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 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1332 * @param[in] mc cursor pointing to the page to be touched
1333 * @return 0 on success, non-zero on failure.
1336 mdb_page_touch(MDB_cursor *mc)
1338 MDB_page *mp = mc->mc_pg[mc->mc_top];
1341 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1343 if ((np = mdb_page_alloc(mc, 1)) == NULL)
1345 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1346 assert(mp->mp_pgno != np->mp_pgno);
1347 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1349 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1352 mp->mp_flags |= P_DIRTY;
1355 /* Adjust other cursors pointing to mp */
1356 if (mc->mc_flags & C_SUB) {
1357 MDB_cursor *m2, *m3;
1358 MDB_dbi dbi = mc->mc_dbi-1;
1360 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1361 if (m2 == mc) continue;
1362 m3 = &m2->mc_xcursor->mx_cursor;
1363 if (m3->mc_snum < mc->mc_snum) continue;
1364 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1365 m3->mc_pg[mc->mc_top] = mp;
1371 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1372 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1373 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1374 m2->mc_pg[mc->mc_top] = mp;
1378 mc->mc_pg[mc->mc_top] = mp;
1379 /** If this page has a parent, update the parent to point to
1383 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1385 mc->mc_db->md_root = mp->mp_pgno;
1386 } else if (mc->mc_txn->mt_parent) {
1389 /* If txn has a parent, make sure the page is in our
1392 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1393 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1394 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1395 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1396 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1397 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1398 mc->mc_pg[mc->mc_top] = mp;
1404 np = mdb_page_malloc(mc);
1405 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1406 mid.mid = np->mp_pgno;
1408 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1416 mdb_env_sync(MDB_env *env, int force)
1419 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1420 if (MDB_FDATASYNC(env->me_fd))
1426 /** Make shadow copies of all of parent txn's cursors */
1428 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1430 MDB_cursor *mc, *m2;
1431 unsigned int i, j, size;
1433 for (i=0;i<src->mt_numdbs; i++) {
1434 if (src->mt_cursors[i]) {
1435 size = sizeof(MDB_cursor);
1436 if (src->mt_cursors[i]->mc_xcursor)
1437 size += sizeof(MDB_xcursor);
1438 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1445 mc->mc_db = &dst->mt_dbs[i];
1446 mc->mc_dbx = m2->mc_dbx;
1447 mc->mc_dbflag = &dst->mt_dbflags[i];
1448 mc->mc_snum = m2->mc_snum;
1449 mc->mc_top = m2->mc_top;
1450 mc->mc_flags = m2->mc_flags | C_SHADOW;
1451 for (j=0; j<mc->mc_snum; j++) {
1452 mc->mc_pg[j] = m2->mc_pg[j];
1453 mc->mc_ki[j] = m2->mc_ki[j];
1455 if (m2->mc_xcursor) {
1456 MDB_xcursor *mx, *mx2;
1457 mx = (MDB_xcursor *)(mc+1);
1458 mc->mc_xcursor = mx;
1459 mx2 = m2->mc_xcursor;
1460 mx->mx_db = mx2->mx_db;
1461 mx->mx_dbx = mx2->mx_dbx;
1462 mx->mx_dbflag = mx2->mx_dbflag;
1463 mx->mx_cursor.mc_txn = dst;
1464 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1465 mx->mx_cursor.mc_db = &mx->mx_db;
1466 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1467 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1468 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1469 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1470 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1471 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1472 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1473 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1476 mc->mc_xcursor = NULL;
1478 mc->mc_next = dst->mt_cursors[i];
1479 dst->mt_cursors[i] = mc;
1486 /** Merge shadow cursors back into parent's */
1488 mdb_cursor_merge(MDB_txn *txn)
1491 for (i=0; i<txn->mt_numdbs; i++) {
1492 if (txn->mt_cursors[i]) {
1494 while ((mc = txn->mt_cursors[i])) {
1495 txn->mt_cursors[i] = mc->mc_next;
1496 if (mc->mc_flags & C_SHADOW) {
1497 MDB_cursor *m2 = mc->mc_orig;
1499 m2->mc_snum = mc->mc_snum;
1500 m2->mc_top = mc->mc_top;
1501 for (j=0; j<mc->mc_snum; j++) {
1502 m2->mc_pg[j] = mc->mc_pg[j];
1503 m2->mc_ki[j] = mc->mc_ki[j];
1506 if (mc->mc_flags & C_ALLOCD)
1514 mdb_txn_reset0(MDB_txn *txn);
1516 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1517 * @param[in] txn the transaction handle to initialize
1518 * @return 0 on success, non-zero on failure. This can only
1519 * fail for read-only transactions, and then only if the
1520 * reader table is full.
1523 mdb_txn_renew0(MDB_txn *txn)
1525 MDB_env *env = txn->mt_env;
1529 txn->mt_numdbs = env->me_numdbs;
1530 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1532 if (txn->mt_flags & MDB_TXN_RDONLY) {
1533 MDB_reader *r = pthread_getspecific(env->me_txkey);
1535 pid_t pid = getpid();
1536 pthread_t tid = pthread_self();
1539 for (i=0; i<env->me_txns->mti_numreaders; i++)
1540 if (env->me_txns->mti_readers[i].mr_pid == 0)
1542 if (i == env->me_maxreaders) {
1543 UNLOCK_MUTEX_R(env);
1546 env->me_txns->mti_readers[i].mr_pid = pid;
1547 env->me_txns->mti_readers[i].mr_tid = tid;
1548 if (i >= env->me_txns->mti_numreaders)
1549 env->me_txns->mti_numreaders = i+1;
1550 UNLOCK_MUTEX_R(env);
1551 r = &env->me_txns->mti_readers[i];
1552 pthread_setspecific(env->me_txkey, r);
1554 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1555 txn->mt_toggle = txn->mt_txnid & 1;
1556 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1557 txn->mt_u.reader = r;
1561 txn->mt_txnid = env->me_txns->mti_txnid;
1562 txn->mt_toggle = txn->mt_txnid & 1;
1563 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1566 if (txn->mt_txnid == mdb_debug_start)
1569 txn->mt_u.dirty_list = env->me_dirty_list;
1570 txn->mt_u.dirty_list[0].mid = 0;
1571 txn->mt_free_pgs = env->me_free_pgs;
1572 txn->mt_free_pgs[0] = 0;
1576 /* Copy the DB info and flags */
1577 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1578 for (i=2; i<txn->mt_numdbs; i++)
1579 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1580 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1581 memset(txn->mt_dbflags+2, DB_STALE, env->me_numdbs-2);
1587 mdb_txn_renew(MDB_txn *txn)
1594 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1595 DPUTS("environment had fatal error, must shutdown!");
1599 rc = mdb_txn_renew0(txn);
1600 if (rc == MDB_SUCCESS) {
1601 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1602 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1603 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1609 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1614 if (env->me_flags & MDB_FATAL_ERROR) {
1615 DPUTS("environment had fatal error, must shutdown!");
1619 /* parent already has an active child txn */
1620 if (parent->mt_child) {
1624 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1625 if (!(flags & MDB_RDONLY))
1626 size += env->me_maxdbs * sizeof(MDB_cursor *);
1628 if ((txn = calloc(1, size)) == NULL) {
1629 DPRINTF("calloc: %s", strerror(ErrCode()));
1632 txn->mt_dbs = (MDB_db *)(txn+1);
1633 if (flags & MDB_RDONLY) {
1634 txn->mt_flags |= MDB_TXN_RDONLY;
1635 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1637 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1638 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1643 txn->mt_free_pgs = mdb_midl_alloc();
1644 if (!txn->mt_free_pgs) {
1648 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1649 if (!txn->mt_u.dirty_list) {
1650 free(txn->mt_free_pgs);
1654 txn->mt_txnid = parent->mt_txnid;
1655 txn->mt_toggle = parent->mt_toggle;
1656 txn->mt_u.dirty_list[0].mid = 0;
1657 txn->mt_free_pgs[0] = 0;
1658 txn->mt_next_pgno = parent->mt_next_pgno;
1659 parent->mt_child = txn;
1660 txn->mt_parent = parent;
1661 txn->mt_numdbs = parent->mt_numdbs;
1662 txn->mt_dbxs = parent->mt_dbxs;
1663 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1664 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1665 mdb_cursor_shadow(parent, txn);
1668 rc = mdb_txn_renew0(txn);
1674 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1675 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1676 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1682 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1683 * @param[in] txn the transaction handle to reset
1686 mdb_txn_reset0(MDB_txn *txn)
1688 MDB_env *env = txn->mt_env;
1690 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1691 txn->mt_u.reader->mr_txnid = 0;
1697 /* close(free) all cursors */
1698 for (i=0; i<txn->mt_numdbs; i++) {
1699 if (txn->mt_cursors[i]) {
1701 while ((mc = txn->mt_cursors[i])) {
1702 txn->mt_cursors[i] = mc->mc_next;
1703 if (mc->mc_flags & C_ALLOCD)
1709 /* return all dirty pages to dpage list */
1710 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1711 dp = txn->mt_u.dirty_list[i].mptr;
1712 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1713 dp->mp_next = txn->mt_env->me_dpages;
1714 VGMEMP_FREE(txn->mt_env, dp);
1715 txn->mt_env->me_dpages = dp;
1717 /* large pages just get freed directly */
1718 VGMEMP_FREE(txn->mt_env, dp);
1723 if (txn->mt_parent) {
1724 txn->mt_parent->mt_child = NULL;
1725 free(txn->mt_free_pgs);
1726 free(txn->mt_u.dirty_list);
1729 if (mdb_midl_shrink(&txn->mt_free_pgs))
1730 env->me_free_pgs = txn->mt_free_pgs;
1733 while ((mop = txn->mt_env->me_pghead)) {
1734 txn->mt_env->me_pghead = mop->mo_next;
1737 txn->mt_env->me_pgfirst = 0;
1738 txn->mt_env->me_pglast = 0;
1741 /* The writer mutex was locked in mdb_txn_begin. */
1742 UNLOCK_MUTEX_W(env);
1747 mdb_txn_reset(MDB_txn *txn)
1752 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1753 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1754 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1756 mdb_txn_reset0(txn);
1760 mdb_txn_abort(MDB_txn *txn)
1765 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1766 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1767 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1770 mdb_txn_abort(txn->mt_child);
1772 mdb_txn_reset0(txn);
1777 mdb_txn_commit(MDB_txn *txn)
1785 pgno_t next, freecnt;
1788 assert(txn != NULL);
1789 assert(txn->mt_env != NULL);
1791 if (txn->mt_child) {
1792 mdb_txn_commit(txn->mt_child);
1793 txn->mt_child = NULL;
1798 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1799 if (txn->mt_numdbs > env->me_numdbs) {
1800 /* update the DB flags */
1802 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
1803 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
1810 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1811 DPUTS("error flag is set, can't commit");
1813 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1818 /* Merge (and close) our cursors with parent's */
1819 mdb_cursor_merge(txn);
1821 if (txn->mt_parent) {
1827 /* Update parent's DB table */
1828 ip = &txn->mt_parent->mt_dbs[2];
1829 jp = &txn->mt_dbs[2];
1830 for (i = 2; i < txn->mt_numdbs; i++) {
1831 if (ip->md_root != jp->md_root)
1835 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1837 /* Append our free list to parent's */
1838 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1840 mdb_midl_free(txn->mt_free_pgs);
1842 /* Merge our dirty list with parent's */
1843 dst = txn->mt_parent->mt_u.dirty_list;
1844 src = txn->mt_u.dirty_list;
1845 x = mdb_mid2l_search(dst, src[1].mid);
1846 for (y=1; y<=src[0].mid; y++) {
1847 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1851 dst[x].mptr = src[y].mptr;
1854 for (; y<=src[0].mid; y++) {
1855 if (++x >= MDB_IDL_UM_MAX) {
1862 free(txn->mt_u.dirty_list);
1863 txn->mt_parent->mt_child = NULL;
1868 if (txn != env->me_txn) {
1869 DPUTS("attempt to commit unknown transaction");
1874 if (!txn->mt_u.dirty_list[0].mid)
1877 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1878 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1880 /* Update DB root pointers. Their pages have already been
1881 * touched so this is all in-place and cannot fail.
1886 data.mv_size = sizeof(MDB_db);
1888 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
1889 for (i = 2; i < txn->mt_numdbs; i++) {
1890 if (txn->mt_dbflags[i] & DB_DIRTY) {
1891 data.mv_data = &txn->mt_dbs[i];
1892 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1897 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1899 /* should only be one record now */
1900 if (env->me_pghead) {
1901 /* make sure first page of freeDB is touched and on freelist */
1902 mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
1905 /* Delete IDLs we used from the free list */
1906 if (env->me_pgfirst) {
1911 key.mv_size = sizeof(cur);
1912 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
1915 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
1916 rc = mdb_cursor_del(&mc, 0);
1922 env->me_pgfirst = 0;
1926 /* save to free list */
1928 freecnt = txn->mt_free_pgs[0];
1929 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1932 /* make sure last page of freeDB is touched and on freelist */
1933 key.mv_size = MAXKEYSIZE+1;
1935 mdb_page_search(&mc, &key, MDB_PS_MODIFY);
1937 mdb_midl_sort(txn->mt_free_pgs);
1941 MDB_IDL idl = txn->mt_free_pgs;
1942 DPRINTF("IDL write txn %zu root %zu num %zu",
1943 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1944 for (i=0; i<idl[0]; i++) {
1945 DPRINTF("IDL %zu", idl[i+1]);
1949 /* write to last page of freeDB */
1950 key.mv_size = sizeof(pgno_t);
1951 key.mv_data = &txn->mt_txnid;
1952 data.mv_data = txn->mt_free_pgs;
1953 /* The free list can still grow during this call,
1954 * despite the pre-emptive touches above. So check
1955 * and make sure the entire thing got written.
