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 */
847 unsigned int mc_flags; /**< @ref mdb_cursor */
848 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
849 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
852 /** Context for sorted-dup records.
853 * We could have gone to a fully recursive design, with arbitrarily
854 * deep nesting of sub-databases. But for now we only handle these
855 * levels - main DB, optional sub-DB, sorted-duplicate DB.
857 typedef struct MDB_xcursor {
858 /** A sub-cursor for traversing the Dup DB */
859 MDB_cursor mx_cursor;
860 /** The database record for this Dup DB */
862 /** The auxiliary DB record for this Dup DB */
864 /** The @ref mt_dbflag for this Dup DB */
865 unsigned char mx_dbflag;
868 /** A set of pages freed by an earlier transaction. */
869 typedef struct MDB_oldpages {
870 /** Usually we only read one record from the FREEDB at a time, but
871 * in case we read more, this will chain them together.
873 struct MDB_oldpages *mo_next;
874 /** The ID of the transaction in which these pages were freed. */
876 /** An #MDB_IDL of the pages */
877 pgno_t mo_pages[1]; /* dynamic */
880 /** The database environment. */
882 HANDLE me_fd; /**< The main data file */
883 HANDLE me_lfd; /**< The lock file */
884 HANDLE me_mfd; /**< just for writing the meta pages */
885 /** Failed to update the meta page. Probably an I/O error. */
886 #define MDB_FATAL_ERROR 0x80000000U
887 uint32_t me_flags; /**< @ref mdb_env */
888 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
889 unsigned int me_maxreaders; /**< size of the reader table */
890 MDB_dbi me_numdbs; /**< number of DBs opened */
891 MDB_dbi me_maxdbs; /**< size of the DB table */
892 char *me_path; /**< path to the DB files */
893 char *me_map; /**< the memory map of the data file */
894 MDB_txninfo *me_txns; /**< the memory map of the lock file */
895 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
896 MDB_txn *me_txn; /**< current write transaction */
897 size_t me_mapsize; /**< size of the data memory map */
898 off_t me_size; /**< current file size */
899 pgno_t me_maxpg; /**< me_mapsize / me_psize */
900 txnid_t me_pgfirst; /**< ID of first old page record we used */
901 txnid_t me_pglast; /**< ID of last old page record we used */
902 MDB_dbx *me_dbxs; /**< array of static DB info */
903 uint16_t *me_dbflags; /**< array of DB flags */
904 MDB_oldpages *me_pghead; /**< list of old page records */
905 MDB_oldpages *me_pgfree; /**< list of page records to free */
906 pthread_key_t me_txkey; /**< thread-key for readers */
907 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
908 /** IDL of pages that became unused in a write txn */
910 /** ID2L of pages that were written during a write txn */
911 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
913 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
917 sem_t *me_rmutex; /* Apple doesn't support shared mutexes */
921 /** max number of pages to commit in one writev() call */
922 #define MDB_COMMIT_PAGES 64
923 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
924 #undef MDB_COMMIT_PAGES
925 #define MDB_COMMIT_PAGES IOV_MAX
928 static MDB_page *mdb_page_alloc(MDB_cursor *mc, int num);
929 static MDB_page *mdb_page_new(MDB_cursor *mc, uint32_t flags, int num);
930 static int mdb_page_touch(MDB_cursor *mc);
932 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
933 static int mdb_page_search_root(MDB_cursor *mc,
934 MDB_val *key, int modify);
935 #define MDB_PS_MODIFY 1
936 #define MDB_PS_ROOTONLY 2
937 static int mdb_page_search(MDB_cursor *mc,
938 MDB_val *key, int flags);
939 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
940 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
941 pgno_t newpgno, unsigned int nflags);
943 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
944 static int mdb_env_pick_meta(const MDB_env *env);
945 static int mdb_env_write_meta(MDB_txn *txn);
947 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
948 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
949 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
950 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
951 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
952 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
953 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
954 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
955 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
957 static int mdb_rebalance(MDB_cursor *mc);
958 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
960 static void mdb_cursor_pop(MDB_cursor *mc);
961 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
963 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
964 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
965 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
966 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
967 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
969 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
970 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
972 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
973 static void mdb_xcursor_init0(MDB_cursor *mc);
974 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
976 static int mdb_drop0(MDB_cursor *mc, int subs);
977 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
980 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
984 static SECURITY_DESCRIPTOR mdb_null_sd;
985 static SECURITY_ATTRIBUTES mdb_all_sa;
986 static int mdb_sec_inited;
989 /** Return the library version info. */
991 mdb_version(int *major, int *minor, int *patch)
993 if (major) *major = MDB_VERSION_MAJOR;
994 if (minor) *minor = MDB_VERSION_MINOR;
995 if (patch) *patch = MDB_VERSION_PATCH;
996 return MDB_VERSION_STRING;
999 /** Table of descriptions for MDB @ref errors */
1000 static char *const mdb_errstr[] = {
1001 "MDB_KEYEXIST: Key/data pair already exists",
1002 "MDB_NOTFOUND: No matching key/data pair found",
1003 "MDB_PAGE_NOTFOUND: Requested page not found",
1004 "MDB_CORRUPTED: Located page was wrong type",
1005 "MDB_PANIC: Update of meta page failed",
1006 "MDB_VERSION_MISMATCH: Database environment version mismatch"
1010 mdb_strerror(int err)
1013 return ("Successful return: 0");
1015 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
1016 return mdb_errstr[err - MDB_KEYEXIST];
1018 return strerror(err);
1022 /** Display a key in hexadecimal and return the address of the result.
1023 * @param[in] key the key to display
1024 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1025 * @return The key in hexadecimal form.
1028 mdb_dkey(MDB_val *key, char *buf)
1031 unsigned char *c = key->mv_data;
1033 if (key->mv_size > MAXKEYSIZE)
1034 return "MAXKEYSIZE";
1035 /* may want to make this a dynamic check: if the key is mostly
1036 * printable characters, print it as-is instead of converting to hex.
1040 for (i=0; i<key->mv_size; i++)
1041 ptr += sprintf(ptr, "%02x", *c++);
1043 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1048 /** Display all the keys in the page. */
1050 mdb_page_keys(MDB_page *mp)
1053 unsigned int i, nkeys;
1057 nkeys = NUMKEYS(mp);
1058 DPRINTF("numkeys %d", nkeys);
1059 for (i=0; i<nkeys; i++) {
1060 node = NODEPTR(mp, i);
1061 key.mv_size = node->mn_ksize;
1062 key.mv_data = node->mn_data;
1063 DPRINTF("key %d: %s", i, DKEY(&key));
1069 /** Count all the pages in each DB and in the freelist
1070 * and make sure it matches the actual number of pages
1073 static void mdb_audit(MDB_txn *txn)
1077 MDB_ID freecount, count;
1082 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1083 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1084 freecount += *(MDB_ID *)data.mv_data;
1085 freecount += txn->mt_dbs[0].md_branch_pages + txn->mt_dbs[0].md_leaf_pages +
1086 txn->mt_dbs[0].md_overflow_pages;
1089 for (i = 0; i<txn->mt_numdbs; i++) {
1090 count += txn->mt_dbs[i].md_branch_pages +
1091 txn->mt_dbs[i].md_leaf_pages +
1092 txn->mt_dbs[i].md_overflow_pages;
1093 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1095 mdb_cursor_init(&mc, txn, i, &mx);
1096 mdb_page_search(&mc, NULL, 0);
1100 mp = mc.mc_pg[mc.mc_top];
1101 for (j=0; j<NUMKEYS(mp); j++) {
1102 MDB_node *leaf = NODEPTR(mp, j);
1103 if (leaf->mn_flags & F_SUBDATA) {
1105 memcpy(&db, NODEDATA(leaf), sizeof(db));
1106 count += db.md_branch_pages + db.md_leaf_pages +
1107 db.md_overflow_pages;
1111 while (mdb_cursor_sibling(&mc, 1) == 0);
1114 assert(freecount + count + 2 >= txn->mt_next_pgno - 1);
1119 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1121 return txn->mt_dbxs[dbi].md_cmp(a, b);
1125 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1127 if (txn->mt_dbxs[dbi].md_dcmp)
1128 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1130 return EINVAL; /* too bad you can't distinguish this from a valid result */
1133 /** Allocate a single page.
1134 * Re-use old malloc'd pages first, otherwise just malloc.
1137 mdb_page_malloc(MDB_cursor *mc) {
1139 size_t sz = mc->mc_txn->mt_env->me_psize;
1140 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1141 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1142 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1143 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1144 } else if ((ret = malloc(sz)) != NULL) {
1145 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1150 /** Allocate pages for writing.
1151 * If there are free pages available from older transactions, they
1152 * will be re-used first. Otherwise a new page will be allocated.
1153 * @param[in] mc cursor A cursor handle identifying the transaction and
1154 * database for which we are allocating.
1155 * @param[in] num the number of pages to allocate.
1156 * @return Address of the allocated page(s). Requests for multiple pages
1157 * will always be satisfied by a single contiguous chunk of memory.
1160 mdb_page_alloc(MDB_cursor *mc, int num)
1162 MDB_txn *txn = mc->mc_txn;
1164 pgno_t pgno = P_INVALID;
1167 /* The free list won't have any content at all until txn 2 has
1168 * committed. The pages freed by txn 2 will be unreferenced
1169 * after txn 3 commits, and so will be safe to re-use in txn 4.
1171 if (txn->mt_txnid > 3) {
1173 if (!txn->mt_env->me_pghead &&
1174 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1175 /* See if there's anything in the free DB */
1179 txnid_t *kptr, oldest, last;
1181 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1182 if (!txn->mt_env->me_pgfirst) {
1183 mdb_page_search(&m2, NULL, 0);
1184 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1185 kptr = (txnid_t *)NODEKEY(leaf);
1192 last = txn->mt_env->me_pglast + 1;
1194 key.mv_data = &last;
1195 key.mv_size = sizeof(last);
1196 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1199 last = *(txnid_t *)key.mv_data;
1204 oldest = txn->mt_txnid - 1;
1205 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
1206 txnid_t mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
1207 if (mr && mr < oldest)
1212 if (oldest > last) {
1213 /* It's usable, grab it.
1218 if (!txn->mt_env->me_pgfirst) {
1219 mdb_node_read(txn, leaf, &data);
1221 txn->mt_env->me_pglast = last;
1222 if (!txn->mt_env->me_pgfirst)
1223 txn->mt_env->me_pgfirst = last;
1224 idl = (MDB_ID *) data.mv_data;
1225 /* We might have a zero-length IDL due to freelist growth
1226 * during a prior commit
1228 if (!idl[0]) goto again;
1229 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1230 mop->mo_next = txn->mt_env->me_pghead;
1231 mop->mo_txnid = last;
1232 txn->mt_env->me_pghead = mop;
1233 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1238 DPRINTF("IDL read txn %zu root %zu num %zu",
1239 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1240 for (i=0; i<idl[0]; i++) {
1241 DPRINTF("IDL %zu", idl[i+1]);
1248 if (txn->mt_env->me_pghead) {
1249 MDB_oldpages *mop = txn->mt_env->me_pghead;
1251 /* FIXME: For now, always use fresh pages. We
1252 * really ought to search the free list for a
1257 /* peel pages off tail, so we only have to truncate the list */
1258 pgno = MDB_IDL_LAST(mop->mo_pages);
1259 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1261 if (mop->mo_pages[2] > mop->mo_pages[1])
1262 mop->mo_pages[0] = 0;
1266 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1267 txn->mt_env->me_pghead = mop->mo_next;
1268 if (mc->mc_dbi == FREE_DBI) {
1269 mop->mo_next = txn->mt_env->me_pgfree;
1270 txn->mt_env->me_pgfree = mop;
1279 if (pgno == P_INVALID) {
1280 /* DB size is maxed out */
1281 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1282 DPUTS("DB size maxed out");
1286 if (txn->mt_env->me_dpages && num == 1) {
1287 np = txn->mt_env->me_dpages;
1288 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1289 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1290 txn->mt_env->me_dpages = np->mp_next;
1292 size_t sz = txn->mt_env->me_psize * num;
1293 if ((np = malloc(sz)) == NULL)
1295 VGMEMP_ALLOC(txn->mt_env, np, sz);
1297 if (pgno == P_INVALID) {
1298 np->mp_pgno = txn->mt_next_pgno;
1299 txn->mt_next_pgno += num;
1303 mid.mid = np->mp_pgno;
1305 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1310 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1311 * @param[in] mc cursor pointing to the page to be touched
1312 * @return 0 on success, non-zero on failure.
