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.
36 #include <sys/types.h>
38 #include <sys/param.h>
44 #ifdef HAVE_SYS_FILE_H
61 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
62 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
65 #if defined(__APPLE__) || defined (BSD)
66 # define MDB_USE_POSIX_SEM 1
67 # define MDB_FDATASYNC fsync
68 #elif defined(ANDROID)
69 # define MDB_FDATASYNC fsync
74 #ifdef MDB_USE_POSIX_SEM
75 #include <semaphore.h>
80 #include <valgrind/memcheck.h>
81 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
82 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
83 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
84 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
85 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
87 #define VGMEMP_CREATE(h,r,z)
88 #define VGMEMP_ALLOC(h,a,s)
89 #define VGMEMP_FREE(h,a)
90 #define VGMEMP_DESTROY(h)
91 #define VGMEMP_DEFINED(a,s)
95 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
96 /* Solaris just defines one or the other */
97 # define LITTLE_ENDIAN 1234
98 # define BIG_ENDIAN 4321
99 # ifdef _LITTLE_ENDIAN
100 # define BYTE_ORDER LITTLE_ENDIAN
102 # define BYTE_ORDER BIG_ENDIAN
105 # define BYTE_ORDER __BYTE_ORDER
109 #ifndef LITTLE_ENDIAN
110 #define LITTLE_ENDIAN __LITTLE_ENDIAN
113 #define BIG_ENDIAN __BIG_ENDIAN
116 #if defined(__i386) || defined(__x86_64)
117 #define MISALIGNED_OK 1
123 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
124 # error "Unknown or unsupported endianness (BYTE_ORDER)"
125 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
126 # error "Two's complement, reasonably sized integer types, please"
129 /** @defgroup internal MDB Internals
132 /** @defgroup compat Windows Compatibility Macros
133 * A bunch of macros to minimize the amount of platform-specific ifdefs
134 * needed throughout the rest of the code. When the features this library
135 * needs are similar enough to POSIX to be hidden in a one-or-two line
136 * replacement, this macro approach is used.
140 #define pthread_t DWORD
141 #define pthread_mutex_t HANDLE
142 #define pthread_key_t DWORD
143 #define pthread_self() GetCurrentThreadId()
144 #define pthread_key_create(x,y) (*(x) = TlsAlloc())
145 #define pthread_key_delete(x) TlsFree(x)
146 #define pthread_getspecific(x) TlsGetValue(x)
147 #define pthread_setspecific(x,y) TlsSetValue(x,y)
148 #define pthread_mutex_unlock(x) ReleaseMutex(x)
149 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
150 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
151 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
152 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
153 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
154 #define getpid() GetCurrentProcessId()
155 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
156 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
157 #define ErrCode() GetLastError()
158 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
159 #define close(fd) CloseHandle(fd)
160 #define munmap(ptr,len) UnmapViewOfFile(ptr)
163 #ifdef MDB_USE_POSIX_SEM
165 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
166 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
167 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
168 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
171 mdb_sem_wait(sem_t *sem)
174 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
179 /** Lock the reader mutex.
181 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
182 /** Unlock the reader mutex.
184 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
186 /** Lock the writer mutex.
187 * Only a single write transaction is allowed at a time. Other writers
188 * will block waiting for this mutex.
190 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
191 /** Unlock the writer mutex.
193 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
194 #endif /* MDB_USE_POSIX_SEM */
196 /** Get the error code for the last failed system function.
198 #define ErrCode() errno
200 /** An abstraction for a file handle.
201 * On POSIX systems file handles are small integers. On Windows
202 * they're opaque pointers.
206 /** A value for an invalid file handle.
207 * Mainly used to initialize file variables and signify that they are
210 #define INVALID_HANDLE_VALUE (-1)
212 /** Get the size of a memory page for the system.
213 * This is the basic size that the platform's memory manager uses, and is
214 * fundamental to the use of memory-mapped files.
216 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
219 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
222 #define MNAME_LEN (sizeof(pthread_mutex_t))
228 /** A flag for opening a file and requesting synchronous data writes.
229 * This is only used when writing a meta page. It's not strictly needed;
230 * we could just do a normal write and then immediately perform a flush.
231 * But if this flag is available it saves us an extra system call.
233 * @note If O_DSYNC is undefined but exists in /usr/include,
234 * preferably set some compiler flag to get the definition.
235 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
238 # define MDB_DSYNC O_DSYNC
242 /** Function for flushing the data of a file. Define this to fsync
243 * if fdatasync() is not supported.
245 #ifndef MDB_FDATASYNC
246 # define MDB_FDATASYNC fdatasync
250 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
261 /** A page number in the database.
262 * Note that 64 bit page numbers are overkill, since pages themselves
263 * already represent 12-13 bits of addressable memory, and the OS will
264 * always limit applications to a maximum of 63 bits of address space.
266 * @note In the #MDB_node structure, we only store 48 bits of this value,
267 * which thus limits us to only 60 bits of addressable data.
269 typedef MDB_ID pgno_t;
271 /** A transaction ID.
272 * See struct MDB_txn.mt_txnid for details.
274 typedef MDB_ID txnid_t;
276 /** @defgroup debug Debug Macros
280 /** Enable debug output.
281 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
282 * read from and written to the database (used for free space management).
287 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
288 # define DPRINTF (void) /* Vararg macros may be unsupported */
290 static int mdb_debug;
291 static txnid_t mdb_debug_start;
293 /** Print a debug message with printf formatting. */
294 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
295 ((void) ((mdb_debug) && \
296 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
298 # define DPRINTF(fmt, ...) ((void) 0)
300 /** Print a debug string.
301 * The string is printed literally, with no format processing.
303 #define DPUTS(arg) DPRINTF("%s", arg)
306 /** A default memory page size.
307 * The actual size is platform-dependent, but we use this for
308 * boot-strapping. We probably should not be using this any more.
309 * The #GET_PAGESIZE() macro is used to get the actual size.
311 * Note that we don't currently support Huge pages. On Linux,
312 * regular data files cannot use Huge pages, and in general
313 * Huge pages aren't actually pageable. We rely on the OS
314 * demand-pager to read our data and page it out when memory
315 * pressure from other processes is high. So until OSs have
316 * actual paging support for Huge pages, they're not viable.
318 #define MDB_PAGESIZE 4096
320 /** The minimum number of keys required in a database page.
321 * Setting this to a larger value will place a smaller bound on the
322 * maximum size of a data item. Data items larger than this size will
323 * be pushed into overflow pages instead of being stored directly in
324 * the B-tree node. This value used to default to 4. With a page size
325 * of 4096 bytes that meant that any item larger than 1024 bytes would
326 * go into an overflow page. That also meant that on average 2-3KB of
327 * each overflow page was wasted space. The value cannot be lower than
328 * 2 because then there would no longer be a tree structure. With this
329 * value, items larger than 2KB will go into overflow pages, and on
330 * average only 1KB will be wasted.
332 #define MDB_MINKEYS 2
334 /** A stamp that identifies a file as an MDB file.
335 * There's nothing special about this value other than that it is easily
336 * recognizable, and it will reflect any byte order mismatches.