1958 freecnt = txn->mt_free_pgs[0];
1959 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1960 rc = mdb_cursor_put(&mc, &key, &data, 0);
1965 } while (freecnt != txn->mt_free_pgs[0]);
1967 /* should only be one record now */
1969 if (env->me_pghead) {
1975 mop = env->me_pghead;
1977 key.mv_size = sizeof(id);
1979 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1980 data.mv_data = mop->mo_pages;
1981 orig = mop->mo_pages[0];
1982 /* These steps may grow the freelist again
1983 * due to freed overflow pages...
1985 mdb_cursor_put(&mc, &key, &data, 0);
1986 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
1987 /* could have been used again here */
1988 if (mop->mo_pages[0] != orig) {
1989 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1990 data.mv_data = mop->mo_pages;
1992 mdb_cursor_put(&mc, &key, &data, 0);
1994 env->me_pghead = NULL;
1997 /* was completely used up */
1998 mdb_cursor_del(&mc, 0);
2002 env->me_pgfirst = 0;
2006 while (env->me_pgfree) {
2007 MDB_oldpages *mop = env->me_pgfree;
2008 env->me_pgfree = mop->mo_next;
2012 /* Check for growth of freelist again */
2013 if (freecnt != txn->mt_free_pgs[0])
2016 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2017 if (mdb_midl_shrink(&txn->mt_free_pgs))
2018 env->me_free_pgs = txn->mt_free_pgs;
2025 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2031 /* Windows actually supports scatter/gather I/O, but only on
2032 * unbuffered file handles. Since we're relying on the OS page
2033 * cache for all our data, that's self-defeating. So we just
2034 * write pages one at a time. We use the ov structure to set
2035 * the write offset, to at least save the overhead of a Seek
2039 memset(&ov, 0, sizeof(ov));
2040 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2042 dp = txn->mt_u.dirty_list[i].mptr;
2043 DPRINTF("committing page %zu", dp->mp_pgno);
2044 size = dp->mp_pgno * env->me_psize;
2045 ov.Offset = size & 0xffffffff;
2046 ov.OffsetHigh = size >> 16;
2047 ov.OffsetHigh >>= 16;
2048 /* clear dirty flag */
2049 dp->mp_flags &= ~P_DIRTY;
2050 wsize = env->me_psize;
2051 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2052 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2055 DPRINTF("WriteFile: %d", n);
2062 struct iovec iov[MDB_COMMIT_PAGES];
2066 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2067 dp = txn->mt_u.dirty_list[i].mptr;
2068 if (dp->mp_pgno != next) {
2070 rc = writev(env->me_fd, iov, n);
2074 DPUTS("short write, filesystem full?");
2076 DPRINTF("writev: %s", strerror(n));
2083 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2086 DPRINTF("committing page %zu", dp->mp_pgno);
2087 iov[n].iov_len = env->me_psize;
2088 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2089 iov[n].iov_base = (char *)dp;
2090 size += iov[n].iov_len;
2091 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2092 /* clear dirty flag */
2093 dp->mp_flags &= ~P_DIRTY;
2094 if (++n >= MDB_COMMIT_PAGES) {
2104 rc = writev(env->me_fd, iov, n);
2108 DPUTS("short write, filesystem full?");
2110 DPRINTF("writev: %s", strerror(n));
2117 /* Drop the dirty pages.
2119 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2120 dp = txn->mt_u.dirty_list[i].mptr;
2121 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2122 dp->mp_next = txn->mt_env->me_dpages;
2123 VGMEMP_FREE(txn->mt_env, dp);
2124 txn->mt_env->me_dpages = dp;
2126 VGMEMP_FREE(txn->mt_env, dp);
2129 txn->mt_u.dirty_list[i].mid = 0;
2131 txn->mt_u.dirty_list[0].mid = 0;
2133 if ((n = mdb_env_sync(env, 0)) != 0 ||
2134 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2141 if (txn->mt_numdbs > env->me_numdbs) {
2142 /* update the DB flags */
2144 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2145 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2149 UNLOCK_MUTEX_W(env);
2155 /** Read the environment parameters of a DB environment before
2156 * mapping it into memory.
2157 * @param[in] env the environment handle
2158 * @param[out] meta address of where to store the meta information
2159 * @return 0 on success, non-zero on failure.
2162 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2169 /* We don't know the page size yet, so use a minimum value.
2173 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2175 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2180 else if (rc != MDB_PAGESIZE) {
2184 DPRINTF("read: %s", strerror(err));
2188 p = (MDB_page *)&pbuf;
2190 if (!F_ISSET(p->mp_flags, P_META)) {
2191 DPRINTF("page %zu not a meta page", p->mp_pgno);
2196 if (m->mm_magic != MDB_MAGIC) {
2197 DPUTS("meta has invalid magic");
2201 if (m->mm_version != MDB_VERSION) {
2202 DPRINTF("database is version %u, expected version %u",
2203 m->mm_version, MDB_VERSION);
2204 return MDB_VERSION_MISMATCH;
2207 memcpy(meta, m, sizeof(*m));
2211 /** Write the environment parameters of a freshly created DB environment.
2212 * @param[in] env the environment handle
2213 * @param[out] meta address of where to store the meta information
2214 * @return 0 on success, non-zero on failure.
2217 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2224 DPUTS("writing new meta page");
2226 GET_PAGESIZE(psize);
2228 meta->mm_magic = MDB_MAGIC;
2229 meta->mm_version = MDB_VERSION;
2230 meta->mm_psize = psize;
2231 meta->mm_last_pg = 1;
2232 meta->mm_flags = env->me_flags & 0xffff;
2233 meta->mm_flags |= MDB_INTEGERKEY;
2234 meta->mm_dbs[0].md_root = P_INVALID;
2235 meta->mm_dbs[1].md_root = P_INVALID;
2237 p = calloc(2, psize);
2239 p->mp_flags = P_META;
2242 memcpy(m, meta, sizeof(*meta));
2244 q = (MDB_page *)((char *)p + psize);
2247 q->mp_flags = P_META;
2250 memcpy(m, meta, sizeof(*meta));
2255 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2256 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2259 rc = write(env->me_fd, p, psize * 2);
2260 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2266 /** Update the environment info to commit a transaction.
2267 * @param[in] txn the transaction that's being committed
2268 * @return 0 on success, non-zero on failure.
2271 mdb_env_write_meta(MDB_txn *txn)
2274 MDB_meta meta, metab;
2276 int rc, len, toggle;
2282 assert(txn != NULL);
2283 assert(txn->mt_env != NULL);
2285 toggle = !txn->mt_toggle;
2286 DPRINTF("writing meta page %d for root page %zu",
2287 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2291 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2292 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2294 ptr = (char *)&meta;
2295 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2296 len = sizeof(MDB_meta) - off;
2299 meta.mm_dbs[0] = txn->mt_dbs[0];
2300 meta.mm_dbs[1] = txn->mt_dbs[1];
2301 meta.mm_last_pg = txn->mt_next_pgno - 1;
2302 meta.mm_txnid = txn->mt_txnid;
2305 off += env->me_psize;
2308 /* Write to the SYNC fd */
2311 memset(&ov, 0, sizeof(ov));
2313 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2316 rc = pwrite(env->me_mfd, ptr, len, off);
2321 DPUTS("write failed, disk error?");
2322 /* On a failure, the pagecache still contains the new data.
2323 * Write some old data back, to prevent it from being used.
2324 * Use the non-SYNC fd; we know it will fail anyway.
2326 meta.mm_last_pg = metab.mm_last_pg;
2327 meta.mm_txnid = metab.mm_txnid;
2329 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2331 r2 = pwrite(env->me_fd, ptr, len, off);
2333 env->me_flags |= MDB_FATAL_ERROR;
2336 /* Memory ordering issues are irrelevant; since the entire writer
2337 * is wrapped by wmutex, all of these changes will become visible
2338 * after the wmutex is unlocked. Since the DB is multi-version,
2339 * readers will get consistent data regardless of how fresh or
2340 * how stale their view of these values is.
2342 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2347 /** Check both meta pages to see which one is newer.
2348 * @param[in] env the environment handle
2349 * @return meta toggle (0 or 1).
2352 mdb_env_pick_meta(const MDB_env *env)
2354 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2358 mdb_env_create(MDB_env **env)
2362 e = calloc(1, sizeof(MDB_env));
2366 e->me_free_pgs = mdb_midl_alloc();
2367 if (!e->me_free_pgs) {
2371 e->me_maxreaders = DEFAULT_READERS;
2373 e->me_fd = INVALID_HANDLE_VALUE;
2374 e->me_lfd = INVALID_HANDLE_VALUE;
2375 e->me_mfd = INVALID_HANDLE_VALUE;
2376 VGMEMP_CREATE(e,0,0);
2382 mdb_env_set_mapsize(MDB_env *env, size_t size)
2386 env->me_mapsize = size;
2388 env->me_maxpg = env->me_mapsize / env->me_psize;
2393 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2397 env->me_maxdbs = dbs;
2402 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2404 if (env->me_map || readers < 1)
2406 env->me_maxreaders = readers;
2411 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2413 if (!env || !readers)
2415 *readers = env->me_maxreaders;
2419 /** Further setup required for opening an MDB environment
2422 mdb_env_open2(MDB_env *env, unsigned int flags)
2428 env->me_flags = flags;
2430 memset(&meta, 0, sizeof(meta));
2432 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2435 DPUTS("new mdbenv");
2439 if (!env->me_mapsize) {
2440 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2446 LONG sizelo, sizehi;
2447 sizelo = env->me_mapsize & 0xffffffff;
2448 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2450 /* Windows won't create mappings for zero length files.
2451 * Just allocate the maxsize right now.
2454 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2455 if (!SetEndOfFile(env->me_fd))
2457 SetFilePointer(env->me_fd, 0, NULL, 0);
2459 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
2460 sizehi, sizelo, NULL);
2463 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
2471 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2473 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
2475 if (env->me_map == MAP_FAILED) {
2482 meta.mm_mapsize = env->me_mapsize;
2483 if (flags & MDB_FIXEDMAP)
2484 meta.mm_address = env->me_map;
2485 i = mdb_env_init_meta(env, &meta);
2486 if (i != MDB_SUCCESS) {
2487 munmap(env->me_map, env->me_mapsize);
2491 env->me_psize = meta.mm_psize;
2493 env->me_maxpg = env->me_mapsize / env->me_psize;
2495 p = (MDB_page *)env->me_map;
2496 env->me_metas[0] = METADATA(p);
2497 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2501 int toggle = mdb_env_pick_meta(env);
2502 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2504 DPRINTF("opened database version %u, pagesize %u",
2505 env->me_metas[0]->mm_version, env->me_psize);
2506 DPRINTF("using meta page %d", toggle);
2507 DPRINTF("depth: %u", db->md_depth);
2508 DPRINTF("entries: %zu", db->md_entries);
2509 DPRINTF("branch pages: %zu", db->md_branch_pages);
2510 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2511 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2512 DPRINTF("root: %zu", db->md_root);
2520 /** Release a reader thread's slot in the reader lock table.
2521 * This function is called automatically when a thread exits.
2522 * @param[in] ptr This points to the slot in the reader lock table.
2525 mdb_env_reader_dest(void *ptr)
2527 MDB_reader *reader = ptr;
2529 reader->mr_txnid = 0;
2535 /** Junk for arranging thread-specific callbacks on Windows. This is
2536 * necessarily platform and compiler-specific. Windows supports up
2537 * to 1088 keys. Let's assume nobody opens more than 64 environments
2538 * in a single process, for now. They can override this if needed.
2540 #ifndef MAX_TLS_KEYS
2541 #define MAX_TLS_KEYS 64
2543 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2544 static int mdb_tls_nkeys;
2546 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2550 case DLL_PROCESS_ATTACH: break;
2551 case DLL_THREAD_ATTACH: break;
2552 case DLL_THREAD_DETACH:
2553 for (i=0; i<mdb_tls_nkeys; i++) {
2554 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2555 mdb_env_reader_dest(r);
2558 case DLL_PROCESS_DETACH: break;
2563 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2565 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2569 /* Force some symbol references.
2570 * _tls_used forces the linker to create the TLS directory if not already done
2571 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2573 #pragma comment(linker, "/INCLUDE:_tls_used")
2574 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2575 #pragma const_seg(".CRT$XLB")
2576 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2577 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2580 #pragma comment(linker, "/INCLUDE:__tls_used")
2581 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2582 #pragma data_seg(".CRT$XLB")
2583 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2585 #endif /* WIN 32/64 */
2586 #endif /* !__GNUC__ */
2589 /** Downgrade the exclusive lock on the region back to shared */
2591 mdb_env_share_locks(MDB_env *env)
2593 int toggle = mdb_env_pick_meta(env);
2595 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2600 /* First acquire a shared lock. The Unlock will
2601 * then release the existing exclusive lock.