1315 mdb_page_touch(MDB_cursor *mc)
1317 MDB_page *mp = mc->mc_pg[mc->mc_top];
1320 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1322 if ((np = mdb_page_alloc(mc, 1)) == NULL)
1324 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1325 assert(mp->mp_pgno != np->mp_pgno);
1326 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1328 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1331 mp->mp_flags |= P_DIRTY;
1334 /* Adjust other cursors pointing to mp */
1335 if (mc->mc_flags & C_SUB) {
1336 MDB_cursor *m2, *m3;
1337 MDB_dbi dbi = mc->mc_dbi-1;
1339 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1340 if (m2 == mc) continue;
1341 m3 = &m2->mc_xcursor->mx_cursor;
1342 if (m3->mc_snum < mc->mc_snum) continue;
1343 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1344 m3->mc_pg[mc->mc_top] = mp;
1350 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1351 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1352 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1353 m2->mc_pg[mc->mc_top] = mp;
1357 mc->mc_pg[mc->mc_top] = mp;
1358 /** If this page has a parent, update the parent to point to
1362 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1364 mc->mc_db->md_root = mp->mp_pgno;
1365 } else if (mc->mc_txn->mt_parent) {
1368 /* If txn has a parent, make sure the page is in our
1371 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1372 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1373 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1374 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1375 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1376 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1377 mc->mc_pg[mc->mc_top] = mp;
1383 np = mdb_page_malloc(mc);
1384 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1385 mid.mid = np->mp_pgno;
1387 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1395 mdb_env_sync(MDB_env *env, int force)
1398 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1399 if (MDB_FDATASYNC(env->me_fd))
1405 /** Make shadow copies of all of parent txn's cursors */
1407 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1409 MDB_cursor *mc, *m2;
1410 unsigned int i, j, size;
1412 for (i=0;i<src->mt_numdbs; i++) {
1413 if (src->mt_cursors[i]) {
1414 size = sizeof(MDB_cursor);
1415 if (src->mt_cursors[i]->mc_xcursor)
1416 size += sizeof(MDB_xcursor);
1417 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1424 mc->mc_db = &dst->mt_dbs[i];
1425 mc->mc_dbx = m2->mc_dbx;
1426 mc->mc_dbflag = &dst->mt_dbflags[i];
1427 mc->mc_snum = m2->mc_snum;
1428 mc->mc_top = m2->mc_top;
1429 mc->mc_flags = m2->mc_flags | C_SHADOW;
1430 for (j=0; j<mc->mc_snum; j++) {
1431 mc->mc_pg[j] = m2->mc_pg[j];
1432 mc->mc_ki[j] = m2->mc_ki[j];
1434 if (m2->mc_xcursor) {
1435 MDB_xcursor *mx, *mx2;
1436 mx = (MDB_xcursor *)(mc+1);
1437 mc->mc_xcursor = mx;
1438 mx2 = m2->mc_xcursor;
1439 mx->mx_db = mx2->mx_db;
1440 mx->mx_dbx = mx2->mx_dbx;
1441 mx->mx_dbflag = mx2->mx_dbflag;
1442 mx->mx_cursor.mc_txn = dst;
1443 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1444 mx->mx_cursor.mc_db = &mx->mx_db;
1445 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1446 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1447 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1448 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1449 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1450 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1451 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1452 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1455 mc->mc_xcursor = NULL;
1457 mc->mc_next = dst->mt_cursors[i];
1458 dst->mt_cursors[i] = mc;
1465 /** Merge shadow cursors back into parent's */
1467 mdb_cursor_merge(MDB_txn *txn)
1470 for (i=0; i<txn->mt_numdbs; i++) {
1471 if (txn->mt_cursors[i]) {
1473 while ((mc = txn->mt_cursors[i])) {
1474 txn->mt_cursors[i] = mc->mc_next;
1475 if (mc->mc_flags & C_SHADOW) {
1476 MDB_cursor *m2 = mc->mc_orig;
1478 m2->mc_snum = mc->mc_snum;
1479 m2->mc_top = mc->mc_top;
1480 for (j=0; j<mc->mc_snum; j++) {
1481 m2->mc_pg[j] = mc->mc_pg[j];
1482 m2->mc_ki[j] = mc->mc_ki[j];
1485 if (mc->mc_flags & C_ALLOCD)
1493 mdb_txn_reset0(MDB_txn *txn);
1495 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1496 * @param[in] txn the transaction handle to initialize
1497 * @return 0 on success, non-zero on failure. This can only
1498 * fail for read-only transactions, and then only if the
1499 * reader table is full.
1502 mdb_txn_renew0(MDB_txn *txn)
1504 MDB_env *env = txn->mt_env;
1508 txn->mt_numdbs = env->me_numdbs;
1509 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1511 if (txn->mt_flags & MDB_TXN_RDONLY) {
1512 MDB_reader *r = pthread_getspecific(env->me_txkey);
1514 pid_t pid = getpid();
1515 pthread_t tid = pthread_self();
1518 for (i=0; i<env->me_txns->mti_numreaders; i++)
1519 if (env->me_txns->mti_readers[i].mr_pid == 0)
1521 if (i == env->me_maxreaders) {
1522 UNLOCK_MUTEX_R(env);
1525 env->me_txns->mti_readers[i].mr_pid = pid;
1526 env->me_txns->mti_readers[i].mr_tid = tid;
1527 if (i >= env->me_txns->mti_numreaders)
1528 env->me_txns->mti_numreaders = i+1;
1529 UNLOCK_MUTEX_R(env);
1530 r = &env->me_txns->mti_readers[i];
1531 pthread_setspecific(env->me_txkey, r);
1533 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1534 txn->mt_toggle = txn->mt_txnid & 1;
1535 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1536 txn->mt_u.reader = r;
1540 txn->mt_txnid = env->me_txns->mti_txnid;
1541 txn->mt_toggle = txn->mt_txnid & 1;
1542 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1545 if (txn->mt_txnid == mdb_debug_start)
1548 txn->mt_u.dirty_list = env->me_dirty_list;
1549 txn->mt_u.dirty_list[0].mid = 0;
1550 txn->mt_free_pgs = env->me_free_pgs;
1551 txn->mt_free_pgs[0] = 0;
1555 /* Copy the DB info and flags */
1556 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1557 for (i=2; i<txn->mt_numdbs; i++)
1558 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1559 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1560 memset(txn->mt_dbflags+2, DB_STALE, env->me_numdbs-2);
1566 mdb_txn_renew(MDB_txn *txn)
1573 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1574 DPUTS("environment had fatal error, must shutdown!");
1578 rc = mdb_txn_renew0(txn);
1579 if (rc == MDB_SUCCESS) {
1580 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1581 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1582 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1588 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1593 if (env->me_flags & MDB_FATAL_ERROR) {
1594 DPUTS("environment had fatal error, must shutdown!");
1598 /* parent already has an active child txn */
1599 if (parent->mt_child) {
1603 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1604 if (!(flags & MDB_RDONLY))
1605 size += env->me_maxdbs * sizeof(MDB_cursor *);
1607 if ((txn = calloc(1, size)) == NULL) {
1608 DPRINTF("calloc: %s", strerror(ErrCode()));
1611 txn->mt_dbs = (MDB_db *)(txn+1);
1612 if (flags & MDB_RDONLY) {
1613 txn->mt_flags |= MDB_TXN_RDONLY;
1614 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1616 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1617 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1622 txn->mt_free_pgs = mdb_midl_alloc();
1623 if (!txn->mt_free_pgs) {
1627 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1628 if (!txn->mt_u.dirty_list) {
1629 free(txn->mt_free_pgs);
1633 txn->mt_txnid = parent->mt_txnid;
1634 txn->mt_toggle = parent->mt_toggle;
1635 txn->mt_u.dirty_list[0].mid = 0;
1636 txn->mt_free_pgs[0] = 0;
1637 txn->mt_next_pgno = parent->mt_next_pgno;
1638 parent->mt_child = txn;
1639 txn->mt_parent = parent;
1640 txn->mt_numdbs = parent->mt_numdbs;
1641 txn->mt_dbxs = parent->mt_dbxs;
1642 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1643 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1644 mdb_cursor_shadow(parent, txn);
1647 rc = mdb_txn_renew0(txn);
1653 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1654 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1655 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1661 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1662 * @param[in] txn the transaction handle to reset
1665 mdb_txn_reset0(MDB_txn *txn)
1667 MDB_env *env = txn->mt_env;
1669 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1670 txn->mt_u.reader->mr_txnid = 0;
1676 /* close(free) all cursors */
1677 for (i=0; i<txn->mt_numdbs; i++) {
1678 if (txn->mt_cursors[i]) {
1680 while ((mc = txn->mt_cursors[i])) {
1681 txn->mt_cursors[i] = mc->mc_next;
1682 if (mc->mc_flags & C_ALLOCD)
1688 /* return all dirty pages to dpage list */
1689 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1690 dp = txn->mt_u.dirty_list[i].mptr;
1691 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1692 dp->mp_next = txn->mt_env->me_dpages;
1693 VGMEMP_FREE(txn->mt_env, dp);
1694 txn->mt_env->me_dpages = dp;
1696 /* large pages just get freed directly */
1697 VGMEMP_FREE(txn->mt_env, dp);
1702 if (txn->mt_parent) {
1703 txn->mt_parent->mt_child = NULL;
1704 free(txn->mt_free_pgs);
1705 free(txn->mt_u.dirty_list);
1708 if (mdb_midl_shrink(&txn->mt_free_pgs))
1709 env->me_free_pgs = txn->mt_free_pgs;
1712 while ((mop = txn->mt_env->me_pghead)) {
1713 txn->mt_env->me_pghead = mop->mo_next;
1716 txn->mt_env->me_pgfirst = 0;
1717 txn->mt_env->me_pglast = 0;
1720 /* The writer mutex was locked in mdb_txn_begin. */
1721 UNLOCK_MUTEX_W(env);
1726 mdb_txn_reset(MDB_txn *txn)
1731 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1732 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1733 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1735 mdb_txn_reset0(txn);
1739 mdb_txn_abort(MDB_txn *txn)
1744 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1745 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1746 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1749 mdb_txn_abort(txn->mt_child);
1751 mdb_txn_reset0(txn);
1756 mdb_txn_commit(MDB_txn *txn)
1764 pgno_t next, freecnt;
1767 assert(txn != NULL);
1768 assert(txn->mt_env != NULL);
1770 if (txn->mt_child) {
1771 mdb_txn_commit(txn->mt_child);
1772 txn->mt_child = NULL;
1777 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1778 if (txn->mt_numdbs > env->me_numdbs) {
1779 /* update the DB flags */
1781 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
1782 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
1789 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1790 DPUTS("error flag is set, can't commit");
1792 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1797 /* Merge (and close) our cursors with parent's */
1798 mdb_cursor_merge(txn);
1800 if (txn->mt_parent) {
1806 /* Update parent's DB table */
1807 ip = &txn->mt_parent->mt_dbs[2];
1808 jp = &txn->mt_dbs[2];
1809 for (i = 2; i < txn->mt_numdbs; i++) {
1810 if (ip->md_root != jp->md_root)
1814 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1816 /* Append our free list to parent's */
1817 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1819 mdb_midl_free(txn->mt_free_pgs);
1821 /* Merge our dirty list with parent's */
1822 dst = txn->mt_parent->mt_u.dirty_list;
1823 src = txn->mt_u.dirty_list;
1824 x = mdb_mid2l_search(dst, src[1].mid);
1825 for (y=1; y<=src[0].mid; y++) {
1826 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1830 dst[x].mptr = src[y].mptr;
1833 for (; y<=src[0].mid; y++) {
1834 if (++x >= MDB_IDL_UM_MAX) {
1841 free(txn->mt_u.dirty_list);
1842 txn->mt_parent->mt_child = NULL;
1847 if (txn != env->me_txn) {
1848 DPUTS("attempt to commit unknown transaction");
1853 if (!txn->mt_u.dirty_list[0].mid)
1856 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1857 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1859 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1861 /* should only be one record now */
1862 if (env->me_pghead) {
1863 /* make sure first page of freeDB is touched and on freelist */
1864 mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
1867 /* Delete IDLs we used from the free list */
1868 if (env->me_pgfirst) {
1873 key.mv_size = sizeof(cur);
1874 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
1877 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
1878 mdb_cursor_del(&mc, 0);
1880 env->me_pgfirst = 0;
1884 /* save to free list */
1886 freecnt = txn->mt_free_pgs[0];
1887 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1890 /* make sure last page of freeDB is touched and on freelist */
1891 key.mv_size = MAXKEYSIZE+1;
1893 mdb_page_search(&mc, &key, MDB_PS_MODIFY);
1895 mdb_midl_sort(txn->mt_free_pgs);
1899 MDB_IDL idl = txn->mt_free_pgs;
1900 DPRINTF("IDL write txn %zu root %zu num %zu",
1901 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1902 for (i=0; i<idl[0]; i++) {
1903 DPRINTF("IDL %zu", idl[i+1]);
1907 /* write to last page of freeDB */
1908 key.mv_size = sizeof(pgno_t);
1909 key.mv_data = &txn->mt_txnid;
1910 data.mv_data = txn->mt_free_pgs;
1911 /* The free list can still grow during this call,
1912 * despite the pre-emptive touches above. So check
1913 * and make sure the entire thing got written.
1916 freecnt = txn->mt_free_pgs[0];
1917 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1918 rc = mdb_cursor_put(&mc, &key, &data, 0);
1923 } while (freecnt != txn->mt_free_pgs[0]);
1925 /* should only be one record now */
1927 if (env->me_pghead) {
1933 mop = env->me_pghead;
1935 key.mv_size = sizeof(id);
1937 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1938 data.mv_data = mop->mo_pages;
1939 orig = mop->mo_pages[0];
1940 /* These steps may grow the freelist again
1941 * due to freed overflow pages...
1943 mdb_cursor_put(&mc, &key, &data, 0);
1944 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
1945 /* could have been used again here */
1946 if (mop->mo_pages[0] != orig) {
1947 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1948 data.mv_data = mop->mo_pages;
1950 mdb_cursor_put(&mc, &key, &data, 0);
1952 env->me_pghead = NULL;
1955 /* was completely used up */
1956 mdb_cursor_del(&mc, 0);
1960 env->me_pgfirst = 0;
1964 while (env->me_pgfree) {
1965 MDB_oldpages *mop = env->me_pgfree;
1966 env->me_pgfree = mop->mo_next;
1970 /* Check for growth of freelist again */
1971 if (freecnt != txn->mt_free_pgs[0])
1974 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1975 if (mdb_midl_shrink(&txn->mt_free_pgs))
1976 env->me_free_pgs = txn->mt_free_pgs;
1979 /* Update DB root pointers. Their pages have already been
1980 * touched so this is all in-place and cannot fail.
1985 data.mv_size = sizeof(MDB_db);
1987 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
1988 for (i = 2; i < txn->mt_numdbs; i++) {
1989 if (txn->mt_dbflags[i] & DB_DIRTY) {
1990 data.mv_data = &txn->mt_dbs[i];
1991 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1999 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2005 /* Windows actually supports scatter/gather I/O, but only on
2006 * unbuffered file handles. Since we're relying on the OS page
2007 * cache for all our data, that's self-defeating. So we just
2008 * write pages one at a time. We use the ov structure to set
2009 * the write offset, to at least save the overhead of a Seek
2013 memset(&ov, 0, sizeof(ov));
2014 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2016 dp = txn->mt_u.dirty_list[i].mptr;
2017 DPRINTF("committing page %zu", dp->mp_pgno);
2018 size = dp->mp_pgno * env->me_psize;
2019 ov.Offset = size & 0xffffffff;
2020 ov.OffsetHigh = size >> 16;
2021 ov.OffsetHigh >>= 16;
2022 /* clear dirty flag */
2023 dp->mp_flags &= ~P_DIRTY;
2024 wsize = env->me_psize;
2025 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2026 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2029 DPRINTF("WriteFile: %d", n);
2036 struct iovec iov[MDB_COMMIT_PAGES];
2040 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2041 dp = txn->mt_u.dirty_list[i].mptr;
2042 if (dp->mp_pgno != next) {
2044 rc = writev(env->me_fd, iov, n);
2048 DPUTS("short write, filesystem full?");
2050 DPRINTF("writev: %s", strerror(n));
2057 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2060 DPRINTF("committing page %zu", dp->mp_pgno);
2061 iov[n].iov_len = env->me_psize;
2062 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2063 iov[n].iov_base = (char *)dp;
2064 size += iov[n].iov_len;
2065 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2066 /* clear dirty flag */
2067 dp->mp_flags &= ~P_DIRTY;
2068 if (++n >= MDB_COMMIT_PAGES) {
2078 rc = writev(env->me_fd, iov, n);
2082 DPUTS("short write, filesystem full?");
2084 DPRINTF("writev: %s", strerror(n));
2091 /* Drop the dirty pages.