338 #define MDB_MAGIC 0xBEEFC0DE
340 /** The version number for a database's file format. */
341 #define MDB_VERSION 1
343 /** The maximum size of a key in the database.
344 * While data items have essentially unbounded size, we require that
345 * keys all fit onto a regular page. This limit could be raised a bit
346 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
348 #define MAXKEYSIZE 511
353 * This is used for printing a hex dump of a key's contents.
355 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
356 /** Display a key in hex.
358 * Invoke a function to display a key in hex.
360 #define DKEY(x) mdb_dkey(x, kbuf)
362 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
366 /** An invalid page number.
367 * Mainly used to denote an empty tree.
369 #define P_INVALID (~(pgno_t)0)
371 /** Test if a flag \b f is set in a flag word \b w. */
372 #define F_ISSET(w, f) (((w) & (f)) == (f))
374 /** Used for offsets within a single page.
375 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
378 typedef uint16_t indx_t;
380 /** Default size of memory map.
381 * This is certainly too small for any actual applications. Apps should always set
382 * the size explicitly using #mdb_env_set_mapsize().
384 #define DEFAULT_MAPSIZE 1048576
386 /** @defgroup readers Reader Lock Table
387 * Readers don't acquire any locks for their data access. Instead, they
388 * simply record their transaction ID in the reader table. The reader
389 * mutex is needed just to find an empty slot in the reader table. The
390 * slot's address is saved in thread-specific data so that subsequent read
391 * transactions started by the same thread need no further locking to proceed.
393 * Since the database uses multi-version concurrency control, readers don't
394 * actually need any locking. This table is used to keep track of which
395 * readers are using data from which old transactions, so that we'll know
396 * when a particular old transaction is no longer in use. Old transactions
397 * that have discarded any data pages can then have those pages reclaimed
398 * for use by a later write transaction.
400 * The lock table is constructed such that reader slots are aligned with the
401 * processor's cache line size. Any slot is only ever used by one thread.
402 * This alignment guarantees that there will be no contention or cache
403 * thrashing as threads update their own slot info, and also eliminates
404 * any need for locking when accessing a slot.
406 * A writer thread will scan every slot in the table to determine the oldest
407 * outstanding reader transaction. Any freed pages older than this will be
408 * reclaimed by the writer. The writer doesn't use any locks when scanning
409 * this table. This means that there's no guarantee that the writer will
410 * see the most up-to-date reader info, but that's not required for correct
411 * operation - all we need is to know the upper bound on the oldest reader,
412 * we don't care at all about the newest reader. So the only consequence of
413 * reading stale information here is that old pages might hang around a
414 * while longer before being reclaimed. That's actually good anyway, because
415 * the longer we delay reclaiming old pages, the more likely it is that a
416 * string of contiguous pages can be found after coalescing old pages from
417 * many old transactions together.
419 * @todo We don't actually do such coalescing yet, we grab pages from one
420 * old transaction at a time.
423 /** Number of slots in the reader table.
424 * This value was chosen somewhat arbitrarily. 126 readers plus a
425 * couple mutexes fit exactly into 8KB on my development machine.
426 * Applications should set the table size using #mdb_env_set_maxreaders().
428 #define DEFAULT_READERS 126
430 /** The size of a CPU cache line in bytes. We want our lock structures
431 * aligned to this size to avoid false cache line sharing in the
433 * This value works for most CPUs. For Itanium this should be 128.
439 /** The information we store in a single slot of the reader table.
440 * In addition to a transaction ID, we also record the process and
441 * thread ID that owns a slot, so that we can detect stale information,
442 * e.g. threads or processes that went away without cleaning up.
443 * @note We currently don't check for stale records. We simply re-init
444 * the table when we know that we're the only process opening the
447 typedef struct MDB_rxbody {
448 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
449 * Multiple readers that start at the same time will probably have the
450 * same ID here. Again, it's not important to exclude them from
451 * anything; all we need to know is which version of the DB they
452 * started from so we can avoid overwriting any data used in that
453 * particular version.
456 /** The process ID of the process owning this reader txn. */
458 /** The thread ID of the thread owning this txn. */
462 /** The actual reader record, with cacheline padding. */
463 typedef struct MDB_reader {
466 /** shorthand for mrb_txnid */
467 #define mr_txnid mru.mrx.mrb_txnid
468 #define mr_pid mru.mrx.mrb_pid
469 #define mr_tid mru.mrx.mrb_tid
470 /** cache line alignment */
471 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
475 /** The header for the reader table.
476 * The table resides in a memory-mapped file. (This is a different file
477 * than is used for the main database.)
479 * For POSIX the actual mutexes reside in the shared memory of this
480 * mapped file. On Windows, mutexes are named objects allocated by the
481 * kernel; we store the mutex names in this mapped file so that other
482 * processes can grab them. This same approach is also used on
483 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
484 * process-shared POSIX mutexes. For these cases where a named object
485 * is used, the object name is derived from a 64 bit FNV hash of the
486 * environment pathname. As such, naming collisions are extremely
487 * unlikely. If a collision occurs, the results are unpredictable.
489 typedef struct MDB_txbody {
490 /** Stamp identifying this as an MDB file. It must be set
493 /** Version number of this lock file. Must be set to #MDB_VERSION. */
494 uint32_t mtb_version;
495 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
496 char mtb_rmname[MNAME_LEN];
498 /** Mutex protecting access to this table.
499 * This is the reader lock that #LOCK_MUTEX_R acquires.
501 pthread_mutex_t mtb_mutex;
503 /** The ID of the last transaction committed to the database.
504 * This is recorded here only for convenience; the value can always
505 * be determined by reading the main database meta pages.
508 /** The number of slots that have been used in the reader table.
509 * This always records the maximum count, it is not decremented
510 * when readers release their slots.
512 unsigned mtb_numreaders;
515 /** The actual reader table definition. */
516 typedef struct MDB_txninfo {
519 #define mti_magic mt1.mtb.mtb_magic
520 #define mti_version mt1.mtb.mtb_version
521 #define mti_mutex mt1.mtb.mtb_mutex
522 #define mti_rmname mt1.mtb.mtb_rmname
523 #define mti_txnid mt1.mtb.mtb_txnid
524 #define mti_numreaders mt1.mtb.mtb_numreaders
525 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
528 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
529 char mt2_wmname[MNAME_LEN];
530 #define mti_wmname mt2.mt2_wmname
532 pthread_mutex_t mt2_wmutex;
533 #define mti_wmutex mt2.mt2_wmutex
535 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
537 MDB_reader mti_readers[1];
541 /** Common header for all page types.
542 * Overflow records occupy a number of contiguous pages with no
543 * headers on any page after the first.
545 typedef struct MDB_page {
546 #define mp_pgno mp_p.p_pgno
547 #define mp_next mp_p.p_next
549 pgno_t p_pgno; /**< page number */
550 void * p_next; /**< for in-memory list of freed structs */
553 /** @defgroup mdb_page Page Flags
555 * Flags for the page headers.