2603 memset(&ov, 0, sizeof(ov));
2604 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2605 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2609 struct flock lock_info;
2610 /* The shared lock replaces the existing lock */
2611 memset((void *)&lock_info, 0, sizeof(lock_info));
2612 lock_info.l_type = F_RDLCK;
2613 lock_info.l_whence = SEEK_SET;
2614 lock_info.l_start = 0;
2615 lock_info.l_len = 1;
2616 fcntl(env->me_lfd, F_SETLK, &lock_info);
2620 #if defined(_WIN32) || defined(__APPLE__)
2622 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2624 * @(#) $Revision: 5.1 $
2625 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2626 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2628 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2632 * Please do not copyright this code. This code is in the public domain.
2634 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2635 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2636 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2637 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2638 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2639 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2640 * PERFORMANCE OF THIS SOFTWARE.
2643 * chongo <Landon Curt Noll> /\oo/\
2644 * http://www.isthe.com/chongo/
2646 * Share and Enjoy! :-)
2649 typedef unsigned long long mdb_hash_t;
2650 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2652 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2653 * @param[in] str string to hash
2654 * @param[in] hval initial value for hash
2655 * @return 64 bit hash
2657 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2658 * hval arg on the first call.
2661 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
2663 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
2664 unsigned char *end = s + val->mv_size;
2666 * FNV-1a hash each octet of the string
2669 /* xor the bottom with the current octet */
2670 hval ^= (mdb_hash_t)*s++;
2672 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2673 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2674 (hval << 7) + (hval << 8) + (hval << 40);
2676 /* return our new hash value */
2680 /** Hash the string and output the hash in hex.
2681 * @param[in] str string to hash
2682 * @param[out] hexbuf an array of 17 chars to hold the hash
2685 mdb_hash_hex(MDB_val *val, char *hexbuf)
2688 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
2689 for (i=0; i<8; i++) {
2690 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2696 /** Open and/or initialize the lock region for the environment.
2697 * @param[in] env The MDB environment.
2698 * @param[in] lpath The pathname of the file used for the lock region.
2699 * @param[in] mode The Unix permissions for the file, if we create it.
2700 * @param[out] excl Set to true if we got an exclusive lock on the region.
2701 * @return 0 on success, non-zero on failure.
2704 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2712 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2713 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2714 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2718 /* Try to get exclusive lock. If we succeed, then
2719 * nobody is using the lock region and we should initialize it.
2722 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2726 memset(&ov, 0, sizeof(ov));
2727 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2733 size = GetFileSize(env->me_lfd, NULL);
2739 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1)
2741 /* Lose record locks when exec*() */
2742 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
2743 fcntl(env->me_lfd, F_SETFD, fdflags);
2745 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
2746 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1)
2750 /* Try to get exclusive lock. If we succeed, then
2751 * nobody is using the lock region and we should initialize it.
2754 struct flock lock_info;
2755 memset((void *)&lock_info, 0, sizeof(lock_info));
2756 lock_info.l_type = F_WRLCK;
2757 lock_info.l_whence = SEEK_SET;
2758 lock_info.l_start = 0;
2759 lock_info.l_len = 1;
2760 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2764 lock_info.l_type = F_RDLCK;
2765 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2772 size = lseek(env->me_lfd, 0, SEEK_END);
2774 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2775 if (size < rsize && *excl) {
2777 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2778 if (!SetEndOfFile(env->me_lfd)) {
2783 if (ftruncate(env->me_lfd, rsize) != 0) {
2790 size = rsize - sizeof(MDB_txninfo);
2791 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2796 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2802 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2804 if (!env->me_txns) {
2809 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2811 if (m == MAP_FAILED) {
2812 env->me_txns = NULL;
2821 BY_HANDLE_FILE_INFORMATION stbuf;
2830 if (!mdb_sec_inited) {
2831 InitializeSecurityDescriptor(&mdb_null_sd,
2832 SECURITY_DESCRIPTOR_REVISION);
2833 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2834 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2835 mdb_all_sa.bInheritHandle = FALSE;
2836 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2839 GetFileInformationByHandle(env->me_lfd, &stbuf);
2840 idbuf.volume = stbuf.dwVolumeSerialNumber;
2841 idbuf.nhigh = stbuf.nFileIndexHigh;
2842 idbuf.nlow = stbuf.nFileIndexLow;
2843 val.mv_data = &idbuf;
2844 val.mv_size = sizeof(idbuf);
2845 mdb_hash_hex(&val, hexbuf);
2846 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
2847 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2848 if (!env->me_rmutex) {
2852 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
2853 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
2854 if (!env->me_wmutex) {
2868 fstat(env->me_lfd, &stbuf);
2869 idbuf.dev = stbuf.st_dev;
2870 idbuf.ino = stbuf.st_ino;
2871 val.mv_data = &idbuf;
2872 val.mv_size = sizeof(idbuf);
2873 mdb_hash_hex(&val, hexbuf);
2874 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
2875 if (sem_unlink(env->me_txns->mti_rmname)) {
2877 if (rc != ENOENT && rc != EINVAL)
2880 env->me_rmutex = sem_open(env->me_txns->mti_rmname, O_CREAT, mode, 1);
2881 if (!env->me_rmutex) {
2885 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
2886 if (sem_unlink(env->me_txns->mti_wmname)) {
2888 if (rc != ENOENT && rc != EINVAL)
2891 env->me_wmutex = sem_open(env->me_txns->mti_wmname, O_CREAT, mode, 1);
2892 if (!env->me_wmutex) {
2896 #else /* __APPLE__ */
2897 pthread_mutexattr_t mattr;
2899 pthread_mutexattr_init(&mattr);
2900 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2904 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2905 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2906 #endif /* __APPLE__ */
2908 env->me_txns->mti_version = MDB_VERSION;
2909 env->me_txns->mti_magic = MDB_MAGIC;
2910 env->me_txns->mti_txnid = 0;
2911 env->me_txns->mti_numreaders = 0;
2914 if (env->me_txns->mti_magic != MDB_MAGIC) {
2915 DPUTS("lock region has invalid magic");
2919 if (env->me_txns->mti_version != MDB_VERSION) {
2920 DPRINTF("lock region is version %u, expected version %u",
2921 env->me_txns->mti_version, MDB_VERSION);
2922 rc = MDB_VERSION_MISMATCH;
2926 if (rc != EACCES && rc != EAGAIN) {
2930 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2931 if (!env->me_rmutex) {
2935 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2936 if (!env->me_wmutex) {
2942 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
2943 if (!env->me_rmutex) {
2947 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
2948 if (!env->me_wmutex) {
2958 env->me_lfd = INVALID_HANDLE_VALUE;
2963 /** The name of the lock file in the DB environment */
2964 #define LOCKNAME "/lock.mdb"
2965 /** The name of the data file in the DB environment */
2966 #define DATANAME "/data.mdb"
2967 /** The suffix of the lock file when no subdir is used */
2968 #define LOCKSUFF "-lock"
2971 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2973 int oflags, rc, len, excl;
2974 char *lpath, *dpath;
2977 if (flags & MDB_NOSUBDIR) {
2978 rc = len + sizeof(LOCKSUFF) + len + 1;
2980 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
2985 if (flags & MDB_NOSUBDIR) {
2986 dpath = lpath + len + sizeof(LOCKSUFF);
2987 sprintf(lpath, "%s" LOCKSUFF, path);
2988 strcpy(dpath, path);
2990 dpath = lpath + len + sizeof(LOCKNAME);
2991 sprintf(lpath, "%s" LOCKNAME, path);
2992 sprintf(dpath, "%s" DATANAME, path);
2995 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3000 if (F_ISSET(flags, MDB_RDONLY)) {
3001 oflags = GENERIC_READ;
3002 len = OPEN_EXISTING;
3004 oflags = GENERIC_READ|GENERIC_WRITE;
3007 mode = FILE_ATTRIBUTE_NORMAL;
3008 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3009 NULL, len, mode, NULL);
3011 if (F_ISSET(flags, MDB_RDONLY))
3014 oflags = O_RDWR | O_CREAT;
3016 env->me_fd = open(dpath, oflags, mode);
3018 if (env->me_fd == INVALID_HANDLE_VALUE) {
3023 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
3024 if (flags & (MDB_RDONLY|MDB_NOSYNC|MDB_NOMETASYNC)) {
3025 env->me_mfd = env->me_fd;
3027 /* synchronous fd for meta writes */
3029 env->me_mfd = CreateFile(dpath, oflags,
3030 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3031 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3033 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3035 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3040 env->me_path = strdup(path);
3041 DPRINTF("opened dbenv %p", (void *) env);
3042 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3044 /* Windows TLS callbacks need help finding their TLS info. */
3045 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3046 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3053 mdb_env_share_locks(env);
3055 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3056 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3057 if (!env->me_dbxs || !env->me_dbflags)
3063 if (env->me_fd != INVALID_HANDLE_VALUE) {
3065 env->me_fd = INVALID_HANDLE_VALUE;
3067 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3069 env->me_lfd = INVALID_HANDLE_VALUE;
3077 mdb_env_close(MDB_env *env)
3084 VGMEMP_DESTROY(env);
3085 while (env->me_dpages) {
3086 dp = env->me_dpages;
3087 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3088 env->me_dpages = dp->mp_next;
3092 free(env->me_dbflags);
3096 pthread_key_delete(env->me_txkey);
3098 /* Delete our key from the global list */
3100 for (i=0; i<mdb_tls_nkeys; i++)
3101 if (mdb_tls_keys[i] == env->me_txkey) {
3102 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3110 munmap(env->me_map, env->me_mapsize);
3112 if (env->me_mfd != env->me_fd)
3116 pid_t pid = getpid();
3118 for (i=0; i<env->me_txns->mti_numreaders; i++)
3119 if (env->me_txns->mti_readers[i].mr_pid == pid)
3120 env->me_txns->mti_readers[i].mr_pid = 0;
3121 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3124 mdb_midl_free(env->me_free_pgs);
3128 /** Compare two items pointing at aligned size_t's */
3130 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3132 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3133 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3136 /** Compare two items pointing at aligned int's */
3138 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3140 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3141 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3144 /** Compare two items pointing at ints of unknown alignment.
3145 * Nodes and keys are guaranteed to be 2-byte aligned.
3148 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3150 #if BYTE_ORDER == LITTLE_ENDIAN
3151 unsigned short *u, *c;
3154 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3155 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3158 } while(!x && u > (unsigned short *)a->mv_data);
3161 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3165 /** Compare two items lexically */
3167 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3174 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3180 diff = memcmp(a->mv_data, b->mv_data, len);
3181 return diff ? diff : len_diff<0 ? -1 : len_diff;
3184 /** Compare two items in reverse byte order */
3186 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3188 const unsigned char *p1, *p2, *p1_lim;
3192 p1_lim = (const unsigned char *)a->mv_data;
3193 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3194 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3196 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3202 while (p1 > p1_lim) {
3203 diff = *--p1 - *--p2;
3207 return len_diff<0 ? -1 : len_diff;
3210 /** Search for key within a page, using binary search.
3211 * Returns the smallest entry larger or equal to the key.
3212 * If exactp is non-null, stores whether the found entry was an exact match
3213 * in *exactp (1 or 0).
3214 * Updates the cursor index with the index of the found entry.
3215 * If no entry larger or equal to the key is found, returns NULL.
3218 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3220 unsigned int i = 0, nkeys;
3223 MDB_page *mp = mc->mc_pg[mc->mc_top];
3224 MDB_node *node = NULL;
3229 nkeys = NUMKEYS(mp);
3234 COPY_PGNO(pgno, mp->mp_pgno);
3235 DPRINTF("searching %u keys in %s %spage %zu",
3236 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3243 low = IS_LEAF(mp) ? 0 : 1;
3245 cmp = mc->mc_dbx->md_cmp;
3247 /* Branch pages have no data, so if using integer keys,
3248 * alignment is guaranteed. Use faster mdb_cmp_int.
3250 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3251 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3258 nodekey.mv_size = mc->mc_db->md_pad;
3259 node = NODEPTR(mp, 0); /* fake */
3260 while (low <= high) {
3261 i = (low + high) >> 1;
3262 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3263 rc = cmp(key, &nodekey);
3264 DPRINTF("found leaf index %u [%s], rc = %i",
3265 i, DKEY(&nodekey), rc);
3274 while (low <= high) {
3275 i = (low + high) >> 1;
3277 node = NODEPTR(mp, i);
3278 nodekey.mv_size = NODEKSZ(node);
3279 nodekey.mv_data = NODEKEY(node);
3281 rc = cmp(key, &nodekey);
3284 DPRINTF("found leaf index %u [%s], rc = %i",
3285 i, DKEY(&nodekey), rc);
3287 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3288 i, DKEY(&nodekey), NODEPGNO(node), rc);
3299 if (rc > 0) { /* Found entry is less than the key. */
3300 i++; /* Skip to get the smallest entry larger than key. */
3302 node = NODEPTR(mp, i);
3305 *exactp = (rc == 0);
3306 /* store the key index */
3307 mc->mc_ki[mc->mc_top] = i;
3309 /* There is no entry larger or equal to the key. */
3312 /* nodeptr is fake for LEAF2 */
3318 mdb_cursor_adjust(MDB_cursor *mc, func)
3322 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3323 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3330 /** Pop a page off the top of the cursor's stack. */
3332 mdb_cursor_pop(MDB_cursor *mc)
3337 top = mc->mc_pg[mc->mc_top];
3342 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3343 mc->mc_dbi, (void *) mc);
3347 /** Push a page onto the top of the cursor's stack. */
3349 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3351 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3352 mc->mc_dbi, (void *) mc);
3354 if (mc->mc_snum >= CURSOR_STACK) {
3355 assert(mc->mc_snum < CURSOR_STACK);
3359 mc->mc_top = mc->mc_snum++;
3360 mc->mc_pg[mc->mc_top] = mp;
3361 mc->mc_ki[mc->mc_top] = 0;
3366 /** Find the address of the page corresponding to a given page number.