2093 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2094 dp = txn->mt_u.dirty_list[i].mptr;
2095 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2096 dp->mp_next = txn->mt_env->me_dpages;
2097 VGMEMP_FREE(txn->mt_env, dp);
2098 txn->mt_env->me_dpages = dp;
2100 VGMEMP_FREE(txn->mt_env, dp);
2103 txn->mt_u.dirty_list[i].mid = 0;
2105 txn->mt_u.dirty_list[0].mid = 0;
2107 if ((n = mdb_env_sync(env, 0)) != 0 ||
2108 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2115 if (txn->mt_numdbs > env->me_numdbs) {
2116 /* update the DB flags */
2118 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2119 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2123 UNLOCK_MUTEX_W(env);
2129 /** Read the environment parameters of a DB environment before
2130 * mapping it into memory.
2131 * @param[in] env the environment handle
2132 * @param[out] meta address of where to store the meta information
2133 * @return 0 on success, non-zero on failure.
2136 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2143 /* We don't know the page size yet, so use a minimum value.
2147 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2149 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2154 else if (rc != MDB_PAGESIZE) {
2158 DPRINTF("read: %s", strerror(err));
2162 p = (MDB_page *)&pbuf;
2164 if (!F_ISSET(p->mp_flags, P_META)) {
2165 DPRINTF("page %zu not a meta page", p->mp_pgno);
2170 if (m->mm_magic != MDB_MAGIC) {
2171 DPUTS("meta has invalid magic");
2175 if (m->mm_version != MDB_VERSION) {
2176 DPRINTF("database is version %u, expected version %u",
2177 m->mm_version, MDB_VERSION);
2178 return MDB_VERSION_MISMATCH;
2181 memcpy(meta, m, sizeof(*m));
2185 /** Write the environment parameters of a freshly created DB environment.
2186 * @param[in] env the environment handle
2187 * @param[out] meta address of where to store the meta information
2188 * @return 0 on success, non-zero on failure.
2191 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2198 DPUTS("writing new meta page");
2200 GET_PAGESIZE(psize);
2202 meta->mm_magic = MDB_MAGIC;
2203 meta->mm_version = MDB_VERSION;
2204 meta->mm_psize = psize;
2205 meta->mm_last_pg = 1;
2206 meta->mm_flags = env->me_flags & 0xffff;
2207 meta->mm_flags |= MDB_INTEGERKEY;
2208 meta->mm_dbs[0].md_root = P_INVALID;
2209 meta->mm_dbs[1].md_root = P_INVALID;
2211 p = calloc(2, psize);
2213 p->mp_flags = P_META;
2216 memcpy(m, meta, sizeof(*meta));
2218 q = (MDB_page *)((char *)p + psize);
2221 q->mp_flags = P_META;
2224 memcpy(m, meta, sizeof(*meta));
2229 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2230 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2233 rc = write(env->me_fd, p, psize * 2);
2234 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2240 /** Update the environment info to commit a transaction.
2241 * @param[in] txn the transaction that's being committed
2242 * @return 0 on success, non-zero on failure.
2245 mdb_env_write_meta(MDB_txn *txn)
2248 MDB_meta meta, metab;
2250 int rc, len, toggle;
2256 assert(txn != NULL);
2257 assert(txn->mt_env != NULL);
2259 toggle = !txn->mt_toggle;
2260 DPRINTF("writing meta page %d for root page %zu",
2261 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2265 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2266 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2268 ptr = (char *)&meta;
2269 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2270 len = sizeof(MDB_meta) - off;
2273 meta.mm_dbs[0] = txn->mt_dbs[0];
2274 meta.mm_dbs[1] = txn->mt_dbs[1];
2275 meta.mm_last_pg = txn->mt_next_pgno - 1;
2276 meta.mm_txnid = txn->mt_txnid;
2279 off += env->me_psize;
2282 /* Write to the SYNC fd */
2285 memset(&ov, 0, sizeof(ov));
2287 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2290 rc = pwrite(env->me_mfd, ptr, len, off);
2295 DPUTS("write failed, disk error?");
2296 /* On a failure, the pagecache still contains the new data.
2297 * Write some old data back, to prevent it from being used.
2298 * Use the non-SYNC fd; we know it will fail anyway.
2300 meta.mm_last_pg = metab.mm_last_pg;
2301 meta.mm_txnid = metab.mm_txnid;
2303 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2305 r2 = pwrite(env->me_fd, ptr, len, off);
2307 env->me_flags |= MDB_FATAL_ERROR;
2310 /* Memory ordering issues are irrelevant; since the entire writer
2311 * is wrapped by wmutex, all of these changes will become visible
2312 * after the wmutex is unlocked. Since the DB is multi-version,
2313 * readers will get consistent data regardless of how fresh or
2314 * how stale their view of these values is.
2316 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2321 /** Check both meta pages to see which one is newer.
2322 * @param[in] env the environment handle
2323 * @return meta toggle (0 or 1).
2326 mdb_env_pick_meta(const MDB_env *env)
2328 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2332 mdb_env_create(MDB_env **env)
2336 e = calloc(1, sizeof(MDB_env));
2340 e->me_free_pgs = mdb_midl_alloc();
2341 if (!e->me_free_pgs) {
2345 e->me_maxreaders = DEFAULT_READERS;
2347 e->me_fd = INVALID_HANDLE_VALUE;
2348 e->me_lfd = INVALID_HANDLE_VALUE;
2349 e->me_mfd = INVALID_HANDLE_VALUE;
2350 VGMEMP_CREATE(e,0,0);
2356 mdb_env_set_mapsize(MDB_env *env, size_t size)
2360 env->me_mapsize = size;
2362 env->me_maxpg = env->me_mapsize / env->me_psize;
2367 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2371 env->me_maxdbs = dbs;
2376 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2378 if (env->me_map || readers < 1)
2380 env->me_maxreaders = readers;
2385 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2387 if (!env || !readers)
2389 *readers = env->me_maxreaders;
2393 /** Further setup required for opening an MDB environment
2396 mdb_env_open2(MDB_env *env, unsigned int flags)
2402 env->me_flags = flags;
2404 memset(&meta, 0, sizeof(meta));
2406 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2409 DPUTS("new mdbenv");
2413 if (!env->me_mapsize) {
2414 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2420 LONG sizelo, sizehi;
2421 sizelo = env->me_mapsize & 0xffffffff;
2422 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2424 /* Windows won't create mappings for zero length files.
2425 * Just allocate the maxsize right now.
2428 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2429 if (!SetEndOfFile(env->me_fd))
2431 SetFilePointer(env->me_fd, 0, NULL, 0);
2433 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
2434 sizehi, sizelo, NULL);
2437 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
2445 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2447 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
2449 if (env->me_map == MAP_FAILED) {
2456 meta.mm_mapsize = env->me_mapsize;
2457 if (flags & MDB_FIXEDMAP)
2458 meta.mm_address = env->me_map;
2459 i = mdb_env_init_meta(env, &meta);
2460 if (i != MDB_SUCCESS) {
2461 munmap(env->me_map, env->me_mapsize);
2465 env->me_psize = meta.mm_psize;
2467 env->me_maxpg = env->me_mapsize / env->me_psize;
2469 p = (MDB_page *)env->me_map;
2470 env->me_metas[0] = METADATA(p);
2471 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2475 int toggle = mdb_env_pick_meta(env);
2476 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2478 DPRINTF("opened database version %u, pagesize %u",
2479 env->me_metas[0]->mm_version, env->me_psize);
2480 DPRINTF("using meta page %d", toggle);
2481 DPRINTF("depth: %u", db->md_depth);
2482 DPRINTF("entries: %zu", db->md_entries);
2483 DPRINTF("branch pages: %zu", db->md_branch_pages);
2484 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2485 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2486 DPRINTF("root: %zu", db->md_root);
2494 /** Release a reader thread's slot in the reader lock table.
2495 * This function is called automatically when a thread exits.
2496 * @param[in] ptr This points to the slot in the reader lock table.
2499 mdb_env_reader_dest(void *ptr)
2501 MDB_reader *reader = ptr;
2503 reader->mr_txnid = 0;
2509 /** Junk for arranging thread-specific callbacks on Windows. This is
2510 * necessarily platform and compiler-specific. Windows supports up
2511 * to 1088 keys. Let's assume nobody opens more than 64 environments
2512 * in a single process, for now. They can override this if needed.
2514 #ifndef MAX_TLS_KEYS
2515 #define MAX_TLS_KEYS 64
2517 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2518 static int mdb_tls_nkeys;
2520 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2524 case DLL_PROCESS_ATTACH: break;
2525 case DLL_THREAD_ATTACH: break;
2526 case DLL_THREAD_DETACH:
2527 for (i=0; i<mdb_tls_nkeys; i++) {
2528 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2529 mdb_env_reader_dest(r);
2532 case DLL_PROCESS_DETACH: break;
2537 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2539 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2543 /* Force some symbol references.
2544 * _tls_used forces the linker to create the TLS directory if not already done
2545 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2547 #pragma comment(linker, "/INCLUDE:_tls_used")
2548 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2549 #pragma const_seg(".CRT$XLB")
2550 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2551 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2554 #pragma comment(linker, "/INCLUDE:__tls_used")
2555 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2556 #pragma data_seg(".CRT$XLB")
2557 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2559 #endif /* WIN 32/64 */
2560 #endif /* !__GNUC__ */
2563 /** Downgrade the exclusive lock on the region back to shared */
2565 mdb_env_share_locks(MDB_env *env)
2567 int toggle = mdb_env_pick_meta(env);
2569 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2574 /* First acquire a shared lock. The Unlock will
2575 * then release the existing exclusive lock.
2577 memset(&ov, 0, sizeof(ov));
2578 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2579 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2583 struct flock lock_info;
2584 /* The shared lock replaces the existing lock */
2585 memset((void *)&lock_info, 0, sizeof(lock_info));
2586 lock_info.l_type = F_RDLCK;
2587 lock_info.l_whence = SEEK_SET;
2588 lock_info.l_start = 0;
2589 lock_info.l_len = 1;
2590 fcntl(env->me_lfd, F_SETLK, &lock_info);
2594 #if defined(_WIN32) || defined(__APPLE__)
2596 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2598 * @(#) $Revision: 5.1 $
2599 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2600 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2602 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2606 * Please do not copyright this code. This code is in the public domain.
2608 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2609 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2610 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2611 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2612 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2613 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2614 * PERFORMANCE OF THIS SOFTWARE.
2617 * chongo <Landon Curt Noll> /\oo/\
2618 * http://www.isthe.com/chongo/
2620 * Share and Enjoy! :-)
2623 typedef unsigned long long mdb_hash_t;
2624 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2626 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2627 * @param[in] str string to hash
2628 * @param[in] hval initial value for hash
2629 * @return 64 bit hash
2631 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2632 * hval arg on the first call.
2635 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
2637 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
2638 unsigned char *end = s + val->mv_size;
2640 * FNV-1a hash each octet of the string
2643 /* xor the bottom with the current octet */
2644 hval ^= (mdb_hash_t)*s++;
2646 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2647 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2648 (hval << 7) + (hval << 8) + (hval << 40);
2650 /* return our new hash value */
2654 /** Hash the string and output the hash in hex.
2655 * @param[in] str string to hash
2656 * @param[out] hexbuf an array of 17 chars to hold the hash
2659 mdb_hash_hex(MDB_val *val, char *hexbuf)
2662 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
2663 for (i=0; i<8; i++) {
2664 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2670 /** Open and/or initialize the lock region for the environment.
2671 * @param[in] env The MDB environment.
2672 * @param[in] lpath The pathname of the file used for the lock region.
2673 * @param[in] mode The Unix permissions for the file, if we create it.
2674 * @param[out] excl Set to true if we got an exclusive lock on the region.
2675 * @return 0 on success, non-zero on failure.
2678 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2686 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2687 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2688 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2692 /* Try to get exclusive lock. If we succeed, then
2693 * nobody is using the lock region and we should initialize it.
2696 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2700 memset(&ov, 0, sizeof(ov));
2701 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2707 size = GetFileSize(env->me_lfd, NULL);
2713 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1)
2715 /* Lose record locks when exec*() */
2716 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
2717 fcntl(env->me_lfd, F_SETFD, fdflags);
2719 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
2720 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1)
2724 /* Try to get exclusive lock. If we succeed, then
2725 * nobody is using the lock region and we should initialize it.