558 #define P_BRANCH 0x01 /**< branch page */
559 #define P_LEAF 0x02 /**< leaf page */
560 #define P_OVERFLOW 0x04 /**< overflow page */
561 #define P_META 0x08 /**< meta page */
562 #define P_DIRTY 0x10 /**< dirty page */
563 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
564 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
566 uint16_t mp_flags; /**< @ref mdb_page */
567 #define mp_lower mp_pb.pb.pb_lower
568 #define mp_upper mp_pb.pb.pb_upper
569 #define mp_pages mp_pb.pb_pages
572 indx_t pb_lower; /**< lower bound of free space */
573 indx_t pb_upper; /**< upper bound of free space */
575 uint32_t pb_pages; /**< number of overflow pages */
577 indx_t mp_ptrs[1]; /**< dynamic size */
580 /** Size of the page header, excluding dynamic data at the end */
581 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
583 /** Address of first usable data byte in a page, after the header */
584 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
586 /** Number of nodes on a page */
587 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
589 /** The amount of space remaining in the page */
590 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
592 /** The percentage of space used in the page, in tenths of a percent. */
593 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
594 ((env)->me_psize - PAGEHDRSZ))
595 /** The minimum page fill factor, in tenths of a percent.
596 * Pages emptier than this are candidates for merging.
598 #define FILL_THRESHOLD 250
600 /** Test if a page is a leaf page */
601 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
602 /** Test if a page is a LEAF2 page */
603 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
604 /** Test if a page is a branch page */
605 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
606 /** Test if a page is an overflow page */
607 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
608 /** Test if a page is a sub page */
609 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
611 /** The number of overflow pages needed to store the given size. */
612 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
614 /** Header for a single key/data pair within a page.
615 * We guarantee 2-byte alignment for nodes.
617 typedef struct MDB_node {
618 /** lo and hi are used for data size on leaf nodes and for
619 * child pgno on branch nodes. On 64 bit platforms, flags
620 * is also used for pgno. (Branch nodes have no flags).
621 * They are in host byte order in case that lets some
622 * accesses be optimized into a 32-bit word access.
624 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
625 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
626 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
627 /** @defgroup mdb_node Node Flags
629 * Flags for node headers.
632 #define F_BIGDATA 0x01 /**< data put on overflow page */
633 #define F_SUBDATA 0x02 /**< data is a sub-database */
634 #define F_DUPDATA 0x04 /**< data has duplicates */
636 /** valid flags for #mdb_node_add() */
637 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
640 unsigned short mn_flags; /**< @ref mdb_node */
641 unsigned short mn_ksize; /**< key size */
642 char mn_data[1]; /**< key and data are appended here */
645 /** Size of the node header, excluding dynamic data at the end */
646 #define NODESIZE offsetof(MDB_node, mn_data)
648 /** Bit position of top word in page number, for shifting mn_flags */
649 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
651 /** Size of a node in a branch page with a given key.
652 * This is just the node header plus the key, there is no data.
654 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
656 /** Size of a node in a leaf page with a given key and data.
657 * This is node header plus key plus data size.
659 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
661 /** Address of node \b i in page \b p */
662 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
664 /** Address of the key for the node */
665 #define NODEKEY(node) (void *)((node)->mn_data)
667 /** Address of the data for a node */
668 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
670 /** Get the page number pointed to by a branch node */
671 #define NODEPGNO(node) \
672 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
673 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
674 /** Set the page number in a branch node */
675 #define SETPGNO(node,pgno) do { \
676 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
677 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
679 /** Get the size of the data in a leaf node */
680 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
681 /** Set the size of the data for a leaf node */
682 #define SETDSZ(node,size) do { \
683 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
684 /** The size of a key in a node */
685 #define NODEKSZ(node) ((node)->mn_ksize)
687 /** Copy a page number from src to dst */
689 #define COPY_PGNO(dst,src) dst = src
691 #if SIZE_MAX > 4294967295UL
692 #define COPY_PGNO(dst,src) do { \
693 unsigned short *s, *d; \
694 s = (unsigned short *)&(src); \
695 d = (unsigned short *)&(dst); \
702 #define COPY_PGNO(dst,src) do { \
703 unsigned short *s, *d; \
704 s = (unsigned short *)&(src); \
705 d = (unsigned short *)&(dst); \
711 /** The address of a key in a LEAF2 page.
712 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
713 * There are no node headers, keys are stored contiguously.
715 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
717 /** Set the \b node's key into \b key, if requested. */
718 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
719 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
721 /** Information about a single database in the environment. */
722 typedef struct MDB_db {
723 uint32_t md_pad; /**< also ksize for LEAF2 pages */
724 uint16_t md_flags; /**< @ref mdb_open */
725 uint16_t md_depth; /**< depth of this tree */
726 pgno_t md_branch_pages; /**< number of internal pages */
727 pgno_t md_leaf_pages; /**< number of leaf pages */
728 pgno_t md_overflow_pages; /**< number of overflow pages */
729 size_t md_entries; /**< number of data items */
730 pgno_t md_root; /**< the root page of this tree */
733 /** Handle for the DB used to track free pages. */
735 /** Handle for the default DB. */
738 /** Meta page content. */
739 typedef struct MDB_meta {
740 /** Stamp identifying this as an MDB file. It must be set
743 /** Version number of this lock file. Must be set to #MDB_VERSION. */
745 void *mm_address; /**< address for fixed mapping */
746 size_t mm_mapsize; /**< size of mmap region */
747 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
748 /** The size of pages used in this DB */
749 #define mm_psize mm_dbs[0].md_pad
750 /** Any persistent environment flags. @ref mdb_env */
751 #define mm_flags mm_dbs[0].md_flags
752 pgno_t mm_last_pg; /**< last used page in file */
753 txnid_t mm_txnid; /**< txnid that committed this page */
756 /** Buffer for a stack-allocated dirty page.
757 * The members define size and alignment, and silence type
758 * aliasing warnings. They are not used directly; that could
759 * mean incorrectly using several union members in parallel.
761 typedef union MDB_pagebuf {
762 char mb_raw[MDB_PAGESIZE];
765 char mm_pad[PAGEHDRSZ];
770 /** Auxiliary DB info.
771 * The information here is mostly static/read-only. There is
772 * only a single copy of this record in the environment.
774 typedef struct MDB_dbx {
775 MDB_val md_name; /**< name of the database */
776 MDB_cmp_func *md_cmp; /**< function for comparing keys */
777 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
778 MDB_rel_func *md_rel; /**< user relocate function */
779 void *md_relctx; /**< user-provided context for md_rel */
782 /** A database transaction.
783 * Every operation requires a transaction handle.
786 MDB_txn *mt_parent; /**< parent of a nested txn */
787 MDB_txn *mt_child; /**< nested txn under this txn */
788 pgno_t mt_next_pgno; /**< next unallocated page */
789 /** The ID of this transaction. IDs are integers incrementing from 1.
790 * Only committed write transactions increment the ID. If a transaction
791 * aborts, the ID may be re-used by the next writer.
794 MDB_env *mt_env; /**< the DB environment */
795 /** The list of pages that became unused during this transaction.