3367 * @param[in] txn the transaction for this access.
3368 * @param[in] pgno the page number for the page to retrieve.
3369 * @param[out] ret address of a pointer where the page's address will be stored.
3370 * @return 0 on success, non-zero on failure.
3373 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3377 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3379 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3380 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3381 p = txn->mt_u.dirty_list[x].mptr;
3385 if (pgno < txn->mt_next_pgno)
3386 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3390 DPRINTF("page %zu not found", pgno);
3393 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3396 /** Search for the page a given key should be in.
3397 * Pushes parent pages on the cursor stack. This function continues a
3398 * search on a cursor that has already been initialized. (Usually by
3399 * #mdb_page_search() but also by #mdb_node_move().)
3400 * @param[in,out] mc the cursor for this operation.
3401 * @param[in] key the key to search for. If NULL, search for the lowest
3402 * page. (This is used by #mdb_cursor_first().)
3403 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3404 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3405 * @return 0 on success, non-zero on failure.
3408 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3410 MDB_page *mp = mc->mc_pg[mc->mc_top];
3415 while (IS_BRANCH(mp)) {
3419 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3420 assert(NUMKEYS(mp) > 1);
3421 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3423 if (key == NULL) /* Initialize cursor to first page. */
3425 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3426 /* cursor to last page */
3430 node = mdb_node_search(mc, key, &exact);
3432 i = NUMKEYS(mp) - 1;
3434 i = mc->mc_ki[mc->mc_top];
3443 DPRINTF("following index %u for key [%s]",
3445 assert(i < NUMKEYS(mp));
3446 node = NODEPTR(mp, i);
3448 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3451 mc->mc_ki[mc->mc_top] = i;
3452 if ((rc = mdb_cursor_push(mc, mp)))
3456 if ((rc = mdb_page_touch(mc)) != 0)
3458 mp = mc->mc_pg[mc->mc_top];
3463 DPRINTF("internal error, index points to a %02X page!?",
3465 return MDB_CORRUPTED;
3468 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3469 key ? DKEY(key) : NULL);
3474 /** Search for the page a given key should be in.
3475 * Pushes parent pages on the cursor stack. This function just sets up
3476 * the search; it finds the root page for \b mc's database and sets this
3477 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3478 * called to complete the search.
3479 * @param[in,out] mc the cursor for this operation.
3480 * @param[in] key the key to search for. If NULL, search for the lowest
3481 * page. (This is used by #mdb_cursor_first().)
3482 * @param[in] modify If true, visited pages are updated with new page numbers.
3483 * @return 0 on success, non-zero on failure.
3486 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
3491 /* Make sure the txn is still viable, then find the root from
3492 * the txn's db table.
3494 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3495 DPUTS("transaction has failed, must abort");
3498 /* Make sure we're using an up-to-date root */
3499 if (mc->mc_dbi > MAIN_DBI) {
3500 if ((*mc->mc_dbflag & DB_STALE) ||
3501 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
3503 unsigned char dbflag = 0;
3504 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3505 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
3508 if (*mc->mc_dbflag & DB_STALE) {
3511 MDB_node *leaf = mdb_node_search(&mc2,
3512 &mc->mc_dbx->md_name, &exact);
3514 return MDB_NOTFOUND;
3515 mdb_node_read(mc->mc_txn, leaf, &data);
3516 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3518 if (flags & MDB_PS_MODIFY)
3520 *mc->mc_dbflag = dbflag;
3523 root = mc->mc_db->md_root;
3525 if (root == P_INVALID) { /* Tree is empty. */
3526 DPUTS("tree is empty");
3527 return MDB_NOTFOUND;
3532 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
3533 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3539 DPRINTF("db %u root page %zu has flags 0x%X",
3540 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3542 if (flags & MDB_PS_MODIFY) {
3543 if ((rc = mdb_page_touch(mc)))
3547 if (flags & MDB_PS_ROOTONLY)
3550 return mdb_page_search_root(mc, key, flags);
3553 /** Return the data associated with a given node.
3554 * @param[in] txn The transaction for this operation.
3555 * @param[in] leaf The node being read.
3556 * @param[out] data Updated to point to the node's data.
3557 * @return 0 on success, non-zero on failure.
3560 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3562 MDB_page *omp; /* overflow page */
3566 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3567 data->mv_size = NODEDSZ(leaf);
3568 data->mv_data = NODEDATA(leaf);
3572 /* Read overflow data.
3574 data->mv_size = NODEDSZ(leaf);
3575 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3576 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3577 DPRINTF("read overflow page %zu failed", pgno);
3580 data->mv_data = METADATA(omp);
3586 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3587 MDB_val *key, MDB_val *data)
3596 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3598 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3601 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3605 mdb_cursor_init(&mc, txn, dbi, &mx);
3606 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3609 /** Find a sibling for a page.
3610 * Replaces the page at the top of the cursor's stack with the
3611 * specified sibling, if one exists.
3612 * @param[in] mc The cursor for this operation.
3613 * @param[in] move_right Non-zero if the right sibling is requested,
3614 * otherwise the left sibling.
3615 * @return 0 on success, non-zero on failure.
3618 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3624 if (mc->mc_snum < 2) {
3625 return MDB_NOTFOUND; /* root has no siblings */
3629 DPRINTF("parent page is page %zu, index %u",
3630 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3632 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3633 : (mc->mc_ki[mc->mc_top] == 0)) {
3634 DPRINTF("no more keys left, moving to %s sibling",
3635 move_right ? "right" : "left");
3636 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3640 mc->mc_ki[mc->mc_top]++;
3642 mc->mc_ki[mc->mc_top]--;
3643 DPRINTF("just moving to %s index key %u",
3644 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3646 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3648 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3649 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3652 mdb_cursor_push(mc, mp);
3657 /** Move the cursor to the next data item. */
3659 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3665 if (mc->mc_flags & C_EOF) {
3666 return MDB_NOTFOUND;
3669 assert(mc->mc_flags & C_INITIALIZED);
3671 mp = mc->mc_pg[mc->mc_top];
3673 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3674 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3675 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3676 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3677 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3678 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3682 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3683 if (op == MDB_NEXT_DUP)
3684 return MDB_NOTFOUND;
3688 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3690 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3691 DPUTS("=====> move to next sibling page");
3692 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3693 mc->mc_flags |= C_EOF;
3694 mc->mc_flags &= ~C_INITIALIZED;
3695 return MDB_NOTFOUND;
3697 mp = mc->mc_pg[mc->mc_top];
3698 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3700 mc->mc_ki[mc->mc_top]++;
3702 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3703 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3706 key->mv_size = mc->mc_db->md_pad;
3707 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3711 assert(IS_LEAF(mp));
3712 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3714 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3715 mdb_xcursor_init1(mc, leaf);
3718 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3721 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3722 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3723 if (rc != MDB_SUCCESS)
3728 MDB_SET_KEY(leaf, key);
3732 /** Move the cursor to the previous data item. */
3734 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3740 assert(mc->mc_flags & C_INITIALIZED);
3742 mp = mc->mc_pg[mc->mc_top];
3744 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3745 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3746 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3747 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3748 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3749 if (op != MDB_PREV || rc == MDB_SUCCESS)
3752 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3753 if (op == MDB_PREV_DUP)
3754 return MDB_NOTFOUND;
3759 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3761 if (mc->mc_ki[mc->mc_top] == 0) {
3762 DPUTS("=====> move to prev sibling page");
3763 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3764 mc->mc_flags &= ~C_INITIALIZED;
3765 return MDB_NOTFOUND;
3767 mp = mc->mc_pg[mc->mc_top];
3768 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3769 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3771 mc->mc_ki[mc->mc_top]--;
3773 mc->mc_flags &= ~C_EOF;
3775 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3776 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3779 key->mv_size = mc->mc_db->md_pad;
3780 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3784 assert(IS_LEAF(mp));
3785 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3787 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3788 mdb_xcursor_init1(mc, leaf);
3791 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3794 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3795 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3796 if (rc != MDB_SUCCESS)
3801 MDB_SET_KEY(leaf, key);
3805 /** Set the cursor on a specific data item. */
3807 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3808 MDB_cursor_op op, int *exactp)
3817 assert(key->mv_size > 0);
3819 /* See if we're already on the right page */
3820 if (mc->mc_flags & C_INITIALIZED) {
3823 mp = mc->mc_pg[mc->mc_top];
3825 mc->mc_ki[mc->mc_top] = 0;
3826 return MDB_NOTFOUND;
3828 if (mp->mp_flags & P_LEAF2) {
3829 nodekey.mv_size = mc->mc_db->md_pad;
3830 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
3832 leaf = NODEPTR(mp, 0);
3833 MDB_SET_KEY(leaf, &nodekey);
3835 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3837 /* Probably happens rarely, but first node on the page
3838 * was the one we wanted.
3840 mc->mc_ki[mc->mc_top] = 0;
3847 unsigned int nkeys = NUMKEYS(mp);
3849 if (mp->mp_flags & P_LEAF2) {
3850 nodekey.mv_data = LEAF2KEY(mp,
3851 nkeys-1, nodekey.mv_size);
3853 leaf = NODEPTR(mp, nkeys-1);
3854 MDB_SET_KEY(leaf, &nodekey);
3856 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3858 /* last node was the one we wanted */
3859 mc->mc_ki[mc->mc_top] = nkeys-1;
3865 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
3866 /* This is definitely the right page, skip search_page */
3867 if (mp->mp_flags & P_LEAF2) {
3868 nodekey.mv_data = LEAF2KEY(mp,
3869 mc->mc_ki[mc->mc_top], nodekey.mv_size);
3871 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3872 MDB_SET_KEY(leaf, &nodekey);
3874 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3876 /* current node was the one we wanted */
3886 /* If any parents have right-sibs, search.
3887 * Otherwise, there's nothing further.
3889 for (i=0; i<mc->mc_top; i++)
3891 NUMKEYS(mc->mc_pg[i])-1)
3893 if (i == mc->mc_top) {
3894 /* There are no other pages */
3895 mc->mc_ki[mc->mc_top] = nkeys;
3896 return MDB_NOTFOUND;
3900 /* There are no other pages */
3901 mc->mc_ki[mc->mc_top] = 0;
3902 return MDB_NOTFOUND;
3906 rc = mdb_page_search(mc, key, 0);
3907 if (rc != MDB_SUCCESS)
3910 mp = mc->mc_pg[mc->mc_top];
3911 assert(IS_LEAF(mp));
3914 leaf = mdb_node_search(mc, key, exactp);
3915 if (exactp != NULL && !*exactp) {
3916 /* MDB_SET specified and not an exact match. */
3917 return MDB_NOTFOUND;
3921 DPUTS("===> inexact leaf not found, goto sibling");
3922 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
3923 return rc; /* no entries matched */
3924 mp = mc->mc_pg[mc->mc_top];
3925 assert(IS_LEAF(mp));
3926 leaf = NODEPTR(mp, 0);
3930 mc->mc_flags |= C_INITIALIZED;
3931 mc->mc_flags &= ~C_EOF;
3934 key->mv_size = mc->mc_db->md_pad;
3935 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3939 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3940 mdb_xcursor_init1(mc, leaf);
3943 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3944 if (op == MDB_SET || op == MDB_SET_RANGE) {
3945 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3948 if (op == MDB_GET_BOTH) {
3954 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
3955 if (rc != MDB_SUCCESS)
3958 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
3960 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
3962 rc = mc->mc_dbx->md_dcmp(data, &d2);
3964 if (op == MDB_GET_BOTH || rc > 0)
3965 return MDB_NOTFOUND;
3970 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3971 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3976 /* The key already matches in all other cases */
3977 if (op == MDB_SET_RANGE)
3978 MDB_SET_KEY(leaf, key);
3979 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3984 /** Move the cursor to the first item in the database. */
3986 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3991 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3992 rc = mdb_page_search(mc, NULL, 0);
3993 if (rc != MDB_SUCCESS)
3996 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3998 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3999 mc->mc_flags |= C_INITIALIZED;
4000 mc->mc_flags &= ~C_EOF;
4002 mc->mc_ki[mc->mc_top] = 0;
4004 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4005 key->mv_size = mc->mc_db->md_pad;
4006 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4011 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4012 mdb_xcursor_init1(mc, leaf);
4013 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4018 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4019 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4023 MDB_SET_KEY(leaf, key);
4027 /** Move the cursor to the last item in the database. */
4029 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4034 if (!(mc->mc_flags & C_EOF)) {
4036 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4039 lkey.mv_size = MAXKEYSIZE+1;
4040 lkey.mv_data = NULL;
4041 rc = mdb_page_search(mc, &lkey, 0);
4042 if (rc != MDB_SUCCESS)
4045 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4047 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4048 mc->mc_flags |= C_INITIALIZED|C_EOF;
4050 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4052 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4053 key->mv_size = mc->mc_db->md_pad;
4054 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4059 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4060 mdb_xcursor_init1(mc, leaf);
4061 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4066 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4067 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4072 MDB_SET_KEY(leaf, key);
4077 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4087 case MDB_GET_BOTH_RANGE:
4088 if (data == NULL || mc->mc_xcursor == NULL) {
4095 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4097 } else if (op == MDB_SET_RANGE)
4098 rc = mdb_cursor_set(mc, key, data, op, NULL);
4100 rc = mdb_cursor_set(mc, key, data, op, &exact);
4102 case MDB_GET_MULTIPLE:
4104 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4105 !(mc->mc_flags & C_INITIALIZED)) {
4110 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4111 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4114 case MDB_NEXT_MULTIPLE:
4116 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4120 if (!(mc->mc_flags & C_INITIALIZED))
4121 rc = mdb_cursor_first(mc, key, data);
4123 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4124 if (rc == MDB_SUCCESS) {
4125 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4128 mx = &mc->mc_xcursor->mx_cursor;
4129 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4131 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4132 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4140 case MDB_NEXT_NODUP:
4141 if (!(mc->mc_flags & C_INITIALIZED))
4142 rc = mdb_cursor_first(mc, key, data);
4144 rc = mdb_cursor_next(mc, key, data, op);
4148 case MDB_PREV_NODUP:
4149 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4150 rc = mdb_cursor_last(mc, key, data);
4151 mc->mc_flags &= ~C_EOF;
4153 rc = mdb_cursor_prev(mc, key, data, op);
4156 rc = mdb_cursor_first(mc, key, data);
4160 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4161 !(mc->mc_flags & C_INITIALIZED) ||
4162 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4166 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4169 rc = mdb_cursor_last(mc, key, data);
4173 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4174 !(mc->mc_flags & C_INITIALIZED) ||
4175 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4179 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4182 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4190 /** Touch all the pages in the cursor stack.