2728 struct flock lock_info;
2729 memset((void *)&lock_info, 0, sizeof(lock_info));
2730 lock_info.l_type = F_WRLCK;
2731 lock_info.l_whence = SEEK_SET;
2732 lock_info.l_start = 0;
2733 lock_info.l_len = 1;
2734 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2738 lock_info.l_type = F_RDLCK;
2739 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2746 size = lseek(env->me_lfd, 0, SEEK_END);
2748 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2749 if (size < rsize && *excl) {
2751 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2752 if (!SetEndOfFile(env->me_lfd)) {
2757 if (ftruncate(env->me_lfd, rsize) != 0) {
2764 size = rsize - sizeof(MDB_txninfo);
2765 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2770 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2776 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2778 if (!env->me_txns) {
2783 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2785 if (m == MAP_FAILED) {
2786 env->me_txns = NULL;
2795 BY_HANDLE_FILE_INFORMATION stbuf;
2804 if (!mdb_sec_inited) {
2805 InitializeSecurityDescriptor(&mdb_null_sd,
2806 SECURITY_DESCRIPTOR_REVISION);
2807 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2808 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2809 mdb_all_sa.bInheritHandle = FALSE;
2810 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2813 GetFileInformationByHandle(env->me_lfd, &stbuf);
2814 idbuf.volume = stbuf.dwVolumeSerialNumber;
2815 idbuf.nhigh = stbuf.nFileIndexHigh;
2816 idbuf.nlow = stbuf.nFileIndexLow;
2817 val.mv_data = &idbuf;
2818 val.mv_size = sizeof(idbuf);
2819 mdb_hash_hex(&val, hexbuf);
2820 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
2821 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2822 if (!env->me_rmutex) {
2826 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
2827 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
2828 if (!env->me_wmutex) {
2842 fstat(env->me_lfd, &stbuf);
2843 idbuf.dev = stbuf.st_dev;
2844 idbuf.ino = stbuf.st_ino;
2845 val.mv_data = &idbuf;
2846 val.mv_size = sizeof(idbuf);
2847 mdb_hash_hex(&val, hexbuf);
2848 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
2849 if (sem_unlink(env->me_txns->mti_rmname)) {
2851 if (rc != ENOENT && rc != EINVAL)
2854 env->me_rmutex = sem_open(env->me_txns->mti_rmname, O_CREAT, mode, 1);
2855 if (!env->me_rmutex) {
2859 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
2860 if (sem_unlink(env->me_txns->mti_wmname)) {
2862 if (rc != ENOENT && rc != EINVAL)
2865 env->me_wmutex = sem_open(env->me_txns->mti_wmname, O_CREAT, mode, 1);
2866 if (!env->me_wmutex) {
2870 #else /* __APPLE__ */
2871 pthread_mutexattr_t mattr;
2873 pthread_mutexattr_init(&mattr);
2874 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2878 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2879 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2880 #endif /* __APPLE__ */
2882 env->me_txns->mti_version = MDB_VERSION;
2883 env->me_txns->mti_magic = MDB_MAGIC;
2884 env->me_txns->mti_txnid = 0;
2885 env->me_txns->mti_numreaders = 0;
2888 if (env->me_txns->mti_magic != MDB_MAGIC) {
2889 DPUTS("lock region has invalid magic");
2893 if (env->me_txns->mti_version != MDB_VERSION) {
2894 DPRINTF("lock region is version %u, expected version %u",
2895 env->me_txns->mti_version, MDB_VERSION);
2896 rc = MDB_VERSION_MISMATCH;
2900 if (rc != EACCES && rc != EAGAIN) {
2904 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2905 if (!env->me_rmutex) {
2909 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2910 if (!env->me_wmutex) {
2916 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
2917 if (!env->me_rmutex) {
2921 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
2922 if (!env->me_wmutex) {
2932 env->me_lfd = INVALID_HANDLE_VALUE;
2937 /** The name of the lock file in the DB environment */
2938 #define LOCKNAME "/lock.mdb"
2939 /** The name of the data file in the DB environment */
2940 #define DATANAME "/data.mdb"
2941 /** The suffix of the lock file when no subdir is used */
2942 #define LOCKSUFF "-lock"
2945 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2947 int oflags, rc, len, excl;
2948 char *lpath, *dpath;
2951 if (flags & MDB_NOSUBDIR) {
2952 rc = len + sizeof(LOCKSUFF) + len + 1;
2954 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
2959 if (flags & MDB_NOSUBDIR) {
2960 dpath = lpath + len + sizeof(LOCKSUFF);
2961 sprintf(lpath, "%s" LOCKSUFF, path);
2962 strcpy(dpath, path);
2964 dpath = lpath + len + sizeof(LOCKNAME);
2965 sprintf(lpath, "%s" LOCKNAME, path);
2966 sprintf(dpath, "%s" DATANAME, path);
2969 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
2974 if (F_ISSET(flags, MDB_RDONLY)) {
2975 oflags = GENERIC_READ;
2976 len = OPEN_EXISTING;
2978 oflags = GENERIC_READ|GENERIC_WRITE;
2981 mode = FILE_ATTRIBUTE_NORMAL;
2982 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2983 NULL, len, mode, NULL);
2985 if (F_ISSET(flags, MDB_RDONLY))
2988 oflags = O_RDWR | O_CREAT;
2990 env->me_fd = open(dpath, oflags, mode);
2992 if (env->me_fd == INVALID_HANDLE_VALUE) {
2997 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
2998 if (flags & (MDB_RDONLY|MDB_NOSYNC|MDB_NOMETASYNC)) {
2999 env->me_mfd = env->me_fd;
3001 /* synchronous fd for meta writes */
3003 env->me_mfd = CreateFile(dpath, oflags,
3004 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3005 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3007 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3009 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3014 env->me_path = strdup(path);
3015 DPRINTF("opened dbenv %p", (void *) env);
3016 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3018 /* Windows TLS callbacks need help finding their TLS info. */
3019 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3020 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3027 mdb_env_share_locks(env);
3029 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3030 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3031 if (!env->me_dbxs || !env->me_dbflags)
3037 if (env->me_fd != INVALID_HANDLE_VALUE) {
3039 env->me_fd = INVALID_HANDLE_VALUE;
3041 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3043 env->me_lfd = INVALID_HANDLE_VALUE;
3051 mdb_env_close(MDB_env *env)
3058 VGMEMP_DESTROY(env);
3059 while (env->me_dpages) {
3060 dp = env->me_dpages;
3061 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3062 env->me_dpages = dp->mp_next;
3066 free(env->me_dbflags);
3070 pthread_key_delete(env->me_txkey);
3072 /* Delete our key from the global list */
3074 for (i=0; i<mdb_tls_nkeys; i++)
3075 if (mdb_tls_keys[i] == env->me_txkey) {
3076 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3084 munmap(env->me_map, env->me_mapsize);
3086 if (env->me_mfd != env->me_fd)
3090 pid_t pid = getpid();
3092 for (i=0; i<env->me_txns->mti_numreaders; i++)
3093 if (env->me_txns->mti_readers[i].mr_pid == pid)
3094 env->me_txns->mti_readers[i].mr_pid = 0;
3095 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3098 mdb_midl_free(env->me_free_pgs);
3102 /** Compare two items pointing at aligned size_t's */
3104 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3106 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3107 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3110 /** Compare two items pointing at aligned int's */
3112 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3114 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3115 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3118 /** Compare two items pointing at ints of unknown alignment.
3119 * Nodes and keys are guaranteed to be 2-byte aligned.
3122 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3124 #if BYTE_ORDER == LITTLE_ENDIAN
3125 unsigned short *u, *c;
3128 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3129 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3132 } while(!x && u > (unsigned short *)a->mv_data);
3135 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3139 /** Compare two items lexically */
3141 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3148 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3154 diff = memcmp(a->mv_data, b->mv_data, len);
3155 return diff ? diff : len_diff<0 ? -1 : len_diff;
3158 /** Compare two items in reverse byte order */
3160 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3162 const unsigned char *p1, *p2, *p1_lim;
3166 p1_lim = (const unsigned char *)a->mv_data;
3167 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3168 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3170 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3176 while (p1 > p1_lim) {
3177 diff = *--p1 - *--p2;
3181 return len_diff<0 ? -1 : len_diff;
3184 /** Search for key within a page, using binary search.
3185 * Returns the smallest entry larger or equal to the key.
3186 * If exactp is non-null, stores whether the found entry was an exact match
3187 * in *exactp (1 or 0).
3188 * Updates the cursor index with the index of the found entry.
3189 * If no entry larger or equal to the key is found, returns NULL.
3192 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3194 unsigned int i = 0, nkeys;
3197 MDB_page *mp = mc->mc_pg[mc->mc_top];
3198 MDB_node *node = NULL;
3203 nkeys = NUMKEYS(mp);
3208 COPY_PGNO(pgno, mp->mp_pgno);
3209 DPRINTF("searching %u keys in %s %spage %zu",
3210 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3217 low = IS_LEAF(mp) ? 0 : 1;
3219 cmp = mc->mc_dbx->md_cmp;
3221 /* Branch pages have no data, so if using integer keys,
3222 * alignment is guaranteed. Use faster mdb_cmp_int.
3224 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3225 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3232 nodekey.mv_size = mc->mc_db->md_pad;
3233 node = NODEPTR(mp, 0); /* fake */
3234 while (low <= high) {
3235 i = (low + high) >> 1;
3236 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3237 rc = cmp(key, &nodekey);
3238 DPRINTF("found leaf index %u [%s], rc = %i",
3239 i, DKEY(&nodekey), rc);
3248 while (low <= high) {
3249 i = (low + high) >> 1;
3251 node = NODEPTR(mp, i);
3252 nodekey.mv_size = NODEKSZ(node);
3253 nodekey.mv_data = NODEKEY(node);
3255 rc = cmp(key, &nodekey);
3258 DPRINTF("found leaf index %u [%s], rc = %i",
3259 i, DKEY(&nodekey), rc);
3261 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3262 i, DKEY(&nodekey), NODEPGNO(node), rc);
3273 if (rc > 0) { /* Found entry is less than the key. */
3274 i++; /* Skip to get the smallest entry larger than key. */
3276 node = NODEPTR(mp, i);
3279 *exactp = (rc == 0);
3280 /* store the key index */
3281 mc->mc_ki[mc->mc_top] = i;
3283 /* There is no entry larger or equal to the key. */
3286 /* nodeptr is fake for LEAF2 */
3292 mdb_cursor_adjust(MDB_cursor *mc, func)
3296 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3297 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3304 /** Pop a page off the top of the cursor's stack. */
3306 mdb_cursor_pop(MDB_cursor *mc)
3311 top = mc->mc_pg[mc->mc_top];
3316 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3317 mc->mc_dbi, (void *) mc);
3321 /** Push a page onto the top of the cursor's stack. */
3323 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3325 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3326 mc->mc_dbi, (void *) mc);
3328 if (mc->mc_snum >= CURSOR_STACK) {
3329 assert(mc->mc_snum < CURSOR_STACK);
3333 mc->mc_top = mc->mc_snum++;
3334 mc->mc_pg[mc->mc_top] = mp;
3335 mc->mc_ki[mc->mc_top] = 0;
3340 /** Find the address of the page corresponding to a given page number.
3341 * @param[in] txn the transaction for this access.
3342 * @param[in] pgno the page number for the page to retrieve.
3343 * @param[out] ret address of a pointer where the page's address will be stored.
3344 * @return 0 on success, non-zero on failure.
3347 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3351 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3353 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3354 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3355 p = txn->mt_u.dirty_list[x].mptr;
3359 if (pgno < txn->mt_next_pgno)
3360 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3364 DPRINTF("page %zu not found", pgno);
3367 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3370 /** Search for the page a given key should be in.
3371 * Pushes parent pages on the cursor stack. This function continues a
3372 * search on a cursor that has already been initialized. (Usually by
3373 * #mdb_page_search() but also by #mdb_node_move().)
3374 * @param[in,out] mc the cursor for this operation.
3375 * @param[in] key the key to search for. If NULL, search for the lowest
3376 * page. (This is used by #mdb_cursor_first().)
3377 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3378 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3379 * @return 0 on success, non-zero on failure.
3382 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3384 MDB_page *mp = mc->mc_pg[mc->mc_top];
3389 while (IS_BRANCH(mp)) {
3393 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3394 assert(NUMKEYS(mp) > 1);
3395 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3397 if (key == NULL) /* Initialize cursor to first page. */
3399 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3400 /* cursor to last page */
3404 node = mdb_node_search(mc, key, &exact);
3406 i = NUMKEYS(mp) - 1;
3408 i = mc->mc_ki[mc->mc_top];
3417 DPRINTF("following index %u for key [%s]",
3419 assert(i < NUMKEYS(mp));
3420 node = NODEPTR(mp, i);
3422 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3425 mc->mc_ki[mc->mc_top] = i;
3426 if ((rc = mdb_cursor_push(mc, mp)))
3430 if ((rc = mdb_page_touch(mc)) != 0)
3432 mp = mc->mc_pg[mc->mc_top];
3437 DPRINTF("internal error, index points to a %02X page!?",
3439 return MDB_CORRUPTED;
3442 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3443 key ? DKEY(key) : NULL);
3448 /** Search for the page a given key should be in.
3449 * Pushes parent pages on the cursor stack. This function just sets up
3450 * the search; it finds the root page for \b mc's database and sets this
3451 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3452 * called to complete the search.
3453 * @param[in,out] mc the cursor for this operation.
3454 * @param[in] key the key to search for. If NULL, search for the lowest
3455 * page. (This is used by #mdb_cursor_first().)
3456 * @param[in] modify If true, visited pages are updated with new page numbers.
3457 * @return 0 on success, non-zero on failure.
3460 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
3465 /* Make sure the txn is still viable, then find the root from
3466 * the txn's db table.
3468 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3469 DPUTS("transaction has failed, must abort");
3472 /* Make sure we're using an up-to-date root */
3473 if (mc->mc_dbi > MAIN_DBI) {
3474 if ((*mc->mc_dbflag & DB_STALE) ||
3475 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
3477 unsigned char dbflag = 0;
3478 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3479 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
3482 if (*mc->mc_dbflag & DB_STALE) {
3485 MDB_node *leaf = mdb_node_search(&mc2,
3486 &mc->mc_dbx->md_name, &exact);
3488 return MDB_NOTFOUND;
3489 mdb_node_read(mc->mc_txn, leaf, &data);
3490 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3492 if (flags & MDB_PS_MODIFY)
3494 *mc->mc_dbflag = dbflag;
3497 root = mc->mc_db->md_root;
3499 if (root == P_INVALID) { /* Tree is empty. */
3500 DPUTS("tree is empty");
3501 return MDB_NOTFOUND;
3506 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3512 DPRINTF("db %u root page %zu has flags 0x%X",
3513 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3515 if (flags & MDB_PS_MODIFY) {
3516 if ((rc = mdb_page_touch(mc)))
3520 if (flags & MDB_PS_ROOTONLY)
3523 return mdb_page_search_root(mc, key, flags);
3526 /** Return the data associated with a given node.
3527 * @param[in] txn The transaction for this operation.
3528 * @param[in] leaf The node being read.
3529 * @param[out] data Updated to point to the node's data.
3530 * @return 0 on success, non-zero on failure.
3533 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3535 MDB_page *omp; /* overflow page */
3539 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3540 data->mv_size = NODEDSZ(leaf);
3541 data->mv_data = NODEDATA(leaf);
3545 /* Read overflow data.
3547 data->mv_size = NODEDSZ(leaf);
3548 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3549 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3550 DPRINTF("read overflow page %zu failed", pgno);
3553 data->mv_data = METADATA(omp);
3559 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3560 MDB_val *key, MDB_val *data)
3569 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3571 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3574 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3578 mdb_cursor_init(&mc, txn, dbi, &mx);
3579 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3582 /** Find a sibling for a page.
3583 * Replaces the page at the top of the cursor's stack with the
3584 * specified sibling, if one exists.
3585 * @param[in] mc The cursor for this operation.
3586 * @param[in] move_right Non-zero if the right sibling is requested,
3587 * otherwise the left sibling.
3588 * @return 0 on success, non-zero on failure.
