dbxs[dbi].md_dcmp =
6293 !(f & MDB_DUPSORT) ? 0 :
6294 ((f & MDB_INTEGERDUP)
6295 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6296 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6299 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6304 int rc, dbflag, exact;
6305 unsigned int unused = 0;
6308 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6309 mdb_default_cmp(txn, FREE_DBI);
6315 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
6316 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
6317 mdb_default_cmp(txn, MAIN_DBI);
6321 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6322 mdb_default_cmp(txn, MAIN_DBI);
6325 /* Is the DB already open? */
6327 for (i=2; i<txn->mt_numdbs; i++) {
6328 if (!txn->mt_dbxs[i].md_name.mv_size) {
6329 /* Remember this free slot */
6330 if (!unused) unused = i;
6333 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6334 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6340 /* If no free slot and max hit, fail */
6341 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
6344 /* Find the DB info */
6348 key.mv_data = (void *)name;
6349 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6350 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6351 if (rc == MDB_SUCCESS) {
6352 /* make sure this is actually a DB */
6353 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6354 if (!(node->mn_flags & F_SUBDATA))
6356 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6357 /* Create if requested */
6359 data.mv_size = sizeof(MDB_db);
6360 data.mv_data = &dummy;
6361 memset(&dummy, 0, sizeof(dummy));
6362 dummy.md_root = P_INVALID;
6363 dummy.md_flags = flags & 0xffff;
6364 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6368 /* OK, got info, add to table */
6369 if (rc == MDB_SUCCESS) {
6370 unsigned int slot = unused ? unused : txn->mt_numdbs;
6371 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
6372 txn->mt_dbxs[slot].md_name.mv_size = len;
6373 txn->mt_dbxs[slot].md_rel = NULL;
6374 txn->mt_dbflags[slot] = dbflag;
6375 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
6377 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
6378 mdb_default_cmp(txn, slot);
6381 txn->mt_env->me_numdbs++;
6388 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6390 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6393 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6396 void mdb_close(MDB_env *env, MDB_dbi dbi)
6399 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6401 ptr = env->me_dbxs[dbi].md_name.mv_data;
6402 env->me_dbxs[dbi].md_name.mv_data = NULL;
6403 env->me_dbxs[dbi].md_name.mv_size = 0;
6407 /** Add all the DB's pages to the free list.
6408 * @param[in] mc Cursor on the DB to free.
6409 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6410 * @return 0 on success, non-zero on failure.
6413 mdb_drop0(MDB_cursor *mc, int subs)
6417 rc = mdb_page_search(mc, NULL, 0);
6418 if (rc == MDB_SUCCESS) {
6423 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6424 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6427 mdb_cursor_copy(mc, &mx);
6428 while (mc->mc_snum > 0) {
6429 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6430 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6431 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6432 if (ni->mn_flags & F_SUBDATA) {
6433 mdb_xcursor_init1(mc, ni);
6434 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6440 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6442 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6445 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6450 rc = mdb_cursor_sibling(mc, 1);
6452 /* no more siblings, go back to beginning
6453 * of previous level. (stack was already popped
6454 * by mdb_cursor_sibling)
6456 for (i=1; i<mc->mc_top; i++)
6457 mc->mc_pg[i] = mx.mc_pg[i];
6461 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6462 mc->mc_db->md_root);
6467 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6472 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6475 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6478 rc = mdb_cursor_open(txn, dbi, &mc);
6482 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6486 /* Can't delete the main DB */
6487 if (del && dbi > MAIN_DBI) {
6488 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6490 mdb_close(txn->mt_env, dbi);
6492 txn->mt_dbflags[dbi] |= DB_DIRTY;
6493 txn->mt_dbs[dbi].md_depth = 0;
6494 txn->mt_dbs[dbi].md_branch_pages = 0;
6495 txn->mt_dbs[dbi].md_leaf_pages = 0;
6496 txn->mt_dbs[dbi].md_overflow_pages = 0;
6497 txn->mt_dbs[dbi].md_entries = 0;
6498 txn->mt_dbs[dbi].md_root = P_INVALID;
6501 mdb_cursor_close(mc);
6505 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6507 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6510 txn->mt_dbxs[dbi].md_cmp = cmp;
6514 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6516 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6519 txn->mt_dbxs[dbi].md_dcmp = cmp;
6523 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6525 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6528 txn->mt_dbxs[dbi].md_rel = rel;
6532 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6534 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6537 txn->mt_dbxs[dbi].md_relctx = ctx;