4191 * Makes sure all the pages are writable, before attempting a write operation.
4192 * @param[in] mc The cursor to operate on.
4195 mdb_cursor_touch(MDB_cursor *mc)
4199 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4201 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4202 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4205 *mc->mc_dbflag = DB_DIRTY;
4207 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4208 rc = mdb_page_touch(mc);
4212 mc->mc_top = mc->mc_snum-1;
4217 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4220 MDB_node *leaf = NULL;
4221 MDB_val xdata, *rdata, dkey;
4224 int do_sub = 0, insert = 0;
4225 unsigned int mcount = 0;
4229 char dbuf[MAXKEYSIZE+1];
4230 unsigned int nflags;
4233 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4236 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4237 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4241 if (flags == MDB_CURRENT) {
4242 if (!(mc->mc_flags & C_INITIALIZED))
4245 } else if (mc->mc_db->md_root == P_INVALID) {
4247 /* new database, write a root leaf page */
4248 DPUTS("allocating new root leaf page");
4249 if ((np = mdb_page_new(mc, P_LEAF, 1)) == NULL) {
4253 mdb_cursor_push(mc, np);
4254 mc->mc_db->md_root = np->mp_pgno;
4255 mc->mc_db->md_depth++;
4256 *mc->mc_dbflag = DB_DIRTY;
4257 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4259 np->mp_flags |= P_LEAF2;
4260 mc->mc_flags |= C_INITIALIZED;
4266 if (flags & MDB_APPEND) {
4268 rc = mdb_cursor_last(mc, &k2, &d2);
4270 rc = mc->mc_dbx->md_cmp(key, &k2);
4273 mc->mc_ki[mc->mc_top]++;
4279 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4281 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4282 DPRINTF("duplicate key [%s]", DKEY(key));
4284 return MDB_KEYEXIST;
4286 if (rc && rc != MDB_NOTFOUND)
4290 /* Cursor is positioned, now make sure all pages are writable */
4291 rc2 = mdb_cursor_touch(mc);
4296 /* The key already exists */
4297 if (rc == MDB_SUCCESS) {
4298 /* there's only a key anyway, so this is a no-op */
4299 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4300 unsigned int ksize = mc->mc_db->md_pad;
4301 if (key->mv_size != ksize)
4303 if (flags == MDB_CURRENT) {
4304 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4305 memcpy(ptr, key->mv_data, ksize);
4310 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4313 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4314 /* Was a single item before, must convert now */
4316 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4317 /* Just overwrite the current item */
4318 if (flags == MDB_CURRENT)
4321 dkey.mv_size = NODEDSZ(leaf);
4322 dkey.mv_data = NODEDATA(leaf);
4323 #if UINT_MAX < SIZE_MAX
4324 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4325 #ifdef MISALIGNED_OK
4326 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4328 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4331 /* if data matches, ignore it */
4332 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4333 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4335 /* create a fake page for the dup items */
4336 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4337 dkey.mv_data = dbuf;
4338 fp = (MDB_page *)&pbuf;
4339 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4340 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4341 fp->mp_lower = PAGEHDRSZ;
4342 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4343 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4344 fp->mp_flags |= P_LEAF2;
4345 fp->mp_pad = data->mv_size;
4346 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4348 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4349 (dkey.mv_size & 1) + (data->mv_size & 1);
4351 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4354 xdata.mv_size = fp->mp_upper;
4359 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4360 /* See if we need to convert from fake page to subDB */
4362 unsigned int offset;
4365 fp = NODEDATA(leaf);
4366 if (flags == MDB_CURRENT) {
4368 fp->mp_flags |= P_DIRTY;
4369 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4370 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4374 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4375 offset = fp->mp_pad;
4376 if (SIZELEFT(fp) >= offset)
4378 offset *= 4; /* space for 4 more */
4380 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4382 offset += offset & 1;
4383 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4384 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4386 /* yes, convert it */
4388 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4389 dummy.md_pad = fp->mp_pad;
4390 dummy.md_flags = MDB_DUPFIXED;
4391 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4392 dummy.md_flags |= MDB_INTEGERKEY;
4395 dummy.md_branch_pages = 0;
4396 dummy.md_leaf_pages = 1;
4397 dummy.md_overflow_pages = 0;
4398 dummy.md_entries = NUMKEYS(fp);
4400 xdata.mv_size = sizeof(MDB_db);
4401 xdata.mv_data = &dummy;
4402 mp = mdb_page_alloc(mc, 1);
4405 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4406 flags |= F_DUPDATA|F_SUBDATA;
4407 dummy.md_root = mp->mp_pgno;
4409 /* no, just grow it */
4411 xdata.mv_size = NODEDSZ(leaf) + offset;
4412 xdata.mv_data = &pbuf;
4413 mp = (MDB_page *)&pbuf;
4414 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4417 mp->mp_flags = fp->mp_flags | P_DIRTY;
4418 mp->mp_pad = fp->mp_pad;
4419 mp->mp_lower = fp->mp_lower;
4420 mp->mp_upper = fp->mp_upper + offset;
4422 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4424 nsize = NODEDSZ(leaf) - fp->mp_upper;
4425 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4426 for (i=0; i<NUMKEYS(fp); i++)
4427 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4429 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4433 /* data is on sub-DB, just store it */
4434 flags |= F_DUPDATA|F_SUBDATA;
4438 /* overflow page overwrites need special handling */
4439 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4442 int ovpages, dpages;
4444 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4445 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4446 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4447 mdb_page_get(mc->mc_txn, pg, &omp);
4448 /* Is the ov page writable and large enough? */
4449 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4450 /* yes, overwrite it. Note in this case we don't
4451 * bother to try shrinking the node if the new data
4452 * is smaller than the overflow threshold.
4454 if (F_ISSET(flags, MDB_RESERVE))
4455 data->mv_data = METADATA(omp);
4457 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4460 /* no, free ovpages */
4462 mc->mc_db->md_overflow_pages -= ovpages;
4463 for (i=0; i<ovpages; i++) {
4464 DPRINTF("freed ov page %zu", pg);
4465 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4469 } else if (NODEDSZ(leaf) == data->mv_size) {
4470 /* same size, just replace it. Note that we could
4471 * also reuse this node if the new data is smaller,
4472 * but instead we opt to shrink the node in that case.
4474 if (F_ISSET(flags, MDB_RESERVE))
4475 data->mv_data = NODEDATA(leaf);
4477 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4480 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4481 mc->mc_db->md_entries--;
4483 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4490 nflags = flags & NODE_ADD_FLAGS;
4491 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4492 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4493 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4494 nflags &= ~MDB_APPEND;
4496 nflags |= MDB_SPLIT_REPLACE;
4497 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4499 /* There is room already in this leaf page. */
4500 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4501 if (rc == 0 && !do_sub && insert) {
4502 /* Adjust other cursors pointing to mp */
4503 MDB_cursor *m2, *m3;
4504 MDB_dbi dbi = mc->mc_dbi;
4505 unsigned i = mc->mc_top;
4506 MDB_page *mp = mc->mc_pg[i];
4508 if (mc->mc_flags & C_SUB)
4511 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4512 if (mc->mc_flags & C_SUB)
4513 m3 = &m2->mc_xcursor->mx_cursor;
4516 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4517 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4524 if (rc != MDB_SUCCESS)
4525 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4527 /* Now store the actual data in the child DB. Note that we're
4528 * storing the user data in the keys field, so there are strict
4529 * size limits on dupdata. The actual data fields of the child
4530 * DB are all zero size.
4537 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4538 if (flags & MDB_CURRENT) {
4539 xflags = MDB_CURRENT;
4541 mdb_xcursor_init1(mc, leaf);
4542 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4544 /* converted, write the original data first */
4546 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4550 /* Adjust other cursors pointing to mp */
4552 unsigned i = mc->mc_top;
4553 MDB_page *mp = mc->mc_pg[i];
4555 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4556 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4557 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4558 mdb_xcursor_init1(m2, leaf);
4563 if (flags & MDB_APPENDDUP)
4564 xflags |= MDB_APPEND;
4565 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4566 if (flags & F_SUBDATA) {
4567 void *db = NODEDATA(leaf);
4568 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4571 /* sub-writes might have failed so check rc again.
4572 * Don't increment count if we just replaced an existing item.
4574 if (!rc && !(flags & MDB_CURRENT))
4575 mc->mc_db->md_entries++;
4576 if (flags & MDB_MULTIPLE) {
4578 if (mcount < data[1].mv_size) {
4579 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4580 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4590 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4595 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4598 if (!mc->mc_flags & C_INITIALIZED)
4601 rc = mdb_cursor_touch(mc);
4605 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4607 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4608 if (flags != MDB_NODUPDATA) {
4609 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4610 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4612 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4613 /* If sub-DB still has entries, we're done */
4614 if (mc->mc_xcursor->mx_db.md_entries) {
4615 if (leaf->mn_flags & F_SUBDATA) {
4616 /* update subDB info */
4617 void *db = NODEDATA(leaf);
4618 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4620 /* shrink fake page */
4621 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4623 mc->mc_db->md_entries--;
4626 /* otherwise fall thru and delete the sub-DB */
4629 if (leaf->mn_flags & F_SUBDATA) {
4630 /* add all the child DB's pages to the free list */
4631 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4632 if (rc == MDB_SUCCESS) {
4633 mc->mc_db->md_entries -=
4634 mc->mc_xcursor->mx_db.md_entries;
4639 return mdb_cursor_del0(mc, leaf);
4642 /** Allocate and initialize new pages for a database.
4643 * @param[in] mc a cursor on the database being added to.
4644 * @param[in] flags flags defining what type of page is being allocated.
4645 * @param[in] num the number of pages to allocate. This is usually 1,
4646 * unless allocating overflow pages for a large record.
4647 * @return Address of a page, or NULL on failure.
4650 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num)
4654 if ((np = mdb_page_alloc(mc, num)) == NULL)
4656 DPRINTF("allocated new mpage %zu, page size %u",
4657 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4658 np->mp_flags = flags | P_DIRTY;
4659 np->mp_lower = PAGEHDRSZ;
4660 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4663 mc->mc_db->md_branch_pages++;
4664 else if (IS_LEAF(np))
4665 mc->mc_db->md_leaf_pages++;
4666 else if (IS_OVERFLOW(np)) {
4667 mc->mc_db->md_overflow_pages += num;
4674 /** Calculate the size of a leaf node.
4675 * The size depends on the environment's page size; if a data item
4676 * is too large it will be put onto an overflow page and the node
4677 * size will only include the key and not the data. Sizes are always
4678 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4679 * of the #MDB_node headers.
4680 * @param[in] env The environment handle.
4681 * @param[in] key The key for the node.
4682 * @param[in] data The data for the node.
4683 * @return The number of bytes needed to store the node.
4686 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4690 sz = LEAFSIZE(key, data);
4691 if (sz >= env->me_psize / MDB_MINKEYS) {
4692 /* put on overflow page */
4693 sz -= data->mv_size - sizeof(pgno_t);
4697 return sz + sizeof(indx_t);
4700 /** Calculate the size of a branch node.
4701 * The size should depend on the environment's page size but since
4702 * we currently don't support spilling large keys onto overflow
4703 * pages, it's simply the size of the #MDB_node header plus the
4704 * size of the key. Sizes are always rounded up to an even number
4705 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4706 * @param[in] env The environment handle.
4707 * @param[in] key The key for the node.
4708 * @return The number of bytes needed to store the node.