799 MDB_ID2L dirty_list; /**< modified pages */
800 MDB_reader *reader; /**< this thread's slot in the reader table */
802 /** Array of records for each DB known in the environment. */
804 /** Array of MDB_db records for each known DB */
806 /** @defgroup mt_dbflag Transaction DB Flags
810 #define DB_DIRTY 0x01 /**< DB was written in this txn */
811 #define DB_STALE 0x02 /**< DB record is older than txnID */
813 /** Array of cursors for each DB */
814 MDB_cursor **mt_cursors;
815 /** Array of flags for each DB */
816 unsigned char *mt_dbflags;
817 /** Number of DB records in use. This number only ever increments;
818 * we don't decrement it when individual DB handles are closed.
822 /** @defgroup mdb_txn Transaction Flags
826 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
827 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
829 unsigned int mt_flags; /**< @ref mdb_txn */
830 /** Tracks which of the two meta pages was used at the start
831 * of this transaction.
833 unsigned int mt_toggle;
836 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
837 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
838 * raise this on a 64 bit machine.
840 #define CURSOR_STACK 32
844 /** Cursors are used for all DB operations */
846 /** Next cursor on this DB in this txn */
848 /** Original cursor if this is a shadow */
850 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
851 struct MDB_xcursor *mc_xcursor;
852 /** The transaction that owns this cursor */
854 /** The database handle this cursor operates on */
856 /** The database record for this cursor */
858 /** The database auxiliary record for this cursor */
860 /** The @ref mt_dbflag for this database */
861 unsigned char *mc_dbflag;
862 unsigned short mc_snum; /**< number of pushed pages */
863 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
864 /** @defgroup mdb_cursor Cursor Flags
866 * Cursor state flags.
869 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
870 #define C_EOF 0x02 /**< No more data */
871 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
872 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
873 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
874 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
876 unsigned int mc_flags; /**< @ref mdb_cursor */
877 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
878 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
881 /** Context for sorted-dup records.
882 * We could have gone to a fully recursive design, with arbitrarily
883 * deep nesting of sub-databases. But for now we only handle these
884 * levels - main DB, optional sub-DB, sorted-duplicate DB.
886 typedef struct MDB_xcursor {
887 /** A sub-cursor for traversing the Dup DB */
888 MDB_cursor mx_cursor;
889 /** The database record for this Dup DB */
891 /** The auxiliary DB record for this Dup DB */
893 /** The @ref mt_dbflag for this Dup DB */
894 unsigned char mx_dbflag;
897 /** A set of pages freed by an earlier transaction. */
898 typedef struct MDB_oldpages {
899 /** Usually we only read one record from the FREEDB at a time, but
900 * in case we read more, this will chain them together.
902 struct MDB_oldpages *mo_next;
903 /** The ID of the transaction in which these pages were freed. */
905 /** An #MDB_IDL of the pages */
906 pgno_t mo_pages[1]; /* dynamic */
909 /** The database environment. */
911 HANDLE me_fd; /**< The main data file */
912 HANDLE me_lfd; /**< The lock file */
913 HANDLE me_mfd; /**< just for writing the meta pages */
914 /** Failed to update the meta page. Probably an I/O error. */
915 #define MDB_FATAL_ERROR 0x80000000U
916 uint32_t me_flags; /**< @ref mdb_env */
917 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
918 unsigned int me_maxreaders; /**< size of the reader table */
919 unsigned int me_numreaders; /**< max numreaders set by this env */
920 MDB_dbi me_numdbs; /**< number of DBs opened */
921 MDB_dbi me_maxdbs; /**< size of the DB table */
922 pid_t me_pid; /**< process ID of this env */
923 char *me_path; /**< path to the DB files */
924 char *me_map; /**< the memory map of the data file */
925 MDB_txninfo *me_txns; /**< the memory map of the lock file */
926 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
927 MDB_txn *me_txn; /**< current write transaction */
928 size_t me_mapsize; /**< size of the data memory map */
929 off_t me_size; /**< current file size */
930 pgno_t me_maxpg; /**< me_mapsize / me_psize */
931 txnid_t me_pgfirst; /**< ID of first old page record we used */
932 txnid_t me_pglast; /**< ID of last old page record we used */
933 MDB_dbx *me_dbxs; /**< array of static DB info */
934 uint16_t *me_dbflags; /**< array of DB flags */
935 MDB_oldpages *me_pghead; /**< list of old page records */
936 MDB_oldpages *me_pgfree; /**< list of page records to free */
937 pthread_key_t me_txkey; /**< thread-key for readers */
938 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
939 /** IDL of pages that became unused in a write txn */
941 /** ID2L of pages that were written during a write txn */
942 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
944 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
946 #elif defined(MDB_USE_POSIX_SEM)
947 sem_t *me_rmutex; /* Shared mutexes are not supported */
951 /** max number of pages to commit in one writev() call */
952 #define MDB_COMMIT_PAGES 64
953 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
954 #undef MDB_COMMIT_PAGES
955 #define MDB_COMMIT_PAGES IOV_MAX
958 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
959 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
960 static int mdb_page_touch(MDB_cursor *mc);
962 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
963 static int mdb_page_search_root(MDB_cursor *mc,
964 MDB_val *key, int modify);
965 #define MDB_PS_MODIFY 1
966 #define MDB_PS_ROOTONLY 2
967 static int mdb_page_search(MDB_cursor *mc,
968 MDB_val *key, int flags);
969 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
971 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
972 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
973 pgno_t newpgno, unsigned int nflags);
975 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
976 static int mdb_env_pick_meta(const MDB_env *env);
977 static int mdb_env_write_meta(MDB_txn *txn);
978 static void mdb_env_close0(MDB_env *env, int excl);
980 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
981 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
982 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
983 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
984 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
985 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
986 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
987 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
988 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
990 static int mdb_rebalance(MDB_cursor *mc);
991 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
993 static void mdb_cursor_pop(MDB_cursor *mc);
994 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
996 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
997 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
998 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
999 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1000 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1002 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1003 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1005 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1006 static void mdb_xcursor_init0(MDB_cursor *mc);
1007 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1009 static int mdb_drop0(MDB_cursor *mc, int subs);
1010 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1013 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1017 static SECURITY_DESCRIPTOR mdb_null_sd;
1018 static SECURITY_ATTRIBUTES mdb_all_sa;
1019 static int mdb_sec_inited;
1022 /** Return the library version info. */
1024 mdb_version(int *major, int *minor, int *patch)
1026 if (major) *major = MDB_VERSION_MAJOR;
1027 if (minor) *minor = MDB_VERSION_MINOR;
1028 if (patch) *patch = MDB_VERSION_PATCH;
1029 return MDB_VERSION_STRING;
1032 /** Table of descriptions for MDB @ref errors */
1033 static char *const mdb_errstr[] = {
1034 "MDB_KEYEXIST: Key/data pair already exists",
1035 "MDB_NOTFOUND: No matching key/data pair found",
1036 "MDB_PAGE_NOTFOUND: Requested page not found",
1037 "MDB_CORRUPTED: Located page was wrong type",
1038 "MDB_PANIC: Update of meta page failed",
1039 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1040 "MDB_INVALID: File is not an MDB file",
1041 "MDB_MAP_FULL: Environment mapsize limit reached",
1042 "MDB_DBS_FULL: Environment maxdbs limit reached",
1043 "MDB_READERS_FULL: Environment maxreaders limit reached",
1044 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1045 "MDB_TXN_FULL: Nested transaction has too many dirty pages - transaction too big",
1046 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1047 "MDB_PAGE_FULL: Internal error - page has no more space"
1051 mdb_strerror(int err)
1055 return ("Successful return: 0");
1057 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1058 i = err - MDB_KEYEXIST;
1059 return mdb_errstr[i];
1062 return strerror(err);
1066 /** Display a key in hexadecimal and return the address of the result.