3591 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3597 if (mc->mc_snum < 2) {
3598 return MDB_NOTFOUND; /* root has no siblings */
3602 DPRINTF("parent page is page %zu, index %u",
3603 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3605 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3606 : (mc->mc_ki[mc->mc_top] == 0)) {
3607 DPRINTF("no more keys left, moving to %s sibling",
3608 move_right ? "right" : "left");
3609 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3613 mc->mc_ki[mc->mc_top]++;
3615 mc->mc_ki[mc->mc_top]--;
3616 DPRINTF("just moving to %s index key %u",
3617 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3619 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3621 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3622 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3625 mdb_cursor_push(mc, mp);
3630 /** Move the cursor to the next data item. */
3632 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3638 if (mc->mc_flags & C_EOF) {
3639 return MDB_NOTFOUND;
3642 assert(mc->mc_flags & C_INITIALIZED);
3644 mp = mc->mc_pg[mc->mc_top];
3646 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3647 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3648 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3649 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3650 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3651 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3655 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3656 if (op == MDB_NEXT_DUP)
3657 return MDB_NOTFOUND;
3661 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3663 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3664 DPUTS("=====> move to next sibling page");
3665 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3666 mc->mc_flags |= C_EOF;
3667 mc->mc_flags &= ~C_INITIALIZED;
3668 return MDB_NOTFOUND;
3670 mp = mc->mc_pg[mc->mc_top];
3671 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3673 mc->mc_ki[mc->mc_top]++;
3675 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3676 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3679 key->mv_size = mc->mc_db->md_pad;
3680 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3684 assert(IS_LEAF(mp));
3685 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3687 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3688 mdb_xcursor_init1(mc, leaf);
3691 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3694 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3695 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3696 if (rc != MDB_SUCCESS)
3701 MDB_SET_KEY(leaf, key);
3705 /** Move the cursor to the previous data item. */
3707 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3713 assert(mc->mc_flags & C_INITIALIZED);
3715 mp = mc->mc_pg[mc->mc_top];
3717 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3718 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3719 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3720 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3721 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3722 if (op != MDB_PREV || rc == MDB_SUCCESS)
3725 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3726 if (op == MDB_PREV_DUP)
3727 return MDB_NOTFOUND;
3732 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3734 if (mc->mc_ki[mc->mc_top] == 0) {
3735 DPUTS("=====> move to prev sibling page");
3736 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3737 mc->mc_flags &= ~C_INITIALIZED;
3738 return MDB_NOTFOUND;
3740 mp = mc->mc_pg[mc->mc_top];
3741 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3742 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3744 mc->mc_ki[mc->mc_top]--;
3746 mc->mc_flags &= ~C_EOF;
3748 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3749 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3752 key->mv_size = mc->mc_db->md_pad;
3753 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3757 assert(IS_LEAF(mp));
3758 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3760 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3761 mdb_xcursor_init1(mc, leaf);
3764 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3767 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3768 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3769 if (rc != MDB_SUCCESS)
3774 MDB_SET_KEY(leaf, key);
3778 /** Set the cursor on a specific data item. */
3780 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3781 MDB_cursor_op op, int *exactp)
3790 assert(key->mv_size > 0);
3792 /* See if we're already on the right page */
3793 if (mc->mc_flags & C_INITIALIZED) {
3796 mp = mc->mc_pg[mc->mc_top];
3798 mc->mc_ki[mc->mc_top] = 0;
3799 return MDB_NOTFOUND;
3801 if (mp->mp_flags & P_LEAF2) {
3802 nodekey.mv_size = mc->mc_db->md_pad;
3803 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
3805 leaf = NODEPTR(mp, 0);
3806 MDB_SET_KEY(leaf, &nodekey);
3808 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3810 /* Probably happens rarely, but first node on the page
3811 * was the one we wanted.
3813 mc->mc_ki[mc->mc_top] = 0;
3814 leaf = NODEPTR(mp, 0);
3821 unsigned int nkeys = NUMKEYS(mp);
3823 if (mp->mp_flags & P_LEAF2) {
3824 nodekey.mv_data = LEAF2KEY(mp,
3825 nkeys-1, nodekey.mv_size);
3827 leaf = NODEPTR(mp, nkeys-1);
3828 MDB_SET_KEY(leaf, &nodekey);
3830 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3832 /* last node was the one we wanted */
3833 mc->mc_ki[mc->mc_top] = nkeys-1;
3834 leaf = NODEPTR(mp, nkeys-1);
3840 /* This is definitely the right page, skip search_page */
3845 /* If any parents have right-sibs, search.
3846 * Otherwise, there's nothing further.
3848 for (i=0; i<mc->mc_top; i++)
3850 NUMKEYS(mc->mc_pg[i])-1)
3852 if (i == mc->mc_top) {
3853 /* There are no other pages */
3854 mc->mc_ki[mc->mc_top] = nkeys;
3855 return MDB_NOTFOUND;
3859 /* There are no other pages */
3860 mc->mc_ki[mc->mc_top] = 0;
3861 return MDB_NOTFOUND;
3865 rc = mdb_page_search(mc, key, 0);
3866 if (rc != MDB_SUCCESS)
3869 mp = mc->mc_pg[mc->mc_top];
3870 assert(IS_LEAF(mp));
3873 leaf = mdb_node_search(mc, key, exactp);
3874 if (exactp != NULL && !*exactp) {
3875 /* MDB_SET specified and not an exact match. */
3876 return MDB_NOTFOUND;
3880 DPUTS("===> inexact leaf not found, goto sibling");
3881 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
3882 return rc; /* no entries matched */
3883 mp = mc->mc_pg[mc->mc_top];
3884 assert(IS_LEAF(mp));
3885 leaf = NODEPTR(mp, 0);
3889 mc->mc_flags |= C_INITIALIZED;
3890 mc->mc_flags &= ~C_EOF;
3893 key->mv_size = mc->mc_db->md_pad;
3894 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3898 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3899 mdb_xcursor_init1(mc, leaf);
3902 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3903 if (op == MDB_SET || op == MDB_SET_RANGE) {
3904 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3907 if (op == MDB_GET_BOTH) {
3913 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
3914 if (rc != MDB_SUCCESS)
3917 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
3919 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
3921 rc = mc->mc_dbx->md_dcmp(data, &d2);
3923 if (op == MDB_GET_BOTH || rc > 0)
3924 return MDB_NOTFOUND;
3929 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3930 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3935 /* The key already matches in all other cases */
3936 if (op == MDB_SET_RANGE)
3937 MDB_SET_KEY(leaf, key);
3938 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3943 /** Move the cursor to the first item in the database. */
3945 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3950 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3951 rc = mdb_page_search(mc, NULL, 0);
3952 if (rc != MDB_SUCCESS)
3955 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3957 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3958 mc->mc_flags |= C_INITIALIZED;
3959 mc->mc_flags &= ~C_EOF;
3961 mc->mc_ki[mc->mc_top] = 0;
3963 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3964 key->mv_size = mc->mc_db->md_pad;
3965 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
3970 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3971 mdb_xcursor_init1(mc, leaf);
3972 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3977 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3978 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3982 MDB_SET_KEY(leaf, key);
3986 /** Move the cursor to the last item in the database. */
3988 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3994 lkey.mv_size = MAXKEYSIZE+1;
3995 lkey.mv_data = NULL;
3997 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3998 rc = mdb_page_search(mc, &lkey, 0);
3999 if (rc != MDB_SUCCESS)
4002 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4004 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
4005 mc->mc_flags |= C_INITIALIZED;
4006 mc->mc_flags &= ~C_EOF;
4008 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4010 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4011 key->mv_size = mc->mc_db->md_pad;
4012 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4017 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4018 mdb_xcursor_init1(mc, leaf);
4019 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4024 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4025 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4030 MDB_SET_KEY(leaf, key);
4035 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4045 case MDB_GET_BOTH_RANGE:
4046 if (data == NULL || mc->mc_xcursor == NULL) {
4053 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4055 } else if (op == MDB_SET_RANGE)
4056 rc = mdb_cursor_set(mc, key, data, op, NULL);
4058 rc = mdb_cursor_set(mc, key, data, op, &exact);
4060 case MDB_GET_MULTIPLE:
4062 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4063 !(mc->mc_flags & C_INITIALIZED)) {
4068 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4069 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4072 case MDB_NEXT_MULTIPLE:
4074 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4078 if (!(mc->mc_flags & C_INITIALIZED))
4079 rc = mdb_cursor_first(mc, key, data);
4081 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4082 if (rc == MDB_SUCCESS) {
4083 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4086 mx = &mc->mc_xcursor->mx_cursor;
4087 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4089 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4090 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4098 case MDB_NEXT_NODUP:
4099 if (!(mc->mc_flags & C_INITIALIZED))
4100 rc = mdb_cursor_first(mc, key, data);
4102 rc = mdb_cursor_next(mc, key, data, op);
4106 case MDB_PREV_NODUP:
4107 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
4108 rc = mdb_cursor_last(mc, key, data);
4110 rc = mdb_cursor_prev(mc, key, data, op);
4113 rc = mdb_cursor_first(mc, key, data);
4117 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4118 !(mc->mc_flags & C_INITIALIZED) ||
4119 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4123 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4126 rc = mdb_cursor_last(mc, key, data);
4130 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4131 !(mc->mc_flags & C_INITIALIZED) ||
4132 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4136 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4139 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4147 /** Touch all the pages in the cursor stack.
4148 * Makes sure all the pages are writable, before attempting a write operation.
4149 * @param[in] mc The cursor to operate on.
4152 mdb_cursor_touch(MDB_cursor *mc)
4156 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4158 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4159 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4162 *mc->mc_dbflag = DB_DIRTY;
4164 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4165 rc = mdb_page_touch(mc);
4169 mc->mc_top = mc->mc_snum-1;
4174 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4177 MDB_node *leaf = NULL;
4178 MDB_val xdata, *rdata, dkey;
4182 unsigned int mcount = 0;
4186 char dbuf[MAXKEYSIZE+1];
4187 unsigned int nflags;
4190 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4193 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4194 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4198 if (flags == MDB_CURRENT) {
4199 if (!(mc->mc_flags & C_INITIALIZED))
4202 } else if (mc->mc_db->md_root == P_INVALID) {
4204 /* new database, write a root leaf page */
4205 DPUTS("allocating new root leaf page");
4206 if ((np = mdb_page_new(mc, P_LEAF, 1)) == NULL) {
4210 mdb_cursor_push(mc, np);
4211 mc->mc_db->md_root = np->mp_pgno;
4212 mc->mc_db->md_depth++;
4213 *mc->mc_dbflag = DB_DIRTY;
4214 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4216 np->mp_flags |= P_LEAF2;
4217 mc->mc_flags |= C_INITIALIZED;
4223 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4224 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4225 DPRINTF("duplicate key [%s]", DKEY(key));
4227 return MDB_KEYEXIST;
4229 if (rc && rc != MDB_NOTFOUND)
4233 /* Cursor is positioned, now make sure all pages are writable */
4234 rc2 = mdb_cursor_touch(mc);
4239 /* The key already exists */
4240 if (rc == MDB_SUCCESS) {
4241 /* there's only a key anyway, so this is a no-op */
4242 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4243 unsigned int ksize = mc->mc_db->md_pad;
4244 if (key->mv_size != ksize)
4246 if (flags == MDB_CURRENT) {
4247 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4248 memcpy(ptr, key->mv_data, ksize);
4253 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4256 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4257 /* Was a single item before, must convert now */
4259 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4260 /* Just overwrite the current item */
4261 if (flags == MDB_CURRENT)
4264 dkey.mv_size = NODEDSZ(leaf);
4265 dkey.mv_data = NODEDATA(leaf);
4266 #if UINT_MAX < SIZE_MAX
4267 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4268 #ifdef MISALIGNED_OK
4269 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4271 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4274 /* if data matches, ignore it */
4275 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4276 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4278 /* create a fake page for the dup items */
4279 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4280 dkey.mv_data = dbuf;
4281 fp = (MDB_page *)&pbuf;
4282 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4283 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4284 fp->mp_lower = PAGEHDRSZ;
4285 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4286 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4287 fp->mp_flags |= P_LEAF2;
4288 fp->mp_pad = data->mv_size;
4290 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4291 (dkey.mv_size & 1) + (data->mv_size & 1);
4293 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4296 xdata.mv_size = fp->mp_upper;
4301 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4302 /* See if we need to convert from fake page to subDB */
4304 unsigned int offset;
4307 fp = NODEDATA(leaf);
4308 if (flags == MDB_CURRENT) {
4309 fp->mp_flags |= P_DIRTY;
4310 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4311 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4315 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4316 offset = fp->mp_pad;
4318 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4320 offset += offset & 1;
4321 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4322 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4324 /* yes, convert it */
4326 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4327 dummy.md_pad = fp->mp_pad;
4328 dummy.md_flags = MDB_DUPFIXED;
4329 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4330 dummy.md_flags |= MDB_INTEGERKEY;
4333 dummy.md_branch_pages = 0;
4334 dummy.md_leaf_pages = 1;
4335 dummy.md_overflow_pages = 0;
4336 dummy.md_entries = NUMKEYS(fp);
4338 xdata.mv_size = sizeof(MDB_db);
4339 xdata.mv_data = &dummy;
4340 mp = mdb_page_alloc(mc, 1);
4343 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4344 flags |= F_DUPDATA|F_SUBDATA;
4345 dummy.md_root = mp->mp_pgno;
4347 /* no, just grow it */
4349 xdata.mv_size = NODEDSZ(leaf) + offset;
4350 xdata.mv_data = &pbuf;
4351 mp = (MDB_page *)&pbuf;
4352 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4355 mp->mp_flags = fp->mp_flags | P_DIRTY;
4356 mp->mp_pad = fp->mp_pad;
4357 mp->mp_lower = fp->mp_lower;
4358 mp->mp_upper = fp->mp_upper + offset;
4360 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4362 nsize = NODEDSZ(leaf) - fp->mp_upper;
4363 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4364 for (i=0; i<NUMKEYS(fp); i++)
4365 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4367 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4371 /* data is on sub-DB, just store it */
4372 flags |= F_DUPDATA|F_SUBDATA;
4376 /* overflow page overwrites need special handling */
4377 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4380 int ovpages, dpages;
4382 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4383 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4384 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4385 mdb_page_get(mc->mc_txn, pg, &omp);
4386 /* Is the ov page writable and large enough? */
4387 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4388 /* yes, overwrite it. Note in this case we don't
4389 * bother to try shrinking the node if the new data
4390 * is smaller than the overflow threshold.
4392 if (F_ISSET(flags, MDB_RESERVE))
4393 data->mv_data = METADATA(omp);
4395 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4398 /* no, free ovpages */
4400 mc->mc_db->md_overflow_pages -= ovpages;
4401 for (i=0; i<ovpages; i++) {
4402 DPRINTF("freed ov page %zu", pg);
4403 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4407 } else if (NODEDSZ(leaf) == data->mv_size) {
4408 /* same size, just replace it. Note that we could
4409 * also reuse this node if the new data is smaller,
4410 * but instead we opt to shrink the node in that case.