4711 mdb_branch_size(MDB_env *env, MDB_val *key)
4716 if (sz >= env->me_psize / MDB_MINKEYS) {
4717 /* put on overflow page */
4718 /* not implemented */
4719 /* sz -= key->size - sizeof(pgno_t); */
4722 return sz + sizeof(indx_t);
4725 /** Add a node to the page pointed to by the cursor.
4726 * @param[in] mc The cursor for this operation.
4727 * @param[in] indx The index on the page where the new node should be added.
4728 * @param[in] key The key for the new node.
4729 * @param[in] data The data for the new node, if any.
4730 * @param[in] pgno The page number, if adding a branch node.
4731 * @param[in] flags Flags for the node.
4732 * @return 0 on success, non-zero on failure. Possible errors are:
4734 * <li>ENOMEM - failed to allocate overflow pages for the node.
4735 * <li>ENOSPC - there is insufficient room in the page. This error
4736 * should never happen since all callers already calculate the
4737 * page's free space before calling this function.
4741 mdb_node_add(MDB_cursor *mc, indx_t indx,
4742 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4745 size_t node_size = NODESIZE;
4748 MDB_page *mp = mc->mc_pg[mc->mc_top];
4749 MDB_page *ofp = NULL; /* overflow page */
4752 assert(mp->mp_upper >= mp->mp_lower);
4754 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4755 IS_LEAF(mp) ? "leaf" : "branch",
4756 IS_SUBP(mp) ? "sub-" : "",
4757 mp->mp_pgno, indx, data ? data->mv_size : 0,
4758 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4761 /* Move higher keys up one slot. */
4762 int ksize = mc->mc_db->md_pad, dif;
4763 char *ptr = LEAF2KEY(mp, indx, ksize);
4764 dif = NUMKEYS(mp) - indx;
4766 memmove(ptr+ksize, ptr, dif*ksize);
4767 /* insert new key */
4768 memcpy(ptr, key->mv_data, ksize);
4770 /* Just using these for counting */
4771 mp->mp_lower += sizeof(indx_t);
4772 mp->mp_upper -= ksize - sizeof(indx_t);
4777 node_size += key->mv_size;
4781 if (F_ISSET(flags, F_BIGDATA)) {
4782 /* Data already on overflow page. */
4783 node_size += sizeof(pgno_t);
4784 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4785 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4786 /* Put data on overflow page. */
4787 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4788 data->mv_size, node_size+data->mv_size);
4789 node_size += sizeof(pgno_t);
4790 if ((ofp = mdb_page_new(mc, P_OVERFLOW, ovpages)) == NULL)
4792 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4795 node_size += data->mv_size;
4798 node_size += node_size & 1;
4800 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4801 DPRINTF("not enough room in page %zu, got %u ptrs",
4802 mp->mp_pgno, NUMKEYS(mp));
4803 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4804 mp->mp_upper - mp->mp_lower);
4805 DPRINTF("node size = %zu", node_size);
4809 /* Move higher pointers up one slot. */
4810 for (i = NUMKEYS(mp); i > indx; i--)
4811 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
4813 /* Adjust free space offsets. */
4814 ofs = mp->mp_upper - node_size;
4815 assert(ofs >= mp->mp_lower + sizeof(indx_t));
4816 mp->mp_ptrs[indx] = ofs;
4818 mp->mp_lower += sizeof(indx_t);
4820 /* Write the node data. */
4821 node = NODEPTR(mp, indx);
4822 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
4823 node->mn_flags = flags;
4825 SETDSZ(node,data->mv_size);
4830 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4835 if (F_ISSET(flags, F_BIGDATA))
4836 memcpy(node->mn_data + key->mv_size, data->mv_data,
4838 else if (F_ISSET(flags, MDB_RESERVE))
4839 data->mv_data = node->mn_data + key->mv_size;
4841 memcpy(node->mn_data + key->mv_size, data->mv_data,
4844 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
4846 if (F_ISSET(flags, MDB_RESERVE))
4847 data->mv_data = METADATA(ofp);
4849 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
4856 /** Delete the specified node from a page.
4857 * @param[in] mp The page to operate on.
4858 * @param[in] indx The index of the node to delete.
4859 * @param[in] ksize The size of a node. Only used if the page is
4860 * part of a #MDB_DUPFIXED database.
4863 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
4866 indx_t i, j, numkeys, ptr;
4873 COPY_PGNO(pgno, mp->mp_pgno);
4874 DPRINTF("delete node %u on %s page %zu", indx,
4875 IS_LEAF(mp) ? "leaf" : "branch", pgno);
4878 assert(indx < NUMKEYS(mp));
4881 int x = NUMKEYS(mp) - 1 - indx;
4882 base = LEAF2KEY(mp, indx, ksize);
4884 memmove(base, base + ksize, x * ksize);
4885 mp->mp_lower -= sizeof(indx_t);
4886 mp->mp_upper += ksize - sizeof(indx_t);
4890 node = NODEPTR(mp, indx);
4891 sz = NODESIZE + node->mn_ksize;
4893 if (F_ISSET(node->mn_flags, F_BIGDATA))
4894 sz += sizeof(pgno_t);
4896 sz += NODEDSZ(node);
4900 ptr = mp->mp_ptrs[indx];
4901 numkeys = NUMKEYS(mp);
4902 for (i = j = 0; i < numkeys; i++) {
4904 mp->mp_ptrs[j] = mp->mp_ptrs[i];
4905 if (mp->mp_ptrs[i] < ptr)
4906 mp->mp_ptrs[j] += sz;
4911 base = (char *)mp + mp->mp_upper;
4912 memmove(base + sz, base, ptr - mp->mp_upper);
4914 mp->mp_lower -= sizeof(indx_t);
4918 /** Compact the main page after deleting a node on a subpage.
4919 * @param[in] mp The main page to operate on.
4920 * @param[in] indx The index of the subpage on the main page.
4923 mdb_node_shrink(MDB_page *mp, indx_t indx)
4930 indx_t i, numkeys, ptr;
4932 node = NODEPTR(mp, indx);
4933 sp = (MDB_page *)NODEDATA(node);
4934 osize = NODEDSZ(node);
4936 delta = sp->mp_upper - sp->mp_lower;
4937 SETDSZ(node, osize - delta);
4938 xp = (MDB_page *)((char *)sp + delta);
4940 /* shift subpage upward */
4942 nsize = NUMKEYS(sp) * sp->mp_pad;
4943 memmove(METADATA(xp), METADATA(sp), nsize);
4946 nsize = osize - sp->mp_upper;
4947 numkeys = NUMKEYS(sp);
4948 for (i=numkeys-1; i>=0; i--)
4949 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
4951 xp->mp_upper = sp->mp_lower;
4952 xp->mp_lower = sp->mp_lower;
4953 xp->mp_flags = sp->mp_flags;
4954 xp->mp_pad = sp->mp_pad;
4955 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
4957 /* shift lower nodes upward */
4958 ptr = mp->mp_ptrs[indx];
4959 numkeys = NUMKEYS(mp);
4960 for (i = 0; i < numkeys; i++) {
4961 if (mp->mp_ptrs[i] <= ptr)
4962 mp->mp_ptrs[i] += delta;
4965 base = (char *)mp + mp->mp_upper;
4966 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
4967 mp->mp_upper += delta;
4970 /** Initial setup of a sorted-dups cursor.
4971 * Sorted duplicates are implemented as a sub-database for the given key.
4972 * The duplicate data items are actually keys of the sub-database.
4973 * Operations on the duplicate data items are performed using a sub-cursor
4974 * initialized when the sub-database is first accessed. This function does
4975 * the preliminary setup of the sub-cursor, filling in the fields that
4976 * depend only on the parent DB.
4977 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4980 mdb_xcursor_init0(MDB_cursor *mc)
4982 MDB_xcursor *mx = mc->mc_xcursor;
4984 mx->mx_cursor.mc_xcursor = NULL;
4985 mx->mx_cursor.mc_txn = mc->mc_txn;
4986 mx->mx_cursor.mc_db = &mx->mx_db;
4987 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
4988 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
4989 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
4990 mx->mx_cursor.mc_snum = 0;
4991 mx->mx_cursor.mc_top = 0;
4992 mx->mx_cursor.mc_flags = C_SUB;
4993 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
4994 mx->mx_dbx.md_dcmp = NULL;
4995 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
4998 /** Final setup of a sorted-dups cursor.
4999 * Sets up the fields that depend on the data from the main cursor.
5000 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5001 * @param[in] node The data containing the #MDB_db record for the
5002 * sorted-dup database.
5005 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5007 MDB_xcursor *mx = mc->mc_xcursor;
5009 if (node->mn_flags & F_SUBDATA) {
5010 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5011 mx->mx_cursor.mc_pg[0] = 0;
5012 mx->mx_cursor.mc_snum = 0;
5013 mx->mx_cursor.mc_flags = C_SUB;
5015 MDB_page *fp = NODEDATA(node);
5016 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5017 mx->mx_db.md_flags = 0;
5018 mx->mx_db.md_depth = 1;
5019 mx->mx_db.md_branch_pages = 0;
5020 mx->mx_db.md_leaf_pages = 1;
5021 mx->mx_db.md_overflow_pages = 0;
5022 mx->mx_db.md_entries = NUMKEYS(fp);
5023 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5024 mx->mx_cursor.mc_snum = 1;
5025 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5026 mx->mx_cursor.mc_top = 0;
5027 mx->mx_cursor.mc_pg[0] = fp;
5028 mx->mx_cursor.mc_ki[0] = 0;
5029 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5030 mx->mx_db.md_flags = MDB_DUPFIXED;
5031 mx->mx_db.md_pad = fp->mp_pad;
5032 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5033 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5036 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5038 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5040 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5041 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5042 #if UINT_MAX < SIZE_MAX
5043 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5044 #ifdef MISALIGNED_OK
5045 mx->mx_dbx.md_cmp = mdb_cmp_long;
5047 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5052 /** Initialize a cursor for a given transaction and database. */
5054 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5059 mc->mc_db = &txn->mt_dbs[dbi];
5060 mc->mc_dbx = &txn->mt_dbxs[dbi];
5061 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5066 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5068 mc->mc_xcursor = mx;
5069 mdb_xcursor_init0(mc);
5071 mc->mc_xcursor = NULL;
5073 if (*mc->mc_dbflag & DB_STALE) {
5074 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5079 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5082 MDB_xcursor *mx = NULL;
5083 size_t size = sizeof(MDB_cursor);
5085 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5088 /* Allow read access to the freelist */
5089 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5092 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5093 size += sizeof(MDB_xcursor);
5095 if ((mc = malloc(size)) != NULL) {
5096 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5097 mx = (MDB_xcursor *)(mc + 1);
5099 mdb_cursor_init(mc, txn, dbi, mx);
5100 if (txn->mt_cursors) {
5101 mc->mc_next = txn->mt_cursors[dbi];
5102 txn->mt_cursors[dbi] = mc;
5104 mc->mc_flags |= C_ALLOCD;
5114 /* Return the count of duplicate data items for the current key */
5116 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5120 if (mc == NULL || countp == NULL)
5123 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5126 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5127 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5130 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5133 *countp = mc->mc_xcursor->mx_db.md_entries;
5139 mdb_cursor_close(MDB_cursor *mc)
5142 /* remove from txn, if tracked */
5143 if (mc->mc_txn->mt_cursors) {
5144 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5145 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5147 *prev = mc->mc_next;
5149 if (mc->mc_flags & C_ALLOCD)
5155 mdb_cursor_txn(MDB_cursor *mc)
5157 if (!mc) return NULL;
5162 mdb_cursor_dbi(MDB_cursor *mc)
5168 /** Replace the key for a node with a new key.
5169 * @param[in] mp The page containing the node to operate on.
5170 * @param[in] indx The index of the node to operate on.
5171 * @param[in] key The new key to use.
5172 * @return 0 on success, non-zero on failure.
5175 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5181 indx_t ptr, i, numkeys;
5184 node = NODEPTR(mp, indx);
5185 ptr = mp->mp_ptrs[indx];
5189 char kbuf2[(MAXKEYSIZE*2+1)];
5190 k2.mv_data = NODEKEY(node);
5191 k2.mv_size = node->mn_ksize;
5192 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5194 mdb_dkey(&k2, kbuf2),
5200 delta0 = delta = key->mv_size - node->mn_ksize;
5202 /* Must be 2-byte aligned. If new key is
5203 * shorter by 1, the shift will be skipped.
5205 delta += (delta & 1);
5207 if (delta > 0 && SIZELEFT(mp) < delta) {
5208 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5212 numkeys = NUMKEYS(mp);
5213 for (i = 0; i < numkeys; i++) {
5214 if (mp->mp_ptrs[i] <= ptr)
5215 mp->mp_ptrs[i] -= delta;
5218 base = (char *)mp + mp->mp_upper;
5219 len = ptr - mp->mp_upper + NODESIZE;
5220 memmove(base - delta, base, len);
5221 mp->mp_upper -= delta;
5223 node = NODEPTR(mp, indx);
5226 /* But even if no shift was needed, update ksize */
5228 node->mn_ksize = key->mv_size;
5231 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5236 /** Move a node from csrc to cdst.