1067 * @param[in] key the key to display
1068 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1069 * @return The key in hexadecimal form.
1072 mdb_dkey(MDB_val *key, char *buf)
1075 unsigned char *c = key->mv_data;
1077 if (key->mv_size > MAXKEYSIZE)
1078 return "MAXKEYSIZE";
1079 /* may want to make this a dynamic check: if the key is mostly
1080 * printable characters, print it as-is instead of converting to hex.
1084 for (i=0; i<key->mv_size; i++)
1085 ptr += sprintf(ptr, "%02x", *c++);
1087 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1092 /** Display all the keys in the page. */
1094 mdb_page_list(MDB_page *mp)
1097 unsigned int i, nkeys, nsize;
1101 nkeys = NUMKEYS(mp);
1102 fprintf(stderr, "numkeys %d\n", nkeys);
1103 for (i=0; i<nkeys; i++) {
1104 node = NODEPTR(mp, i);
1105 key.mv_size = node->mn_ksize;
1106 key.mv_data = node->mn_data;
1107 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1108 if (F_ISSET(node->mn_flags, F_BIGDATA))
1109 nsize += sizeof(pgno_t);
1111 nsize += NODEDSZ(node);
1112 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1117 mdb_cursor_chk(MDB_cursor *mc)
1123 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1124 for (i=0; i<mc->mc_top; i++) {
1126 node = NODEPTR(mp, mc->mc_ki[i]);
1127 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1130 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1136 /** Count all the pages in each DB and in the freelist
1137 * and make sure it matches the actual number of pages
1140 static void mdb_audit(MDB_txn *txn)
1144 MDB_ID freecount, count;
1149 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1150 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1151 freecount += *(MDB_ID *)data.mv_data;
1154 for (i = 0; i<txn->mt_numdbs; i++) {
1155 MDB_xcursor mx, *mxp;
1156 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1157 mdb_cursor_init(&mc, txn, i, mxp);
1158 if (txn->mt_dbs[i].md_root == P_INVALID)
1160 count += txn->mt_dbs[i].md_branch_pages +
1161 txn->mt_dbs[i].md_leaf_pages +
1162 txn->mt_dbs[i].md_overflow_pages;
1163 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1164 mdb_page_search(&mc, NULL, 0);
1168 mp = mc.mc_pg[mc.mc_top];
1169 for (j=0; j<NUMKEYS(mp); j++) {
1170 MDB_node *leaf = NODEPTR(mp, j);
1171 if (leaf->mn_flags & F_SUBDATA) {
1173 memcpy(&db, NODEDATA(leaf), sizeof(db));
1174 count += db.md_branch_pages + db.md_leaf_pages +
1175 db.md_overflow_pages;
1179 while (mdb_cursor_sibling(&mc, 1) == 0);
1182 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1183 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1184 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1190 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1192 return txn->mt_dbxs[dbi].md_cmp(a, b);
1196 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1198 if (txn->mt_dbxs[dbi].md_dcmp)
1199 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1201 return EINVAL; /* too bad you can't distinguish this from a valid result */
1204 /** Allocate a single page.
1205 * Re-use old malloc'd pages first, otherwise just malloc.
1208 mdb_page_malloc(MDB_cursor *mc) {
1210 size_t sz = mc->mc_txn->mt_env->me_psize;
1211 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1212 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1213 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1214 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1215 } else if ((ret = malloc(sz)) != NULL) {
1216 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1221 /** Allocate pages for writing.
1222 * If there are free pages available from older transactions, they
1223 * will be re-used first. Otherwise a new page will be allocated.
1224 * @param[in] mc cursor A cursor handle identifying the transaction and
1225 * database for which we are allocating.
1226 * @param[in] num the number of pages to allocate.
1227 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1228 * will always be satisfied by a single contiguous chunk of memory.
1229 * @return 0 on success, non-zero on failure.
1232 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1234 MDB_txn *txn = mc->mc_txn;
1236 pgno_t pgno = P_INVALID;
1241 /* The free list won't have any content at all until txn 2 has
1242 * committed. The pages freed by txn 2 will be unreferenced
1243 * after txn 3 commits, and so will be safe to re-use in txn 4.
1245 if (txn->mt_txnid > 3) {
1247 if (!txn->mt_env->me_pghead &&
1248 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1249 /* See if there's anything in the free DB */
1255 txnid_t *kptr, last;
1257 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1258 if (!txn->mt_env->me_pgfirst) {
1259 mdb_page_search(&m2, NULL, 0);
1260 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1261 kptr = (txnid_t *)NODEKEY(leaf);
1268 last = txn->mt_env->me_pglast + 1;
1270 key.mv_data = &last;
1271 key.mv_size = sizeof(last);
1272 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1275 last = *(txnid_t *)key.mv_data;
1278 /* Unusable if referred by a meta page or reader... */
1280 if (last < txn->mt_txnid-1) {
1281 j = txn->mt_env->me_txns->mti_numreaders;
1282 r = txn->mt_env->me_txns->mti_readers + j;
1283 for (j = -j; j && (last<r[j].mr_txnid || !r[j].mr_pid); j++) ;
1287 /* It's usable, grab it.