4412 if (F_ISSET(flags, MDB_RESERVE))
4413 data->mv_data = NODEDATA(leaf);
4415 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4418 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4419 mc->mc_db->md_entries--;
4421 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4427 nflags = flags & NODE_ADD_FLAGS;
4428 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4429 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4430 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4431 nflags &= ~MDB_APPEND;
4432 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4434 /* There is room already in this leaf page. */
4435 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4436 if (rc == 0 && !do_sub) {
4437 /* Adjust other cursors pointing to mp */
4438 MDB_cursor *m2, *m3;
4439 MDB_dbi dbi = mc->mc_dbi;
4440 unsigned i = mc->mc_top;
4441 MDB_page *mp = mc->mc_pg[i];
4443 if (mc->mc_flags & C_SUB)
4446 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4447 if (mc->mc_flags & C_SUB)
4448 m3 = &m2->mc_xcursor->mx_cursor;
4451 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4452 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4459 if (rc != MDB_SUCCESS)
4460 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4462 /* Now store the actual data in the child DB. Note that we're
4463 * storing the user data in the keys field, so there are strict
4464 * size limits on dupdata. The actual data fields of the child
4465 * DB are all zero size.
4472 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4473 if (flags & MDB_CURRENT) {
4474 xflags = MDB_CURRENT;
4476 mdb_xcursor_init1(mc, leaf);
4477 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4479 /* converted, write the original data first */
4481 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4485 /* Adjust other cursors pointing to mp */
4487 unsigned i = mc->mc_top;
4488 MDB_page *mp = mc->mc_pg[i];
4490 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4491 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4492 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4493 mdb_xcursor_init1(m2, leaf);
4498 xflags |= (flags & MDB_APPEND);
4499 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4500 if (flags & F_SUBDATA) {
4501 void *db = NODEDATA(leaf);
4502 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4505 /* sub-writes might have failed so check rc again.
4506 * Don't increment count if we just replaced an existing item.
4508 if (!rc && !(flags & MDB_CURRENT))
4509 mc->mc_db->md_entries++;
4510 if (flags & MDB_MULTIPLE) {
4512 if (mcount < data[1].mv_size) {
4513 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4514 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4524 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4529 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4532 if (!mc->mc_flags & C_INITIALIZED)
4535 rc = mdb_cursor_touch(mc);
4539 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4541 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4542 if (flags != MDB_NODUPDATA) {
4543 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4544 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4546 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4547 /* If sub-DB still has entries, we're done */
4548 if (mc->mc_xcursor->mx_db.md_entries) {
4549 if (leaf->mn_flags & F_SUBDATA) {
4550 /* update subDB info */
4551 void *db = NODEDATA(leaf);
4552 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4554 /* shrink fake page */
4555 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4557 mc->mc_db->md_entries--;
4560 /* otherwise fall thru and delete the sub-DB */
4563 if (leaf->mn_flags & F_SUBDATA) {
4564 /* add all the child DB's pages to the free list */
4565 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4566 if (rc == MDB_SUCCESS) {
4567 mc->mc_db->md_entries -=
4568 mc->mc_xcursor->mx_db.md_entries;
4573 return mdb_cursor_del0(mc, leaf);
4576 /** Allocate and initialize new pages for a database.
4577 * @param[in] mc a cursor on the database being added to.
4578 * @param[in] flags flags defining what type of page is being allocated.
4579 * @param[in] num the number of pages to allocate. This is usually 1,
4580 * unless allocating overflow pages for a large record.
4581 * @return Address of a page, or NULL on failure.
4584 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num)
4588 if ((np = mdb_page_alloc(mc, num)) == NULL)
4590 DPRINTF("allocated new mpage %zu, page size %u",
4591 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4592 np->mp_flags = flags | P_DIRTY;
4593 np->mp_lower = PAGEHDRSZ;
4594 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4597 mc->mc_db->md_branch_pages++;
4598 else if (IS_LEAF(np))
4599 mc->mc_db->md_leaf_pages++;
4600 else if (IS_OVERFLOW(np)) {
4601 mc->mc_db->md_overflow_pages += num;
4608 /** Calculate the size of a leaf node.
4609 * The size depends on the environment's page size; if a data item
4610 * is too large it will be put onto an overflow page and the node
4611 * size will only include the key and not the data. Sizes are always
4612 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4613 * of the #MDB_node headers.
4614 * @param[in] env The environment handle.
4615 * @param[in] key The key for the node.
4616 * @param[in] data The data for the node.
4617 * @return The number of bytes needed to store the node.
4620 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4624 sz = LEAFSIZE(key, data);
4625 if (sz >= env->me_psize / MDB_MINKEYS) {
4626 /* put on overflow page */
4627 sz -= data->mv_size - sizeof(pgno_t);
4631 return sz + sizeof(indx_t);
4634 /** Calculate the size of a branch node.
4635 * The size should depend on the environment's page size but since
4636 * we currently don't support spilling large keys onto overflow
4637 * pages, it's simply the size of the #MDB_node header plus the
4638 * size of the key. Sizes are always rounded up to an even number
4639 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4640 * @param[in] env The environment handle.
4641 * @param[in] key The key for the node.
4642 * @return The number of bytes needed to store the node.
4645 mdb_branch_size(MDB_env *env, MDB_val *key)
4650 if (sz >= env->me_psize / MDB_MINKEYS) {
4651 /* put on overflow page */
4652 /* not implemented */
4653 /* sz -= key->size - sizeof(pgno_t); */
4656 return sz + sizeof(indx_t);
4659 /** Add a node to the page pointed to by the cursor.
4660 * @param[in] mc The cursor for this operation.
4661 * @param[in] indx The index on the page where the new node should be added.
4662 * @param[in] key The key for the new node.
4663 * @param[in] data The data for the new node, if any.
4664 * @param[in] pgno The page number, if adding a branch node.
4665 * @param[in] flags Flags for the node.
4666 * @return 0 on success, non-zero on failure. Possible errors are:
4668 * <li>ENOMEM - failed to allocate overflow pages for the node.
4669 * <li>ENOSPC - there is insufficient room in the page. This error
4670 * should never happen since all callers already calculate the
4671 * page's free space before calling this function.
4675 mdb_node_add(MDB_cursor *mc, indx_t indx,
4676 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4679 size_t node_size = NODESIZE;
4682 MDB_page *mp = mc->mc_pg[mc->mc_top];
4683 MDB_page *ofp = NULL; /* overflow page */
4686 assert(mp->mp_upper >= mp->mp_lower);
4688 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4689 IS_LEAF(mp) ? "leaf" : "branch",
4690 IS_SUBP(mp) ? "sub-" : "",
4691 mp->mp_pgno, indx, data ? data->mv_size : 0,
4692 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4695 /* Move higher keys up one slot. */
4696 int ksize = mc->mc_db->md_pad, dif;
4697 char *ptr = LEAF2KEY(mp, indx, ksize);
4698 dif = NUMKEYS(mp) - indx;
4700 memmove(ptr+ksize, ptr, dif*ksize);
4701 /* insert new key */
4702 memcpy(ptr, key->mv_data, ksize);
4704 /* Just using these for counting */
4705 mp->mp_lower += sizeof(indx_t);
4706 mp->mp_upper -= ksize - sizeof(indx_t);
4711 node_size += key->mv_size;
4715 if (F_ISSET(flags, F_BIGDATA)) {
4716 /* Data already on overflow page. */
4717 node_size += sizeof(pgno_t);
4718 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4719 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4720 /* Put data on overflow page. */
4721 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4722 data->mv_size, node_size+data->mv_size);
4723 node_size += sizeof(pgno_t);
4724 if ((ofp = mdb_page_new(mc, P_OVERFLOW, ovpages)) == NULL)
4726 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4729 node_size += data->mv_size;
4732 node_size += node_size & 1;
4734 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4735 DPRINTF("not enough room in page %zu, got %u ptrs",
4736 mp->mp_pgno, NUMKEYS(mp));
4737 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4738 mp->mp_upper - mp->mp_lower);
4739 DPRINTF("node size = %zu", node_size);
4743 /* Move higher pointers up one slot. */
4744 for (i = NUMKEYS(mp); i > indx; i--)
4745 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
4747 /* Adjust free space offsets. */
4748 ofs = mp->mp_upper - node_size;
4749 assert(ofs >= mp->mp_lower + sizeof(indx_t));
4750 mp->mp_ptrs[indx] = ofs;
4752 mp->mp_lower += sizeof(indx_t);
4754 /* Write the node data. */
4755 node = NODEPTR(mp, indx);
4756 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
4757 node->mn_flags = flags;
4759 SETDSZ(node,data->mv_size);
4764 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4769 if (F_ISSET(flags, F_BIGDATA))
4770 memcpy(node->mn_data + key->mv_size, data->mv_data,
4772 else if (F_ISSET(flags, MDB_RESERVE))
4773 data->mv_data = node->mn_data + key->mv_size;
4775 memcpy(node->mn_data + key->mv_size, data->mv_data,
4778 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
4780 if (F_ISSET(flags, MDB_RESERVE))
4781 data->mv_data = METADATA(ofp);
4783 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
4790 /** Delete the specified node from a page.
4791 * @param[in] mp The page to operate on.
4792 * @param[in] indx The index of the node to delete.
4793 * @param[in] ksize The size of a node. Only used if the page is
4794 * part of a #MDB_DUPFIXED database.
4797 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
4800 indx_t i, j, numkeys, ptr;
4807 COPY_PGNO(pgno, mp->mp_pgno);
4808 DPRINTF("delete node %u on %s page %zu", indx,
4809 IS_LEAF(mp) ? "leaf" : "branch", pgno);
4812 assert(indx < NUMKEYS(mp));
4815 int x = NUMKEYS(mp) - 1 - indx;
4816 base = LEAF2KEY(mp, indx, ksize);
4818 memmove(base, base + ksize, x * ksize);
4819 mp->mp_lower -= sizeof(indx_t);
4820 mp->mp_upper += ksize - sizeof(indx_t);
4824 node = NODEPTR(mp, indx);
4825 sz = NODESIZE + node->mn_ksize;
4827 if (F_ISSET(node->mn_flags, F_BIGDATA))
4828 sz += sizeof(pgno_t);
4830 sz += NODEDSZ(node);
4834 ptr = mp->mp_ptrs[indx];
4835 numkeys = NUMKEYS(mp);
4836 for (i = j = 0; i < numkeys; i++) {
4838 mp->mp_ptrs[j] = mp->mp_ptrs[i];
4839 if (mp->mp_ptrs[i] < ptr)
4840 mp->mp_ptrs[j] += sz;
4845 base = (char *)mp + mp->mp_upper;
4846 memmove(base + sz, base, ptr - mp->mp_upper);
4848 mp->mp_lower -= sizeof(indx_t);
4852 /** Compact the main page after deleting a node on a subpage.
4853 * @param[in] mp The main page to operate on.
4854 * @param[in] indx The index of the subpage on the main page.
4857 mdb_node_shrink(MDB_page *mp, indx_t indx)
4864 indx_t i, numkeys, ptr;
4866 node = NODEPTR(mp, indx);
4867 sp = (MDB_page *)NODEDATA(node);
4868 osize = NODEDSZ(node);
4870 delta = sp->mp_upper - sp->mp_lower;
4871 SETDSZ(node, osize - delta);
4872 xp = (MDB_page *)((char *)sp + delta);
4874 /* shift subpage upward */
4876 nsize = NUMKEYS(sp) * sp->mp_pad;
4877 memmove(METADATA(xp), METADATA(sp), nsize);
4880 nsize = osize - sp->mp_upper;
4881 numkeys = NUMKEYS(sp);
4882 for (i=numkeys-1; i>=0; i--)
4883 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
4885 xp->mp_upper = sp->mp_lower;
4886 xp->mp_lower = sp->mp_lower;
4887 xp->mp_flags = sp->mp_flags;
4888 xp->mp_pad = sp->mp_pad;
4889 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
4891 /* shift lower nodes upward */
4892 ptr = mp->mp_ptrs[indx];
4893 numkeys = NUMKEYS(mp);
4894 for (i = 0; i < numkeys; i++) {
4895 if (mp->mp_ptrs[i] <= ptr)
4896 mp->mp_ptrs[i] += delta;
4899 base = (char *)mp + mp->mp_upper;
4900 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
4901 mp->mp_upper += delta;
4904 /** Initial setup of a sorted-dups cursor.
4905 * Sorted duplicates are implemented as a sub-database for the given key.
4906 * The duplicate data items are actually keys of the sub-database.
4907 * Operations on the duplicate data items are performed using a sub-cursor
4908 * initialized when the sub-database is first accessed. This function does
4909 * the preliminary setup of the sub-cursor, filling in the fields that
4910 * depend only on the parent DB.
4911 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4914 mdb_xcursor_init0(MDB_cursor *mc)
4916 MDB_xcursor *mx = mc->mc_xcursor;
4918 mx->mx_cursor.mc_xcursor = NULL;
4919 mx->mx_cursor.mc_txn = mc->mc_txn;
4920 mx->mx_cursor.mc_db = &mx->mx_db;
4921 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
4922 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
4923 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
4924 mx->mx_cursor.mc_snum = 0;
4925 mx->mx_cursor.mc_top = 0;
4926 mx->mx_cursor.mc_flags = C_SUB;
4927 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
4928 mx->mx_dbx.md_dcmp = NULL;
4929 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
4932 /** Final setup of a sorted-dups cursor.
4933 * Sets up the fields that depend on the data from the main cursor.
4934 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4935 * @param[in] node The data containing the #MDB_db record for the
4936 * sorted-dup database.
4939 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
4941 MDB_xcursor *mx = mc->mc_xcursor;
4943 if (node->mn_flags & F_SUBDATA) {
4944 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
4945 mx->mx_cursor.mc_snum = 0;
4946 mx->mx_cursor.mc_flags = C_SUB;
4948 MDB_page *fp = NODEDATA(node);
4949 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
4950 mx->mx_db.md_flags = 0;
4951 mx->mx_db.md_depth = 1;
4952 mx->mx_db.md_branch_pages = 0;
4953 mx->mx_db.md_leaf_pages = 1;
4954 mx->mx_db.md_overflow_pages = 0;
4955 mx->mx_db.md_entries = NUMKEYS(fp);
4956 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
4957 mx->mx_cursor.mc_snum = 1;
4958 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
4959 mx->mx_cursor.mc_top = 0;
4960 mx->mx_cursor.mc_pg[0] = fp;
4961 mx->mx_cursor.mc_ki[0] = 0;
4962 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4963 mx->mx_db.md_flags = MDB_DUPFIXED;
4964 mx->mx_db.md_pad = fp->mp_pad;
4965 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4966 mx->mx_db.md_flags |= MDB_INTEGERKEY;
4969 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
4971 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
4973 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
4974 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
4975 #if UINT_MAX < SIZE_MAX
4976 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
4977 #ifdef MISALIGNED_OK
4978 mx->mx_dbx.md_cmp = mdb_cmp_long;
4980 mx->mx_dbx.md_cmp = mdb_cmp_cint;
4985 /** Initialize a cursor for a given transaction and database. */
4987 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
4992 mc->mc_db = &txn->mt_dbs[dbi];
4993 mc->mc_dbx = &txn->mt_dbxs[dbi];
4994 mc->mc_dbflag = &txn->mt_dbflags[dbi];
4998 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5000 mc->mc_xcursor = mx;
5001 mdb_xcursor_init0(mc);
5003 mc->mc_xcursor = NULL;
5005 if (*mc->mc_dbflag & DB_STALE) {
5006 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5011 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5014 MDB_xcursor *mx = NULL;
5015 size_t size = sizeof(MDB_cursor);
5017 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5020 /* Allow read access to the freelist */
5021 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5024 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5025 size += sizeof(MDB_xcursor);
5027 if ((mc = malloc(size)) != NULL) {
5028 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5029 mx = (MDB_xcursor *)(mc + 1);
5031 mdb_cursor_init(mc, txn, dbi, mx);
5032 if (txn->mt_cursors) {
5033 mc->mc_next = txn->mt_cursors[dbi];
5034 txn->mt_cursors[dbi] = mc;
5036 mc->mc_flags |= C_ALLOCD;
5046 /* Return the count of duplicate data items for the current key */
5048 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5052 if (mc == NULL || countp == NULL)
5055 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5058 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5059 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5062 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5065 *countp = mc->mc_xcursor->mx_db.md_entries;
5071 mdb_cursor_close(MDB_cursor *mc)
5074 /* remove from txn, if tracked */
5075 if (mc->mc_txn->mt_cursors) {
5076 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5077 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5079 *prev = mc->mc_next;
5081 if (mc->mc_flags & C_ALLOCD)
5087 mdb_cursor_txn(MDB_cursor *mc)
5089 if (!mc) return NULL;
5094 mdb_cursor_dbi(MDB_cursor *mc)
5100 /** Replace the key for a node with a new key.