5239 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5245 unsigned short flags;
5249 /* Mark src and dst as dirty. */
5250 if ((rc = mdb_page_touch(csrc)) ||
5251 (rc = mdb_page_touch(cdst)))
5254 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5255 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5256 key.mv_size = csrc->mc_db->md_pad;
5257 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5259 data.mv_data = NULL;
5263 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5264 assert(!((long)srcnode&1));
5265 srcpg = NODEPGNO(srcnode);
5266 flags = srcnode->mn_flags;
5267 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5268 unsigned int snum = csrc->mc_snum;
5270 /* must find the lowest key below src */
5271 mdb_page_search_root(csrc, NULL, 0);
5272 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5273 key.mv_size = csrc->mc_db->md_pad;
5274 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5276 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5277 key.mv_size = NODEKSZ(s2);
5278 key.mv_data = NODEKEY(s2);
5280 csrc->mc_snum = snum--;
5281 csrc->mc_top = snum;
5283 key.mv_size = NODEKSZ(srcnode);
5284 key.mv_data = NODEKEY(srcnode);
5286 data.mv_size = NODEDSZ(srcnode);
5287 data.mv_data = NODEDATA(srcnode);
5289 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5290 unsigned int snum = cdst->mc_snum;
5293 /* must find the lowest key below dst */
5294 mdb_page_search_root(cdst, NULL, 0);
5295 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5296 bkey.mv_size = cdst->mc_db->md_pad;
5297 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5299 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5300 bkey.mv_size = NODEKSZ(s2);
5301 bkey.mv_data = NODEKEY(s2);
5303 cdst->mc_snum = snum--;
5304 cdst->mc_top = snum;
5305 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5308 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5309 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5310 csrc->mc_ki[csrc->mc_top],
5312 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5313 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5315 /* Add the node to the destination page.
5317 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5318 if (rc != MDB_SUCCESS)
5321 /* Delete the node from the source page.
5323 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5326 /* Adjust other cursors pointing to mp */
5327 MDB_cursor *m2, *m3;
5328 MDB_dbi dbi = csrc->mc_dbi;
5329 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5331 if (csrc->mc_flags & C_SUB)
5334 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5335 if (m2 == csrc) continue;
5336 if (csrc->mc_flags & C_SUB)
5337 m3 = &m2->mc_xcursor->mx_cursor;
5340 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5341 csrc->mc_ki[csrc->mc_top]) {
5342 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5343 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5348 /* Update the parent separators.
5350 if (csrc->mc_ki[csrc->mc_top] == 0) {
5351 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5352 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5353 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5355 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5356 key.mv_size = NODEKSZ(srcnode);
5357 key.mv_data = NODEKEY(srcnode);
5359 DPRINTF("update separator for source page %zu to [%s]",
5360 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5361 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5362 &key)) != MDB_SUCCESS)
5365 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5367 nullkey.mv_size = 0;
5368 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5369 assert(rc == MDB_SUCCESS);
5373 if (cdst->mc_ki[cdst->mc_top] == 0) {
5374 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5375 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5376 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5378 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5379 key.mv_size = NODEKSZ(srcnode);
5380 key.mv_data = NODEKEY(srcnode);
5382 DPRINTF("update separator for destination page %zu to [%s]",
5383 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5384 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5385 &key)) != MDB_SUCCESS)
5388 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5390 nullkey.mv_size = 0;
5391 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5392 assert(rc == MDB_SUCCESS);
5399 /** Merge one page into another.
5400 * The nodes from the page pointed to by \b csrc will
5401 * be copied to the page pointed to by \b cdst and then
5402 * the \b csrc page will be freed.
5403 * @param[in] csrc Cursor pointing to the source page.
5404 * @param[in] cdst Cursor pointing to the destination page.
5407 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5415 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5416 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5418 assert(csrc->mc_snum > 1); /* can't merge root page */
5419 assert(cdst->mc_snum > 1);
5421 /* Mark dst as dirty. */
5422 if ((rc = mdb_page_touch(cdst)))
5425 /* Move all nodes from src to dst.
5427 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5428 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5429 key.mv_size = csrc->mc_db->md_pad;
5430 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5431 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5432 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5433 if (rc != MDB_SUCCESS)
5435 key.mv_data = (char *)key.mv_data + key.mv_size;
5438 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5439 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5440 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5441 unsigned int snum = csrc->mc_snum;
5443 /* must find the lowest key below src */
5444 mdb_page_search_root(csrc, NULL, 0);
5445 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5446 key.mv_size = csrc->mc_db->md_pad;
5447 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5449 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5450 key.mv_size = NODEKSZ(s2);
5451 key.mv_data = NODEKEY(s2);
5453 csrc->mc_snum = snum--;
5454 csrc->mc_top = snum;
5456 key.mv_size = srcnode->mn_ksize;
5457 key.mv_data = NODEKEY(srcnode);
5460 data.mv_size = NODEDSZ(srcnode);
5461 data.mv_data = NODEDATA(srcnode);
5462 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5463 if (rc != MDB_SUCCESS)
5468 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5469 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);
5471 /* Unlink the src page from parent and add to free list.
5473 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5474 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5476 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5480 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5481 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5482 csrc->mc_db->md_leaf_pages--;
5484 csrc->mc_db->md_branch_pages--;
5486 /* Adjust other cursors pointing to mp */
5487 MDB_cursor *m2, *m3;
5488 MDB_dbi dbi = csrc->mc_dbi;
5489 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5491 if (csrc->mc_flags & C_SUB)
5494 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5495 if (csrc->mc_flags & C_SUB)
5496 m3 = &m2->mc_xcursor->mx_cursor;
5499 if (m3 == csrc) continue;
5500 if (m3->mc_snum < csrc->mc_snum) continue;
5501 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5502 m3->mc_pg[csrc->mc_top] = mp;
5503 m3->mc_ki[csrc->mc_top] += nkeys;
5507 mdb_cursor_pop(csrc);
5509 return mdb_rebalance(csrc);
5512 /** Copy the contents of a cursor.
5513 * @param[in] csrc The cursor to copy from.
5514 * @param[out] cdst The cursor to copy to.
5517 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5521 cdst->mc_txn = csrc->mc_txn;
5522 cdst->mc_dbi = csrc->mc_dbi;
5523 cdst->mc_db = csrc->mc_db;
5524 cdst->mc_dbx = csrc->mc_dbx;
5525 cdst->mc_snum = csrc->mc_snum;
5526 cdst->mc_top = csrc->mc_top;
5527 cdst->mc_flags = csrc->mc_flags;
5529 for (i=0; i<csrc->mc_snum; i++) {
5530 cdst->mc_pg[i] = csrc->mc_pg[i];
5531 cdst->mc_ki[i] = csrc->mc_ki[i];
5535 /** Rebalance the tree after a delete operation.
5536 * @param[in] mc Cursor pointing to the page where rebalancing
5538 * @return 0 on success, non-zero on failure.
5541 mdb_rebalance(MDB_cursor *mc)
5551 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5552 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5553 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5554 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5558 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5561 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5562 DPRINTF("no need to rebalance page %zu, above fill threshold",
5568 if (mc->mc_snum < 2) {
5569 MDB_page *mp = mc->mc_pg[0];
5570 if (NUMKEYS(mp) == 0) {
5571 DPUTS("tree is completely empty");
5572 mc->mc_db->md_root = P_INVALID;
5573 mc->mc_db->md_depth = 0;
5574 mc->mc_db->md_leaf_pages = 0;
5575 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5579 /* Adjust other cursors pointing to mp */
5580 MDB_cursor *m2, *m3;
5581 MDB_dbi dbi = mc->mc_dbi;
5583 if (mc->mc_flags & C_SUB)
5586 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5587 if (m2 == mc) continue;
5588 if (mc->mc_flags & C_SUB)
5589 m3 = &m2->mc_xcursor->mx_cursor;
5592 if (m3->mc_snum < mc->mc_snum) continue;
5593 if (m3->mc_pg[0] == mp) {
5599 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5600 DPUTS("collapsing root page!");
5601 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5602 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5603 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5606 mc->mc_db->md_depth--;
5607 mc->mc_db->md_branch_pages--;
5609 /* Adjust other cursors pointing to mp */
5610 MDB_cursor *m2, *m3;
5611 MDB_dbi dbi = mc->mc_dbi;
5613 if (mc->mc_flags & C_SUB)
5616 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5617 if (m2 == mc) continue;
5618 if (mc->mc_flags & C_SUB)
5619 m3 = &m2->mc_xcursor->mx_cursor;
5622 if (m3->mc_snum < mc->mc_snum) continue;
5623 if (m3->mc_pg[0] == mp) {
5624 m3->mc_pg[0] = mc->mc_pg[0];
5629 DPUTS("root page doesn't need rebalancing");
5633 /* The parent (branch page) must have at least 2 pointers,
5634 * otherwise the tree is invalid.
5636 ptop = mc->mc_top-1;
5637 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5639 /* Leaf page fill factor is below the threshold.
5640 * Try to move keys from left or right neighbor, or
5641 * merge with a neighbor page.
5646 mdb_cursor_copy(mc, &mn);
5647 mn.mc_xcursor = NULL;
5649 if (mc->mc_ki[ptop] == 0) {
5650 /* We're the leftmost leaf in our parent.
5652 DPUTS("reading right neighbor");
5654 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5655 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5657 mn.mc_ki[mn.mc_top] = 0;
5658 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5660 /* There is at least one neighbor to the left.
5662 DPUTS("reading left neighbor");
5664 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5665 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5667 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5668 mc->mc_ki[mc->mc_top] = 0;
5671 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5672 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);
5674 /* If the neighbor page is above threshold and has at least two
5675 * keys, move one key from it.
5677 * Otherwise we should try to merge them.
5679 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5680 return mdb_node_move(&mn, mc);
5681 else { /* FIXME: if (has_enough_room()) */
5682 mc->mc_flags &= ~C_INITIALIZED;
5683 if (mc->mc_ki[ptop] == 0)
5684 return mdb_page_merge(&mn, mc);
5686 return mdb_page_merge(mc, &mn);
5690 /** Complete a delete operation started by #mdb_cursor_del(). */
5692 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5696 /* add overflow pages to free list */
5697 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5701 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5702 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5703 mc->mc_db->md_overflow_pages -= ovpages;
5704 for (i=0; i<ovpages; i++) {
5705 DPRINTF("freed ov page %zu", pg);
5706 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5710 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5711 mc->mc_db->md_entries--;
5712 rc = mdb_rebalance(mc);
5713 if (rc != MDB_SUCCESS)
5714 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5720 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5721 MDB_val *key, MDB_val *data)
5726 MDB_val rdata, *xdata;
5730 assert(key != NULL);
5732 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5734 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5737 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5741 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5745 mdb_cursor_init(&mc, txn, dbi, &mx);
5756 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5758 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5762 /** Split a page and insert a new node.
5763 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5764 * The cursor will be updated to point to the actual page and index where
5765 * the node got inserted after the split.
5766 * @param[in] newkey The key for the newly inserted node.
5767 * @param[in] newdata The data for the newly inserted node.
5768 * @param[in] newpgno The page number, if the new node is a branch node.
5769 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
5770 * @return 0 on success, non-zero on failure.
5773 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5774 unsigned int nflags)
5777 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
5780 unsigned int i, j, split_indx, nkeys, pmax;
5782 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5784 MDB_page *mp, *rp, *pp;
5789 mp = mc->mc_pg[mc->mc_top];
5790 newindx = mc->mc_ki[mc->mc_top];
5792 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
5793 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
5794 DKEY(newkey), mc->mc_ki[mc->mc_top]);
5796 /* Create a right sibling. */
5797 if ((rp = mdb_page_new(mc, mp->mp_flags, 1)) == NULL)
5799 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
5801 if (mc->mc_snum < 2) {
5802 if ((pp = mdb_page_new(mc, P_BRANCH, 1)) == NULL)
5804 /* shift current top to make room for new parent */
5805 mc->mc_pg[1] = mc->mc_pg[0];
5806 mc->mc_ki[1] = mc->mc_ki[0];
5809 mc->mc_db->md_root = pp->mp_pgno;
5810 DPRINTF("root split! new root = %zu", pp->mp_pgno);
5811 mc->mc_db->md_depth++;
5814 /* Add left (implicit) pointer. */
5815 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
5816 /* undo the pre-push */
5817 mc->mc_pg[0] = mc->mc_pg[1];
5818 mc->mc_ki[0] = mc->mc_ki[1];
5819 mc->mc_db->md_root = mp->mp_pgno;
5820 mc->mc_db->md_depth--;
5827 ptop = mc->mc_top-1;
5828 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
5831 mc->mc_flags |= C_SPLITTING;
5832 mdb_cursor_copy(mc, &mn);
5833 mn.mc_pg[mn.mc_top] = rp;
5834 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
5836 if (nflags & MDB_APPEND) {
5837 mn.mc_ki[mn.mc_top] = 0;
5839 split_indx = newindx;
5844 nkeys = NUMKEYS(mp);
5845 split_indx = (nkeys + 1) / 2;
5846 if (newindx < split_indx)
5852 unsigned int lsize, rsize, ksize;
5853 /* Move half of the keys to the right sibling */
5855 x = mc->mc_ki[mc->mc_top] - split_indx;
5856 ksize = mc->mc_db->md_pad;
5857 split = LEAF2KEY(mp, split_indx, ksize);
5858 rsize = (nkeys - split_indx) * ksize;
5859 lsize = (nkeys - split_indx) * sizeof(indx_t);
5860 mp->mp_lower -= lsize;
5861 rp->mp_lower += lsize;
5862 mp->mp_upper += rsize - lsize;
5863 rp->mp_upper -= rsize - lsize;
5864 sepkey.mv_size = ksize;
5865 if (newindx == split_indx) {
5866 sepkey.mv_data = newkey->mv_data;
5868 sepkey.mv_data = split;
5871 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
5872 memcpy(rp->mp_ptrs, split, rsize);
5873 sepkey.mv_data = rp->mp_ptrs;
5874 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
5875 memcpy(ins, newkey->mv_data, ksize);
5876 mp->mp_lower += sizeof(indx_t);
5877 mp->mp_upper -= ksize - sizeof(indx_t);
5880 memcpy(rp->mp_ptrs, split, x * ksize);
5881 ins = LEAF2KEY(rp, x, ksize);
5882 memcpy(ins, newkey->mv_data, ksize);
5883 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
5884 rp->mp_lower += sizeof(indx_t);
5885 rp->mp_upper -= ksize - sizeof(indx_t);
5886 mc->mc_ki[mc->mc_top] = x;
5887 mc->mc_pg[mc->mc_top] = rp;
5892 /* For leaf pages, check the split point based on what
5893 * fits where, since otherwise mdb_node_add can fail.