1292 if (!txn->mt_env->me_pgfirst) {
1293 mdb_node_read(txn, leaf, &data);
1295 txn->mt_env->me_pglast = last;
1296 if (!txn->mt_env->me_pgfirst)
1297 txn->mt_env->me_pgfirst = last;
1298 idl = (MDB_ID *) data.mv_data;
1299 /* We might have a zero-length IDL due to freelist growth
1300 * during a prior commit
1302 if (!idl[0]) goto again;
1303 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1306 mop->mo_next = txn->mt_env->me_pghead;
1307 mop->mo_txnid = last;
1308 txn->mt_env->me_pghead = mop;
1309 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1314 DPRINTF("IDL read txn %zu root %zu num %zu",
1315 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1316 for (i=0; i<idl[0]; i++) {
1317 DPRINTF("IDL %zu", idl[i+1]);
1324 if (txn->mt_env->me_pghead) {
1325 MDB_oldpages *mop = txn->mt_env->me_pghead;
1327 /* FIXME: For now, always use fresh pages. We
1328 * really ought to search the free list for a
1333 /* peel pages off tail, so we only have to truncate the list */
1334 pgno = MDB_IDL_LAST(mop->mo_pages);
1335 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1337 if (mop->mo_pages[2] > mop->mo_pages[1])
1338 mop->mo_pages[0] = 0;
1342 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1343 txn->mt_env->me_pghead = mop->mo_next;
1344 if (mc->mc_dbi == FREE_DBI) {
1345 mop->mo_next = txn->mt_env->me_pgfree;
1346 txn->mt_env->me_pgfree = mop;
1355 if (pgno == P_INVALID) {
1356 /* DB size is maxed out */
1357 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1358 DPUTS("DB size maxed out");
1359 return MDB_MAP_FULL;
1362 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1363 if (pgno == P_INVALID) {
1364 pgno = txn->mt_next_pgno;
1365 txn->mt_next_pgno += num;
1367 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1370 if (txn->mt_env->me_dpages && num == 1) {
1371 np = txn->mt_env->me_dpages;
1372 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1373 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1374 txn->mt_env->me_dpages = np->mp_next;
1376 size_t sz = txn->mt_env->me_psize * num;
1377 if ((np = malloc(sz)) == NULL)
1379 VGMEMP_ALLOC(txn->mt_env, np, sz);
1381 if (pgno == P_INVALID) {
1382 np->mp_pgno = txn->mt_next_pgno;
1383 txn->mt_next_pgno += num;
1388 mid.mid = np->mp_pgno;
1390 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1391 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1393 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1400 /** Copy a page: avoid copying unused portions of the page.
1401 * @param[in] dst page to copy into
1402 * @param[in] src page to copy from
1405 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1407 dst->mp_flags = src->mp_flags | P_DIRTY;
1408 dst->mp_pages = src->mp_pages;
1410 if (IS_LEAF2(src)) {
1411 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1413 unsigned int i, nkeys = NUMKEYS(src);
1414 for (i=0; i<nkeys; i++)
1415 dst->mp_ptrs[i] = src->mp_ptrs[i];
1416 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1417 psize - src->mp_upper);
1421 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1422 * @param[in] mc cursor pointing to the page to be touched
1423 * @return 0 on success, non-zero on failure.
1426 mdb_page_touch(MDB_cursor *mc)
1428 MDB_page *mp = mc->mc_pg[mc->mc_top];
1432 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1434 if ((rc = mdb_page_alloc(mc, 1, &np)))
1436 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1437 assert(mp->mp_pgno != np->mp_pgno);
1438 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1440 /* If page isn't full, just copy the used portion */
1441 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1444 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1446 np->mp_flags |= P_DIRTY;
1451 /* Adjust other cursors pointing to mp */
1452 if (mc->mc_flags & C_SUB) {
1453 MDB_cursor *m2, *m3;
1454 MDB_dbi dbi = mc->mc_dbi-1;
1456 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1457 if (m2 == mc) continue;
1458 m3 = &m2->mc_xcursor->mx_cursor;
1459 if (m3->mc_snum < mc->mc_snum) continue;
1460 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1461 m3->mc_pg[mc->mc_top] = mp;
1467 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1468 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1469 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1470 m2->mc_pg[mc->mc_top] = mp;
1474 mc->mc_pg[mc->mc_top] = mp;
1475 /** If this page has a parent, update the parent to point to
1479 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1481 mc->mc_db->md_root = mp->mp_pgno;
1482 } else if (mc->mc_txn->mt_parent) {
1485 /* If txn has a parent, make sure the page is in our
1488 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1489 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1490 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1491 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1492 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1493 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1494 mc->mc_pg[mc->mc_top] = mp;
1500 np = mdb_page_malloc(mc);
1503 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1504 mid.mid = np->mp_pgno;
1506 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1514 mdb_env_sync(MDB_env *env, int force)
1517 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1518 if (env->me_flags & MDB_WRITEMAP) {
1519 int flags = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
1520 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1523 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1527 if (MDB_FDATASYNC(env->me_fd))
1534 /** Make shadow copies of all of parent txn's cursors */
1536 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1538 MDB_cursor *mc, *m2;
1539 unsigned int i, j, size;
1541 for (i=0;i<src->mt_numdbs; i++) {
1542 if (src->mt_cursors[i]) {
1543 size = sizeof(MDB_cursor);
1544 if (src->mt_cursors[i]->mc_xcursor)
1545 size += sizeof(MDB_xcursor);
1546 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1553 mc->mc_db = &dst->mt_dbs[i];
1554 mc->mc_dbx = m2->mc_dbx;
1555 mc->mc_dbflag = &dst->mt_dbflags[i];
1556 mc->mc_snum = m2->mc_snum;
1557 mc->mc_top = m2->mc_top;
1558 mc->mc_flags = m2->mc_flags | C_SHADOW;
1559 for (j=0; j<mc->mc_snum; j++) {
1560 mc->mc_pg[j] = m2->mc_pg[j];
1561 mc->mc_ki[j] = m2->mc_ki[j];
1563 if (m2->mc_xcursor) {
1564 MDB_xcursor *mx, *mx2;
1565 mx = (MDB_xcursor *)(mc+1);
1566 mc->mc_xcursor = mx;
1567 mx2 = m2->mc_xcursor;
1568 mx->mx_db = mx2->mx_db;
1569 mx->mx_dbx = mx2->mx_dbx;
1570 mx->mx_dbflag = mx2->mx_dbflag;
1571 mx->mx_cursor.mc_txn = dst;
1572 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1573 mx->mx_cursor.mc_db = &mx->mx_db;
1574 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1575 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1576 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1577 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1578 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1579 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1580 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1581 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1584 mc->mc_xcursor = NULL;
1586 mc->mc_next = dst->mt_cursors[i];
1587 dst->mt_cursors[i] = mc;
1594 /** Merge shadow cursors back into parent's */
1596 mdb_cursor_merge(MDB_txn *txn)
1599 for (i=0; i<txn->mt_numdbs; i++) {
1600 if (txn->mt_cursors[i]) {
1602 while ((mc = txn->mt_cursors[i])) {
1603 txn->mt_cursors[i] = mc->mc_next;
1604 if (mc->mc_flags & C_SHADOW) {
1605 MDB_cursor *m2 = mc->mc_orig;
1607 m2->mc_snum = mc->mc_snum;
1608 m2->mc_top = mc->mc_top;
1609 for (j=0; j<mc->mc_snum; j++) {
1610 m2->mc_pg[j] = mc->mc_pg[j];
1611 m2->mc_ki[j] = mc->mc_ki[j];
1614 if (mc->mc_flags & C_ALLOCD)
1622 mdb_txn_reset0(MDB_txn *txn);
1624 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1625 * @param[in] txn the transaction handle to initialize
1626 * @return 0 on success, non-zero on failure. This can only
1627 * fail for read-only transactions, and then only if the
1628 * reader table is full.