5101 * @param[in] mp The page containing the node to operate on.
5102 * @param[in] indx The index of the node to operate on.
5103 * @param[in] key The new key to use.
5104 * @return 0 on success, non-zero on failure.
5107 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5113 indx_t ptr, i, numkeys;
5116 node = NODEPTR(mp, indx);
5117 ptr = mp->mp_ptrs[indx];
5121 char kbuf2[(MAXKEYSIZE*2+1)];
5122 k2.mv_data = NODEKEY(node);
5123 k2.mv_size = node->mn_ksize;
5124 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5126 mdb_dkey(&k2, kbuf2),
5132 delta0 = delta = key->mv_size - node->mn_ksize;
5134 /* Must be 2-byte aligned. If new key is
5135 * shorter by 1, the shift will be skipped.
5137 delta += (delta & 1);
5139 if (delta > 0 && SIZELEFT(mp) < delta) {
5140 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5144 numkeys = NUMKEYS(mp);
5145 for (i = 0; i < numkeys; i++) {
5146 if (mp->mp_ptrs[i] <= ptr)
5147 mp->mp_ptrs[i] -= delta;
5150 base = (char *)mp + mp->mp_upper;
5151 len = ptr - mp->mp_upper + NODESIZE;
5152 memmove(base - delta, base, len);
5153 mp->mp_upper -= delta;
5155 node = NODEPTR(mp, indx);
5158 /* But even if no shift was needed, update ksize */
5160 node->mn_ksize = key->mv_size;
5163 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5168 /** Move a node from csrc to cdst.
5171 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5177 unsigned short flags;
5181 /* Mark src and dst as dirty. */
5182 if ((rc = mdb_page_touch(csrc)) ||
5183 (rc = mdb_page_touch(cdst)))
5186 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5187 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5188 key.mv_size = csrc->mc_db->md_pad;
5189 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5191 data.mv_data = NULL;
5195 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5196 assert(!((long)srcnode&1));
5197 srcpg = NODEPGNO(srcnode);
5198 flags = srcnode->mn_flags;
5199 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5200 unsigned int snum = csrc->mc_snum;
5202 /* must find the lowest key below src */
5203 mdb_page_search_root(csrc, NULL, 0);
5204 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5205 key.mv_size = csrc->mc_db->md_pad;
5206 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5208 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5209 key.mv_size = NODEKSZ(s2);
5210 key.mv_data = NODEKEY(s2);
5212 csrc->mc_snum = snum--;
5213 csrc->mc_top = snum;
5215 key.mv_size = NODEKSZ(srcnode);
5216 key.mv_data = NODEKEY(srcnode);
5218 data.mv_size = NODEDSZ(srcnode);
5219 data.mv_data = NODEDATA(srcnode);
5221 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5222 unsigned int snum = cdst->mc_snum;
5225 /* must find the lowest key below dst */
5226 mdb_page_search_root(cdst, NULL, 0);
5227 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5228 bkey.mv_size = cdst->mc_db->md_pad;
5229 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5231 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5232 bkey.mv_size = NODEKSZ(s2);
5233 bkey.mv_data = NODEKEY(s2);
5235 cdst->mc_snum = snum--;
5236 cdst->mc_top = snum;
5237 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5240 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5241 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5242 csrc->mc_ki[csrc->mc_top],
5244 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5245 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5247 /* Add the node to the destination page.
5249 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5250 if (rc != MDB_SUCCESS)
5253 /* Delete the node from the source page.
5255 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5258 /* Adjust other cursors pointing to mp */
5259 MDB_cursor *m2, *m3;
5260 MDB_dbi dbi = csrc->mc_dbi;
5261 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5263 if (csrc->mc_flags & C_SUB)
5266 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5267 if (m2 == csrc) continue;
5268 if (csrc->mc_flags & C_SUB)
5269 m3 = &m2->mc_xcursor->mx_cursor;
5272 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5273 csrc->mc_ki[csrc->mc_top]) {
5274 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5275 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5280 /* Update the parent separators.
5282 if (csrc->mc_ki[csrc->mc_top] == 0) {
5283 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5284 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5285 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5287 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5288 key.mv_size = NODEKSZ(srcnode);
5289 key.mv_data = NODEKEY(srcnode);
5291 DPRINTF("update separator for source page %zu to [%s]",
5292 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5293 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5294 &key)) != MDB_SUCCESS)
5297 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5299 nullkey.mv_size = 0;
5300 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5301 assert(rc == MDB_SUCCESS);
5305 if (cdst->mc_ki[cdst->mc_top] == 0) {
5306 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5307 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5308 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5310 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5311 key.mv_size = NODEKSZ(srcnode);
5312 key.mv_data = NODEKEY(srcnode);
5314 DPRINTF("update separator for destination page %zu to [%s]",
5315 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5316 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5317 &key)) != MDB_SUCCESS)
5320 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5322 nullkey.mv_size = 0;
5323 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5324 assert(rc == MDB_SUCCESS);
5331 /** Merge one page into another.
5332 * The nodes from the page pointed to by \b csrc will
5333 * be copied to the page pointed to by \b cdst and then
5334 * the \b csrc page will be freed.
5335 * @param[in] csrc Cursor pointing to the source page.
5336 * @param[in] cdst Cursor pointing to the destination page.
5339 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5347 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5348 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5350 assert(csrc->mc_snum > 1); /* can't merge root page */
5351 assert(cdst->mc_snum > 1);
5353 /* Mark dst as dirty. */
5354 if ((rc = mdb_page_touch(cdst)))
5357 /* Move all nodes from src to dst.
5359 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5360 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5361 key.mv_size = csrc->mc_db->md_pad;
5362 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5363 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5364 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5365 if (rc != MDB_SUCCESS)
5367 key.mv_data = (char *)key.mv_data + key.mv_size;
5370 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5371 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5372 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5373 unsigned int snum = csrc->mc_snum;
5375 /* must find the lowest key below src */
5376 mdb_page_search_root(csrc, NULL, 0);
5377 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5378 key.mv_size = csrc->mc_db->md_pad;
5379 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5381 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5382 key.mv_size = NODEKSZ(s2);
5383 key.mv_data = NODEKEY(s2);
5385 csrc->mc_snum = snum--;
5386 csrc->mc_top = snum;
5388 key.mv_size = srcnode->mn_ksize;
5389 key.mv_data = NODEKEY(srcnode);
5392 data.mv_size = NODEDSZ(srcnode);
5393 data.mv_data = NODEDATA(srcnode);
5394 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5395 if (rc != MDB_SUCCESS)
5400 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5401 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);
5403 /* Unlink the src page from parent and add to free list.
5405 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5406 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5408 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5412 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5413 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5414 csrc->mc_db->md_leaf_pages--;
5416 csrc->mc_db->md_branch_pages--;
5418 /* Adjust other cursors pointing to mp */
5419 MDB_cursor *m2, *m3;
5420 MDB_dbi dbi = csrc->mc_dbi;
5421 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5423 if (csrc->mc_flags & C_SUB)
5426 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5427 if (csrc->mc_flags & C_SUB)
5428 m3 = &m2->mc_xcursor->mx_cursor;
5431 if (m3 == csrc) continue;
5432 if (m3->mc_snum < csrc->mc_snum) continue;
5433 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5434 m3->mc_pg[csrc->mc_top] = mp;
5435 m3->mc_ki[csrc->mc_top] += nkeys;
5439 mdb_cursor_pop(csrc);
5441 return mdb_rebalance(csrc);
5444 /** Copy the contents of a cursor.
5445 * @param[in] csrc The cursor to copy from.
5446 * @param[out] cdst The cursor to copy to.
5449 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5453 cdst->mc_txn = csrc->mc_txn;
5454 cdst->mc_dbi = csrc->mc_dbi;
5455 cdst->mc_db = csrc->mc_db;
5456 cdst->mc_dbx = csrc->mc_dbx;
5457 cdst->mc_snum = csrc->mc_snum;
5458 cdst->mc_top = csrc->mc_top;
5459 cdst->mc_flags = csrc->mc_flags;
5461 for (i=0; i<csrc->mc_snum; i++) {
5462 cdst->mc_pg[i] = csrc->mc_pg[i];
5463 cdst->mc_ki[i] = csrc->mc_ki[i];
5467 /** Rebalance the tree after a delete operation.
5468 * @param[in] mc Cursor pointing to the page where rebalancing
5470 * @return 0 on success, non-zero on failure.
5473 mdb_rebalance(MDB_cursor *mc)
5483 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5484 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5485 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5486 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5490 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5493 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5494 DPRINTF("no need to rebalance page %zu, above fill threshold",
5500 if (mc->mc_snum < 2) {
5501 MDB_page *mp = mc->mc_pg[0];
5502 if (NUMKEYS(mp) == 0) {
5503 DPUTS("tree is completely empty");
5504 mc->mc_db->md_root = P_INVALID;
5505 mc->mc_db->md_depth = 0;
5506 mc->mc_db->md_leaf_pages = 0;
5507 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5511 /* Adjust other cursors pointing to mp */
5512 MDB_cursor *m2, *m3;
5513 MDB_dbi dbi = mc->mc_dbi;
5515 if (mc->mc_flags & C_SUB)
5518 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5519 if (m2 == mc) continue;
5520 if (mc->mc_flags & C_SUB)
5521 m3 = &m2->mc_xcursor->mx_cursor;
5524 if (m3->mc_snum < mc->mc_snum) continue;
5525 if (m3->mc_pg[0] == mp) {
5531 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5532 DPUTS("collapsing root page!");
5533 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5534 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5535 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5538 mc->mc_db->md_depth--;
5539 mc->mc_db->md_branch_pages--;
5541 /* Adjust other cursors pointing to mp */
5542 MDB_cursor *m2, *m3;
5543 MDB_dbi dbi = mc->mc_dbi;
5545 if (mc->mc_flags & C_SUB)
5548 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5549 if (m2 == mc) continue;
5550 if (mc->mc_flags & C_SUB)
5551 m3 = &m2->mc_xcursor->mx_cursor;
5554 if (m3->mc_snum < mc->mc_snum) continue;
5555 if (m3->mc_pg[0] == mp) {
5556 m3->mc_pg[0] = mc->mc_pg[0];
5561 DPUTS("root page doesn't need rebalancing");
5565 /* The parent (branch page) must have at least 2 pointers,
5566 * otherwise the tree is invalid.
5568 ptop = mc->mc_top-1;
5569 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5571 /* Leaf page fill factor is below the threshold.
5572 * Try to move keys from left or right neighbor, or
5573 * merge with a neighbor page.
5578 mdb_cursor_copy(mc, &mn);
5579 mn.mc_xcursor = NULL;
5581 if (mc->mc_ki[ptop] == 0) {
5582 /* We're the leftmost leaf in our parent.
5584 DPUTS("reading right neighbor");
5586 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5587 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5589 mn.mc_ki[mn.mc_top] = 0;
5590 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5592 /* There is at least one neighbor to the left.
5594 DPUTS("reading left neighbor");
5596 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5597 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5599 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5600 mc->mc_ki[mc->mc_top] = 0;
5603 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5604 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);
5606 /* If the neighbor page is above threshold and has at least two
5607 * keys, move one key from it.
5609 * Otherwise we should try to merge them.
5611 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5612 return mdb_node_move(&mn, mc);
5613 else { /* FIXME: if (has_enough_room()) */
5614 mc->mc_flags &= ~C_INITIALIZED;
5615 if (mc->mc_ki[ptop] == 0)
5616 return mdb_page_merge(&mn, mc);
5618 return mdb_page_merge(mc, &mn);
5622 /** Complete a delete operation started by #mdb_cursor_del(). */
5624 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5628 /* add overflow pages to free list */
5629 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5633 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5634 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5635 mc->mc_db->md_overflow_pages -= ovpages;
5636 for (i=0; i<ovpages; i++) {
5637 DPRINTF("freed ov page %zu", pg);
5638 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5642 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5643 mc->mc_db->md_entries--;
5644 rc = mdb_rebalance(mc);
5645 if (rc != MDB_SUCCESS)
5646 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5652 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5653 MDB_val *key, MDB_val *data)
5658 MDB_val rdata, *xdata;
5662 assert(key != NULL);
5664 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5666 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5669 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5673 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5677 mdb_cursor_init(&mc, txn, dbi, &mx);
5688 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5690 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5694 /** Split a page and insert a new node.
5695 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5696 * The cursor will be updated to point to the actual page and index where
5697 * the node got inserted after the split.
5698 * @param[in] newkey The key for the newly inserted node.
5699 * @param[in] newdata The data for the newly inserted node.
5700 * @param[in] newpgno The page number, if the new node is a branch node.
5701 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
5702 * @return 0 on success, non-zero on failure.