5895 * This check is only needed when the data items are
5896 * relatively large, such that being off by one will
5897 * make the difference between success or failure.
5898 * When the size of the data items is much smaller than
5899 * one-half of a page, this check is irrelevant.
5902 unsigned int psize, nsize;
5903 /* Maximum free space in an empty page */
5904 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
5905 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
5906 if ((nkeys < 20) || (nsize > pmax/4)) {
5907 if (newindx <= split_indx) {
5910 for (i=0; i<split_indx; i++) {
5911 node = NODEPTR(mp, i);
5912 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5913 if (F_ISSET(node->mn_flags, F_BIGDATA))
5914 psize += sizeof(pgno_t);
5916 psize += NODEDSZ(node);
5919 if (i == split_indx - 1 && newindx == split_indx)
5928 for (i=nkeys-1; i>=split_indx; i--) {
5929 node = NODEPTR(mp, i);
5930 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5931 if (F_ISSET(node->mn_flags, F_BIGDATA))
5932 psize += sizeof(pgno_t);
5934 psize += NODEDSZ(node);
5945 /* First find the separating key between the split pages.
5946 * The case where newindx == split_indx is ambiguous; the
5947 * new item could go to the new page or stay on the original
5948 * page. If newpos == 1 it goes to the new page.
5950 if (newindx == split_indx && newpos) {
5951 sepkey.mv_size = newkey->mv_size;
5952 sepkey.mv_data = newkey->mv_data;
5954 node = NODEPTR(mp, split_indx);
5955 sepkey.mv_size = node->mn_ksize;
5956 sepkey.mv_data = NODEKEY(node);
5960 DPRINTF("separator is [%s]", DKEY(&sepkey));
5962 /* Copy separator key to the parent.
5964 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
5968 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
5971 if (mn.mc_snum == mc->mc_snum) {
5972 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
5973 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
5974 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
5975 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
5980 /* Right page might now have changed parent.
5981 * Check if left page also changed parent.
5983 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
5984 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
5985 for (i=0; i<ptop; i++) {
5986 mc->mc_pg[i] = mn.mc_pg[i];
5987 mc->mc_ki[i] = mn.mc_ki[i];
5989 mc->mc_pg[ptop] = mn.mc_pg[ptop];
5990 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
5994 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
5997 mc->mc_flags ^= C_SPLITTING;
5998 if (rc != MDB_SUCCESS) {
6001 if (nflags & MDB_APPEND) {
6002 mc->mc_pg[mc->mc_top] = rp;
6003 mc->mc_ki[mc->mc_top] = 0;
6004 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6007 for (i=0; i<mc->mc_top; i++)
6008 mc->mc_ki[i] = mn.mc_ki[i];
6015 /* Move half of the keys to the right sibling. */
6017 /* grab a page to hold a temporary copy */
6018 copy = mdb_page_malloc(mc);
6022 copy->mp_pgno = mp->mp_pgno;
6023 copy->mp_flags = mp->mp_flags;
6024 copy->mp_lower = PAGEHDRSZ;
6025 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6026 mc->mc_pg[mc->mc_top] = copy;
6027 for (i = j = 0; i <= nkeys; j++) {
6028 if (i == split_indx) {
6029 /* Insert in right sibling. */
6030 /* Reset insert index for right sibling. */
6031 if (i != newindx || (newpos ^ ins_new)) {
6033 mc->mc_pg[mc->mc_top] = rp;
6037 if (i == newindx && !ins_new) {
6038 /* Insert the original entry that caused the split. */
6039 rkey.mv_data = newkey->mv_data;
6040 rkey.mv_size = newkey->mv_size;
6049 /* Update index for the new key. */
6050 mc->mc_ki[mc->mc_top] = j;
6051 } else if (i == nkeys) {
6054 node = NODEPTR(mp, i);
6055 rkey.mv_data = NODEKEY(node);
6056 rkey.mv_size = node->mn_ksize;
6058 xdata.mv_data = NODEDATA(node);
6059 xdata.mv_size = NODEDSZ(node);
6062 pgno = NODEPGNO(node);
6063 flags = node->mn_flags;
6068 if (!IS_LEAF(mp) && j == 0) {
6069 /* First branch index doesn't need key data. */
6073 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6077 nkeys = NUMKEYS(copy);
6078 for (i=0; i<nkeys; i++)
6079 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6080 mp->mp_lower = copy->mp_lower;
6081 mp->mp_upper = copy->mp_upper;
6082 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6083 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6085 /* reset back to original page */
6086 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6087 mc->mc_pg[mc->mc_top] = mp;
6088 if (nflags & MDB_RESERVE) {
6089 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6090 if (!(node->mn_flags & F_BIGDATA))
6091 newdata->mv_data = NODEDATA(node);
6097 /* return tmp page to freelist */
6098 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
6099 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
6100 mc->mc_txn->mt_env->me_dpages = copy;
6103 /* Adjust other cursors pointing to mp */
6104 MDB_cursor *m2, *m3;
6105 MDB_dbi dbi = mc->mc_dbi;
6106 int fixup = NUMKEYS(mp);
6108 if (mc->mc_flags & C_SUB)
6111 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6112 if (m2 == mc) continue;
6113 if (mc->mc_flags & C_SUB)
6114 m3 = &m2->mc_xcursor->mx_cursor;
6117 if (!(m3->mc_flags & C_INITIALIZED))
6119 if (m3->mc_flags & C_SPLITTING)
6124 for (k=m3->mc_top; k>=0; k--) {
6125 m3->mc_ki[k+1] = m3->mc_ki[k];
6126 m3->mc_pg[k+1] = m3->mc_pg[k];
6128 if (m3->mc_ki[0] >= split_indx) {
6133 m3->mc_pg[0] = mc->mc_pg[0];
6137 if (m3->mc_pg[mc->mc_top] == mp) {
6138 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6139 m3->mc_ki[mc->mc_top]++;
6140 if (m3->mc_ki[mc->mc_top] >= fixup) {
6141 m3->mc_pg[mc->mc_top] = rp;
6142 m3->mc_ki[mc->mc_top] -= fixup;
6143 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6145 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6146 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6155 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6156 MDB_val *key, MDB_val *data, unsigned int flags)
6161 assert(key != NULL);
6162 assert(data != NULL);
6164 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6167 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6171 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6175 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6178 mdb_cursor_init(&mc, txn, dbi, &mx);
6179 return mdb_cursor_put(&mc, key, data, flags);
6182 /** Only a subset of the @ref mdb_env flags can be changed
6183 * at runtime. Changing other flags requires closing the environment
6184 * and re-opening it with the new flags.
6186 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC)
6188 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6190 if ((flag & CHANGEABLE) != flag)
6193 env->me_flags |= flag;
6195 env->me_flags &= ~flag;
6200 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6205 *arg = env->me_flags;
6210 mdb_env_get_path(MDB_env *env, const char **arg)
6215 *arg = env->me_path;
6219 /** Common code for #mdb_stat() and #mdb_env_stat().
6220 * @param[in] env the environment to operate in.
6221 * @param[in] db the #MDB_db record containing the stats to return.
6222 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6223 * @return 0, this function always succeeds.
6226 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6228 arg->ms_psize = env->me_psize;
6229 arg->ms_depth = db->md_depth;
6230 arg->ms_branch_pages = db->md_branch_pages;
6231 arg->ms_leaf_pages = db->md_leaf_pages;
6232 arg->ms_overflow_pages = db->md_overflow_pages;
6233 arg->ms_entries = db->md_entries;
6238 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6242 if (env == NULL || arg == NULL)
6245 toggle = mdb_env_pick_meta(env);
6247 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6250 /** Set the default comparison functions for a database.
6251 * Called immediately after a database is opened to set the defaults.
6252 * The user can then override them with #mdb_set_compare() or
6253 * #mdb_set_dupsort().
6254 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6255 * @param[in] dbi A database handle returned by #mdb_open()
6258 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6260 uint16_t f = txn->mt_dbs[dbi].md_flags;
6262 txn->mt_dbxs[dbi].md_cmp =
6263 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6264 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6266 txn->mt_dbxs[dbi].md_dcmp =
6267 !(f & MDB_DUPSORT) ? 0 :
6268 ((f & MDB_INTEGERDUP)
6269 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6270 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6273 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6278 int rc, dbflag, exact;
6279 unsigned int unused = 0;
6282 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6283 mdb_default_cmp(txn, FREE_DBI);
6289 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
6290 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
6291 mdb_default_cmp(txn, MAIN_DBI);
6295 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6296 mdb_default_cmp(txn, MAIN_DBI);
6299 /* Is the DB already open? */
6301 for (i=2; i<txn->mt_numdbs; i++) {
6302 if (!txn->mt_dbxs[i].md_name.mv_size) {
6303 /* Remember this free slot */
6304 if (!unused) unused = i;
6307 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6308 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6314 /* If no free slot and max hit, fail */
6315 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
6318 /* Find the DB info */
6322 key.mv_data = (void *)name;
6323 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6324 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6325 if (rc == MDB_SUCCESS) {
6326 /* make sure this is actually a DB */
6327 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6328 if (!(node->mn_flags & F_SUBDATA))
6330 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6331 /* Create if requested */
6333 data.mv_size = sizeof(MDB_db);
6334 data.mv_data = &dummy;
6335 memset(&dummy, 0, sizeof(dummy));
6336 dummy.md_root = P_INVALID;
6337 dummy.md_flags = flags & 0xffff;
6338 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6342 /* OK, got info, add to table */
6343 if (rc == MDB_SUCCESS) {
6344 unsigned int slot = unused ? unused : txn->mt_numdbs;
6345 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
6346 txn->mt_dbxs[slot].md_name.mv_size = len;
6347 txn->mt_dbxs[slot].md_rel = NULL;
6348 txn->mt_dbflags[slot] = dbflag;
6349 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
6351 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
6352 mdb_default_cmp(txn, slot);
6355 txn->mt_env->me_numdbs++;
6362 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6364 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6367 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6370 void mdb_close(MDB_env *env, MDB_dbi dbi)
6373 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6375 ptr = env->me_dbxs[dbi].md_name.mv_data;
6376 env->me_dbxs[dbi].md_name.mv_data = NULL;
6377 env->me_dbxs[dbi].md_name.mv_size = 0;
6381 /** Add all the DB's pages to the free list.
6382 * @param[in] mc Cursor on the DB to free.
6383 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6384 * @return 0 on success, non-zero on failure.
6387 mdb_drop0(MDB_cursor *mc, int subs)
6391 rc = mdb_page_search(mc, NULL, 0);
6392 if (rc == MDB_SUCCESS) {
6397 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6398 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6401 mdb_cursor_copy(mc, &mx);
6402 while (mc->mc_snum > 0) {
6403 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6404 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6405 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6406 if (ni->mn_flags & F_SUBDATA) {
6407 mdb_xcursor_init1(mc, ni);
6408 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6414 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6416 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6419 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6424 rc = mdb_cursor_sibling(mc, 1);
6426 /* no more siblings, go back to beginning
6427 * of previous level. (stack was already popped
6428 * by mdb_cursor_sibling)
6430 for (i=1; i<mc->mc_top; i++)
6431 mc->mc_pg[i] = mx.mc_pg[i];
6435 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6436 mc->mc_db->md_root);
6441 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6446 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6449 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6452 rc = mdb_cursor_open(txn, dbi, &mc);
6456 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6460 /* Can't delete the main DB */
6461 if (del && dbi > MAIN_DBI) {
6462 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6464 mdb_close(txn->mt_env, dbi);
6466 txn->mt_dbflags[dbi] |= DB_DIRTY;
6467 txn->mt_dbs[dbi].md_depth = 0;
6468 txn->mt_dbs[dbi].md_branch_pages = 0;
6469 txn->mt_dbs[dbi].md_leaf_pages = 0;
6470 txn->mt_dbs[dbi].md_overflow_pages = 0;
6471 txn->mt_dbs[dbi].md_entries = 0;
6472 txn->mt_dbs[dbi].md_root = P_INVALID;
6475 mdb_cursor_close(mc);
6479 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6481 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6484 txn->mt_dbxs[dbi].md_cmp = cmp;
6488 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6490 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6493 txn->mt_dbxs[dbi].md_dcmp = cmp;
6497 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6499 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6502 txn->mt_dbxs[dbi].md_rel = rel;
6506 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6508 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6511 txn->mt_dbxs[dbi].md_relctx = ctx;