1631 mdb_txn_renew0(MDB_txn *txn)
1633 MDB_env *env = txn->mt_env;
1637 txn->mt_numdbs = env->me_numdbs;
1638 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1640 if (txn->mt_flags & MDB_TXN_RDONLY) {
1641 MDB_reader *r = pthread_getspecific(env->me_txkey);
1643 pid_t pid = env->me_pid;
1644 pthread_t tid = pthread_self();
1647 for (i=0; i<env->me_txns->mti_numreaders; i++)
1648 if (env->me_txns->mti_readers[i].mr_pid == 0)
1650 if (i == env->me_maxreaders) {
1651 UNLOCK_MUTEX_R(env);
1652 return MDB_READERS_FULL;
1654 env->me_txns->mti_readers[i].mr_pid = pid;
1655 env->me_txns->mti_readers[i].mr_tid = tid;
1656 if (i >= env->me_txns->mti_numreaders)
1657 env->me_txns->mti_numreaders = i+1;
1658 /* Save numreaders for un-mutexed mdb_env_close() */
1659 env->me_numreaders = env->me_txns->mti_numreaders;
1660 UNLOCK_MUTEX_R(env);
1661 r = &env->me_txns->mti_readers[i];
1662 pthread_setspecific(env->me_txkey, r);
1664 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1665 txn->mt_toggle = txn->mt_txnid & 1;
1666 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1667 txn->mt_u.reader = r;
1671 txn->mt_txnid = env->me_txns->mti_txnid;
1672 txn->mt_toggle = txn->mt_txnid & 1;
1673 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1676 if (txn->mt_txnid == mdb_debug_start)
1679 txn->mt_u.dirty_list = env->me_dirty_list;
1680 txn->mt_u.dirty_list[0].mid = 0;
1681 txn->mt_free_pgs = env->me_free_pgs;
1682 txn->mt_free_pgs[0] = 0;
1686 /* Copy the DB info and flags */
1687 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1688 for (i=2; i<txn->mt_numdbs; i++)
1689 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1690 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1691 if (txn->mt_numdbs > 2)
1692 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1698 mdb_txn_renew(MDB_txn *txn)
1702 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1705 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1706 DPUTS("environment had fatal error, must shutdown!");
1710 rc = mdb_txn_renew0(txn);
1711 if (rc == MDB_SUCCESS) {
1712 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1713 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1714 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1720 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1725 if (env->me_flags & MDB_FATAL_ERROR) {
1726 DPUTS("environment had fatal error, must shutdown!");
1729 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1732 /* Nested transactions: Max 1 child, write txns only, no writemap */
1733 if (parent->mt_child ||
1734 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1735 (env->me_flags & MDB_WRITEMAP))
1740 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1741 if (!(flags & MDB_RDONLY))
1742 size += env->me_maxdbs * sizeof(MDB_cursor *);
1744 if ((txn = calloc(1, size)) == NULL) {
1745 DPRINTF("calloc: %s", strerror(ErrCode()));
1748 txn->mt_dbs = (MDB_db *)(txn+1);
1749 if (flags & MDB_RDONLY) {
1750 txn->mt_flags |= MDB_TXN_RDONLY;
1751 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1753 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1754 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1759 txn->mt_free_pgs = mdb_midl_alloc();
1760 if (!txn->mt_free_pgs) {
1764 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1765 if (!txn->mt_u.dirty_list) {
1766 free(txn->mt_free_pgs);
1770 txn->mt_txnid = parent->mt_txnid;
1771 txn->mt_toggle = parent->mt_toggle;
1772 txn->mt_u.dirty_list[0].mid = 0;
1773 txn->mt_free_pgs[0] = 0;
1774 txn->mt_next_pgno = parent->mt_next_pgno;
1775 parent->mt_child = txn;
1776 txn->mt_parent = parent;
1777 txn->mt_numdbs = parent->mt_numdbs;
1778 txn->mt_dbxs = parent->mt_dbxs;
1779 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1780 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1781 mdb_cursor_shadow(parent, txn);
1784 rc = mdb_txn_renew0(txn);
1790 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1791 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1792 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1798 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1799 * @param[in] txn the transaction handle to reset
1802 mdb_txn_reset0(MDB_txn *txn)
1804 MDB_env *env = txn->mt_env;
1806 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1807 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1813 /* close(free) all cursors */
1814 for (i=0; i<txn->mt_numdbs; i++) {
1815 if (txn->mt_cursors[i]) {
1817 while ((mc = txn->mt_cursors[i])) {
1818 txn->mt_cursors[i] = mc->mc_next;
1819 if (mc->mc_flags & C_ALLOCD)
1825 if (!(env->me_flags & MDB_WRITEMAP)) {
1826 /* return all dirty pages to dpage list */
1827 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1828 dp = txn->mt_u.dirty_list[i].mptr;
1829 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1830 dp->mp_next = txn->mt_env->me_dpages;
1831 VGMEMP_FREE(txn->mt_env, dp);
1832 txn->mt_env->me_dpages = dp;
1834 /* large pages just get freed directly */
1835 VGMEMP_FREE(txn->mt_env, dp);
1841 if (txn->mt_parent) {
1842 txn->mt_parent->mt_child = NULL;
1843 mdb_midl_free(txn->mt_free_pgs);
1844 free(txn->mt_u.dirty_list);
1847 if (mdb_midl_shrink(&txn->mt_free_pgs))
1848 env->me_free_pgs = txn->mt_free_pgs;
1851 while ((mop = txn->mt_env->me_pghead)) {
1852 txn->mt_env->me_pghead = mop->mo_next;
1855 txn->mt_env->me_pgfirst = 0;
1856 txn->mt_env->me_pglast = 0;
1859 /* The writer mutex was locked in mdb_txn_begin. */
1860 UNLOCK_MUTEX_W(env);
1865 mdb_txn_reset(MDB_txn *txn)
1870 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1871 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1872 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1874 mdb_txn_reset0(txn);
1878 mdb_txn_abort(MDB_txn *txn)
1883 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1884 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1885 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1888 mdb_txn_abort(txn->mt_child);
1890 mdb_txn_reset0(txn);
1895 mdb_txn_commit(MDB_txn *txn)
1903 pgno_t next, freecnt;
1906 assert(txn != NULL);
1907 assert(txn->mt_env != NULL);
1909 if (txn->mt_child) {
1910 mdb_txn_commit(txn->mt_child);
1911 txn->mt_child = NULL;
1916 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1917 if (txn->mt_numdbs > env->me_numdbs) {
1918 /* update the DB flags */
1920 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
1921 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
1928 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1929 DPUTS("error flag is set, can't commit");
1931 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1936 /* Merge (and close) our cursors with parent's */
1937 mdb_cursor_merge(txn);
1939 if (txn->mt_parent) {
1945 /* Update parent's DB table */
1946 ip = &txn->mt_parent->mt_dbs[2];
1947 jp = &txn->mt_dbs[2];
1948 for (i = 2; i < txn->mt_numdbs; i++) {
1949 if (ip->md_root != jp->md_root)
1953 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1955 /* Append our free list to parent's */
1956 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1958 mdb_midl_free(txn->mt_free_pgs);
1960 /* Merge our dirty list with parent's */
1961 dst = txn->mt_parent->mt_u.dirty_list;
1962 src = txn->mt_u.dirty_list;
1963 x = mdb_mid2l_search(dst, src[1].mid);
1964 for (y=1; y<=src[0].mid; y++) {
1965 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1969 dst[x].mptr = src[y].mptr;
1972 for (; y<=src[0].mid; y++) {
1973 if (++x >= MDB_IDL_UM_MAX) {
1975 return MDB_TXN_FULL;
1980 free(txn->mt_u.dirty_list);
1981 txn->mt_parent->mt_child = NULL;
1986 if (txn != env->me_txn) {
1987 DPUTS("attempt to commit unknown transaction");
1992 if (!txn->mt_u.dirty_list[0].mid)
1995 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1996 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1998 /* Update DB root pointers. Their pages have already been
1999 * touched so this is all in-place and cannot fail.