5705 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5706 unsigned int nflags)
5709 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1;
5712 unsigned int i, j, split_indx, nkeys, pmax;
5714 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5716 MDB_page *mp, *rp, *pp;
5721 mp = mc->mc_pg[mc->mc_top];
5722 newindx = mc->mc_ki[mc->mc_top];
5724 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
5725 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
5726 DKEY(newkey), mc->mc_ki[mc->mc_top]);
5728 /* Create a right sibling. */
5729 if ((rp = mdb_page_new(mc, mp->mp_flags, 1)) == NULL)
5731 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
5733 if (mc->mc_snum < 2) {
5734 if ((pp = mdb_page_new(mc, P_BRANCH, 1)) == NULL)
5736 /* shift current top to make room for new parent */
5737 mc->mc_pg[1] = mc->mc_pg[0];
5738 mc->mc_ki[1] = mc->mc_ki[0];
5741 mc->mc_db->md_root = pp->mp_pgno;
5742 DPRINTF("root split! new root = %zu", pp->mp_pgno);
5743 mc->mc_db->md_depth++;
5746 /* Add left (implicit) pointer. */
5747 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
5748 /* undo the pre-push */
5749 mc->mc_pg[0] = mc->mc_pg[1];
5750 mc->mc_ki[0] = mc->mc_ki[1];
5751 mc->mc_db->md_root = mp->mp_pgno;
5752 mc->mc_db->md_depth--;
5759 ptop = mc->mc_top-1;
5760 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
5763 mdb_cursor_copy(mc, &mn);
5764 mn.mc_pg[mn.mc_top] = rp;
5765 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
5767 if (nflags & MDB_APPEND) {
5768 mn.mc_ki[mn.mc_top] = 0;
5775 nkeys = NUMKEYS(mp);
5776 split_indx = (nkeys + 1) / 2;
5781 unsigned int lsize, rsize, ksize;
5782 /* Move half of the keys to the right sibling */
5784 x = mc->mc_ki[mc->mc_top] - split_indx;
5785 ksize = mc->mc_db->md_pad;
5786 split = LEAF2KEY(mp, split_indx, ksize);
5787 rsize = (nkeys - split_indx) * ksize;
5788 lsize = (nkeys - split_indx) * sizeof(indx_t);
5789 mp->mp_lower -= lsize;
5790 rp->mp_lower += lsize;
5791 mp->mp_upper += rsize - lsize;
5792 rp->mp_upper -= rsize - lsize;
5793 sepkey.mv_size = ksize;
5794 if (newindx == split_indx) {
5795 sepkey.mv_data = newkey->mv_data;
5797 sepkey.mv_data = split;
5800 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
5801 memcpy(rp->mp_ptrs, split, rsize);
5802 sepkey.mv_data = rp->mp_ptrs;
5803 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
5804 memcpy(ins, newkey->mv_data, ksize);
5805 mp->mp_lower += sizeof(indx_t);
5806 mp->mp_upper -= ksize - sizeof(indx_t);
5809 memcpy(rp->mp_ptrs, split, x * ksize);
5810 ins = LEAF2KEY(rp, x, ksize);
5811 memcpy(ins, newkey->mv_data, ksize);
5812 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
5813 rp->mp_lower += sizeof(indx_t);
5814 rp->mp_upper -= ksize - sizeof(indx_t);
5815 mc->mc_ki[mc->mc_top] = x;
5816 mc->mc_pg[mc->mc_top] = rp;
5821 /* For leaf pages, check the split point based on what
5822 * fits where, since otherwise mdb_node_add can fail.
5824 * This check is only needed when the data items are
5825 * relatively large, such that being off by one will
5826 * make the difference between success or failure.
5827 * When the size of the data items is much smaller than
5828 * one-half of a page, this check is irrelevant.
5831 unsigned int psize, nsize;
5832 /* Maximum free space in an empty page */
5833 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
5834 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
5835 if ((nkeys < 20) || (nsize > pmax/4)) {
5836 if (newindx <= split_indx) {
5839 for (i=0; i<split_indx; i++) {
5840 node = NODEPTR(mp, i);
5841 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5842 if (F_ISSET(node->mn_flags, F_BIGDATA))
5843 psize += sizeof(pgno_t);
5845 psize += NODEDSZ(node);
5848 if (i == split_indx - 1 && newindx == split_indx)
5857 for (i=nkeys-1; i>=split_indx; i--) {
5858 node = NODEPTR(mp, i);
5859 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5860 if (F_ISSET(node->mn_flags, F_BIGDATA))
5861 psize += sizeof(pgno_t);
5863 psize += NODEDSZ(node);
5874 /* First find the separating key between the split pages.
5875 * The case where newindx == split_indx is ambiguous; the
5876 * new item could go to the new page or stay on the original
5877 * page. If newpos == 1 it goes to the new page.
5879 if (newindx == split_indx && newpos) {
5880 sepkey.mv_size = newkey->mv_size;
5881 sepkey.mv_data = newkey->mv_data;
5883 node = NODEPTR(mp, split_indx);
5884 sepkey.mv_size = node->mn_ksize;
5885 sepkey.mv_data = NODEKEY(node);
5889 DPRINTF("separator is [%s]", DKEY(&sepkey));
5891 /* Copy separator key to the parent.
5893 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
5896 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
5898 /* Right page might now have changed parent.
5899 * Check if left page also changed parent.
5901 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
5902 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
5903 mc->mc_pg[ptop] = mn.mc_pg[ptop];
5904 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
5908 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
5911 if (rc != MDB_SUCCESS) {
5914 if (nflags & MDB_APPEND) {
5915 mc->mc_pg[mc->mc_top] = rp;
5916 mc->mc_ki[mc->mc_top] = 0;
5917 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
5926 /* Move half of the keys to the right sibling. */
5928 /* grab a page to hold a temporary copy */
5929 copy = mdb_page_malloc(mc);
5933 copy->mp_pgno = mp->mp_pgno;
5934 copy->mp_flags = mp->mp_flags;
5935 copy->mp_lower = PAGEHDRSZ;
5936 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
5937 mc->mc_pg[mc->mc_top] = copy;
5938 for (i = j = 0; i <= nkeys; j++) {
5939 if (i == split_indx) {
5940 /* Insert in right sibling. */
5941 /* Reset insert index for right sibling. */
5942 if (i != newindx || (newpos ^ ins_new)) {
5944 mc->mc_pg[mc->mc_top] = rp;
5948 if (i == newindx && !ins_new) {
5949 /* Insert the original entry that caused the split. */
5950 rkey.mv_data = newkey->mv_data;
5951 rkey.mv_size = newkey->mv_size;
5960 /* Update index for the new key. */
5961 mc->mc_ki[mc->mc_top] = j;
5962 } else if (i == nkeys) {
5965 node = NODEPTR(mp, i);
5966 rkey.mv_data = NODEKEY(node);
5967 rkey.mv_size = node->mn_ksize;
5969 xdata.mv_data = NODEDATA(node);
5970 xdata.mv_size = NODEDSZ(node);
5973 pgno = NODEPGNO(node);
5974 flags = node->mn_flags;
5979 if (!IS_LEAF(mp) && j == 0) {
5980 /* First branch index doesn't need key data. */
5984 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
5988 nkeys = NUMKEYS(copy);
5989 for (i=0; i<nkeys; i++)
5990 mp->mp_ptrs[i] = copy->mp_ptrs[i];
5991 mp->mp_lower = copy->mp_lower;
5992 mp->mp_upper = copy->mp_upper;
5993 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
5994 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
5996 /* reset back to original page */
5997 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
5998 mc->mc_pg[mc->mc_top] = mp;
5999 if (nflags & MDB_RESERVE) {
6000 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6001 if (!(node->mn_flags & F_BIGDATA))
6002 newdata->mv_data = NODEDATA(node);
6006 /* return tmp page to freelist */
6007 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
6008 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
6009 mc->mc_txn->mt_env->me_dpages = copy;
6012 /* Adjust other cursors pointing to mp */
6013 MDB_cursor *m2, *m3;
6014 MDB_dbi dbi = mc->mc_dbi;
6016 if (mc->mc_flags & C_SUB)
6019 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6020 if (m2 == mc) continue;
6021 if (mc->mc_flags & C_SUB)
6022 m3 = &m2->mc_xcursor->mx_cursor;
6025 if (!(m3->mc_flags & C_INITIALIZED))
6030 for (k=m3->mc_top; k>=0; k--) {
6031 m3->mc_ki[k+1] = m3->mc_ki[k];
6032 m3->mc_pg[k+1] = m3->mc_pg[k];
6034 m3->mc_ki[0] = mc->mc_ki[0];
6035 m3->mc_pg[0] = mc->mc_pg[0];
6039 if (m3->mc_pg[mc->mc_top] == mp) {
6040 if (m3->mc_ki[m3->mc_top] >= split_indx) {
6041 m3->mc_pg[m3->mc_top] = rp;
6042 m3->mc_ki[m3->mc_top] -= split_indx;
6051 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6052 MDB_val *key, MDB_val *data, unsigned int flags)
6057 assert(key != NULL);
6058 assert(data != NULL);
6060 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6063 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6067 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6071 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6074 mdb_cursor_init(&mc, txn, dbi, &mx);
6075 return mdb_cursor_put(&mc, key, data, flags);
6078 /** Only a subset of the @ref mdb_env flags can be changed
6079 * at runtime. Changing other flags requires closing the environment
6080 * and re-opening it with the new flags.
6082 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC)
6084 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6086 if ((flag & CHANGEABLE) != flag)
6089 env->me_flags |= flag;
6091 env->me_flags &= ~flag;
6096 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6101 *arg = env->me_flags;
6106 mdb_env_get_path(MDB_env *env, const char **arg)
6111 *arg = env->me_path;
6115 /** Common code for #mdb_stat() and #mdb_env_stat().
6116 * @param[in] env the environment to operate in.
6117 * @param[in] db the #MDB_db record containing the stats to return.
6118 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6119 * @return 0, this function always succeeds.
6122 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6124 arg->ms_psize = env->me_psize;
6125 arg->ms_depth = db->md_depth;
6126 arg->ms_branch_pages = db->md_branch_pages;
6127 arg->ms_leaf_pages = db->md_leaf_pages;
6128 arg->ms_overflow_pages = db->md_overflow_pages;
6129 arg->ms_entries = db->md_entries;
6134 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6138 if (env == NULL || arg == NULL)
6141 toggle = mdb_env_pick_meta(env);
6143 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6146 /** Set the default comparison functions for a database.
6147 * Called immediately after a database is opened to set the defaults.
6148 * The user can then override them with #mdb_set_compare() or
6149 * #mdb_set_dupsort().
6150 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6151 * @param[in] dbi A database handle returned by #mdb_open()
6154 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6156 uint16_t f = txn->mt_dbs[dbi].md_flags;
6158 txn->mt_dbxs[dbi].md_cmp =
6159 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6160 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6162 txn->mt_dbxs[dbi].md_dcmp =
6163 !(f & MDB_DUPSORT) ? 0 :
6164 ((f & MDB_INTEGERDUP)
6165 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6166 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6169 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6174 int rc, dbflag, exact;
6175 unsigned int unused = 0;
6178 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6179 mdb_default_cmp(txn, FREE_DBI);
6185 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
6186 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
6187 mdb_default_cmp(txn, MAIN_DBI);
6191 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6192 mdb_default_cmp(txn, MAIN_DBI);
6195 /* Is the DB already open? */
6197 for (i=2; i<txn->mt_numdbs; i++) {
6198 if (!txn->mt_dbxs[i].md_name.mv_size) {
6199 /* Remember this free slot */
6200 if (!unused) unused = i;
6203 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6204 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6210 /* If no free slot and max hit, fail */
6211 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
6214 /* Find the DB info */
6218 key.mv_data = (void *)name;
6219 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6220 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6221 if (rc == MDB_SUCCESS) {
6222 /* make sure this is actually a DB */
6223 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6224 if (!(node->mn_flags & F_SUBDATA))
6226 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6227 /* Create if requested */
6229 data.mv_size = sizeof(MDB_db);
6230 data.mv_data = &dummy;
6231 memset(&dummy, 0, sizeof(dummy));
6232 dummy.md_root = P_INVALID;
6233 dummy.md_flags = flags & 0xffff;
6234 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6238 /* OK, got info, add to table */
6239 if (rc == MDB_SUCCESS) {
6240 unsigned int slot = unused ? unused : txn->mt_numdbs;
6241 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
6242 txn->mt_dbxs[slot].md_name.mv_size = len;
6243 txn->mt_dbxs[slot].md_rel = NULL;
6244 txn->mt_dbflags[slot] = dbflag;
6245 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
6247 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
6248 mdb_default_cmp(txn, slot);
6251 txn->mt_env->me_numdbs++;
6258 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6260 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6263 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6266 void mdb_close(MDB_env *env, MDB_dbi dbi)
6269 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6271 ptr = env->me_dbxs[dbi].md_name.mv_data;
6272 env->me_dbxs[dbi].md_name.mv_data = NULL;
6273 env->me_dbxs[dbi].md_name.mv_size = 0;
6277 /** Add all the DB's pages to the free list.
6278 * @param[in] mc Cursor on the DB to free.
6279 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6280 * @return 0 on success, non-zero on failure.
6283 mdb_drop0(MDB_cursor *mc, int subs)
6287 rc = mdb_page_search(mc, NULL, 0);
6288 if (rc == MDB_SUCCESS) {
6293 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6294 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6297 mdb_cursor_copy(mc, &mx);
6298 while (mc->mc_snum > 0) {
6299 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6300 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6301 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6302 if (ni->mn_flags & F_SUBDATA) {
6303 mdb_xcursor_init1(mc, ni);
6304 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6310 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6312 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6315 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6320 rc = mdb_cursor_sibling(mc, 1);
6322 /* no more siblings, go back to beginning
6323 * of previous level. (stack was already popped
6324 * by mdb_cursor_sibling)
6326 for (i=1; i<mc->mc_top; i++)
6327 mc->mc_pg[i] = mx.mc_pg[i];
6331 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6332 mc->mc_db->md_root);
6337 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6342 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6345 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6348 rc = mdb_cursor_open(txn, dbi, &mc);
6352 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6356 /* Can't delete the main DB */
6357 if (del && dbi > MAIN_DBI) {
6358 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6360 mdb_close(txn->mt_env, dbi);
6362 txn->mt_dbflags[dbi] |= DB_DIRTY;
6363 txn->mt_dbs[dbi].md_depth = 0;
6364 txn->mt_dbs[dbi].md_branch_pages = 0;
6365 txn->mt_dbs[dbi].md_leaf_pages = 0;
6366 txn->mt_dbs[dbi].md_overflow_pages = 0;
6367 txn->mt_dbs[dbi].md_entries = 0;
6368 txn->mt_dbs[dbi].md_root = P_INVALID;
6371 mdb_cursor_close(mc);
6375 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6377 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6380 txn->mt_dbxs[dbi].md_cmp = cmp;
6384 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6386 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6389 txn->mt_dbxs[dbi].md_dcmp = cmp;
6393 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6395 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6398 txn->mt_dbxs[dbi].md_rel = rel;
6402 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6404 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6407 txn->mt_dbxs[dbi].md_relctx = ctx;