2001 if (txn->mt_numdbs > 2) {
2004 data.mv_size = sizeof(MDB_db);
2006 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2007 for (i = 2; i < txn->mt_numdbs; i++) {
2008 if (txn->mt_dbflags[i] & DB_DIRTY) {
2009 data.mv_data = &txn->mt_dbs[i];
2010 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2015 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2017 /* should only be one record now */
2018 if (env->me_pghead) {
2019 /* make sure first page of freeDB is touched and on freelist */
2020 mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2023 /* Delete IDLs we used from the free list */
2024 if (env->me_pgfirst) {
2029 key.mv_size = sizeof(cur);
2030 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
2033 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
2034 rc = mdb_cursor_del(&mc, 0);
2040 env->me_pgfirst = 0;
2044 /* save to free list */
2046 freecnt = txn->mt_free_pgs[0];
2047 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2050 /* make sure last page of freeDB is touched and on freelist */
2051 key.mv_size = MAXKEYSIZE+1;
2053 mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2055 mdb_midl_sort(txn->mt_free_pgs);
2059 MDB_IDL idl = txn->mt_free_pgs;
2060 DPRINTF("IDL write txn %zu root %zu num %zu",
2061 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2062 for (i=0; i<idl[0]; i++) {
2063 DPRINTF("IDL %zu", idl[i+1]);
2067 /* write to last page of freeDB */
2068 key.mv_size = sizeof(pgno_t);
2069 key.mv_data = &txn->mt_txnid;
2070 data.mv_data = txn->mt_free_pgs;
2071 /* The free list can still grow during this call,
2072 * despite the pre-emptive touches above. So check
2073 * and make sure the entire thing got written.
2076 freecnt = txn->mt_free_pgs[0];
2077 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2078 rc = mdb_cursor_put(&mc, &key, &data, 0);
2083 } while (freecnt != txn->mt_free_pgs[0]);
2085 /* should only be one record now */
2087 if (env->me_pghead) {
2093 mop = env->me_pghead;
2095 key.mv_size = sizeof(id);
2097 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2098 data.mv_data = mop->mo_pages;
2099 orig = mop->mo_pages[0];
2100 /* These steps may grow the freelist again
2101 * due to freed overflow pages...
2103 mdb_cursor_put(&mc, &key, &data, 0);
2104 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2105 /* could have been used again here */
2106 if (mop->mo_pages[0] != orig) {
2107 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2108 data.mv_data = mop->mo_pages;
2110 mdb_cursor_put(&mc, &key, &data, 0);
2112 env->me_pghead = NULL;
2115 /* was completely used up */
2116 mdb_cursor_del(&mc, 0);
2120 env->me_pgfirst = 0;
2124 while (env->me_pgfree) {
2125 MDB_oldpages *mop = env->me_pgfree;
2126 env->me_pgfree = mop->mo_next;
2130 /* Check for growth of freelist again */
2131 if (freecnt != txn->mt_free_pgs[0])
2134 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2135 if (mdb_midl_shrink(&txn->mt_free_pgs))
2136 env->me_free_pgs = txn->mt_free_pgs;
2143 if (env->me_flags & MDB_WRITEMAP) {
2144 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2145 dp = txn->mt_u.dirty_list[i].mptr;
2146 /* clear dirty flag */
2147 dp->mp_flags &= ~P_DIRTY;
2148 txn->mt_u.dirty_list[i].mid = 0;
2150 txn->mt_u.dirty_list[0].mid = 0;
2154 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2160 /* Windows actually supports scatter/gather I/O, but only on
2161 * unbuffered file handles. Since we're relying on the OS page
2162 * cache for all our data, that's self-defeating. So we just
2163 * write pages one at a time. We use the ov structure to set
2164 * the write offset, to at least save the overhead of a Seek
2168 memset(&ov, 0, sizeof(ov));
2169 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2171 dp = txn->mt_u.dirty_list[i].mptr;
2172 DPRINTF("committing page %zu", dp->mp_pgno);
2173 size = dp->mp_pgno * env->me_psize;
2174 ov.Offset = size & 0xffffffff;
2175 ov.OffsetHigh = size >> 16;
2176 ov.OffsetHigh >>= 16;
2177 /* clear dirty flag */
2178 dp->mp_flags &= ~P_DIRTY;
2179 wsize = env->me_psize;
2180 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2181 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2184 DPRINTF("WriteFile: %d", n);
2191 struct iovec iov[MDB_COMMIT_PAGES];
2195 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2196 dp = txn->mt_u.dirty_list[i].mptr;
2197 if (dp->mp_pgno != next) {
2199 rc = writev(env->me_fd, iov, n);
2203 DPUTS("short write, filesystem full?");
2205 DPRINTF("writev: %s", strerror(n));
2212 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2215 DPRINTF("committing page %zu", dp->mp_pgno);
2216 iov[n].iov_len = env->me_psize;
2217 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2218 iov[n].iov_base = (char *)dp;
2219 size += iov[n].iov_len;
2220 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2221 /* clear dirty flag */
2222 dp->mp_flags &= ~P_DIRTY;
2223 if (++n >= MDB_COMMIT_PAGES) {
2233 rc = writev(env->me_fd, iov, n);
2237 DPUTS("short write, filesystem full?");
2239 DPRINTF("writev: %s", strerror(n));
2246 /* Drop the dirty pages.
2248 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2249 dp = txn->mt_u.dirty_list[i].mptr;
2250 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2251 dp->mp_next = txn->mt_env->me_dpages;
2252 VGMEMP_FREE(txn->mt_env, dp);
2253 txn->mt_env->me_dpages = dp;
2255 VGMEMP_FREE(txn->mt_env, dp);
2258 txn->mt_u.dirty_list[i].mid = 0;
2260 txn->mt_u.dirty_list[0].mid = 0;
2263 if ((n = mdb_env_sync(env, 0)) != 0 ||
2264 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2271 if (txn->mt_numdbs > env->me_numdbs) {
2272 /* update the DB flags */
2274 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2275 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2279 UNLOCK_MUTEX_W(env);
2285 /** Read the environment parameters of a DB environment before
2286 * mapping it into memory.
2287 * @param[in] env the environment handle
2288 * @param[out] meta address of where to store the meta information
2289 * @return 0 on success, non-zero on failure.
2292 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2299 /* We don't know the page size yet, so use a minimum value.
2303 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2305 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2310 else if (rc != MDB_PAGESIZE) {
2314 DPRINTF("read: %s", strerror(err));
2318 p = (MDB_page *)&pbuf;
2320 if (!F_ISSET(p->mp_flags, P_META)) {
2321 DPRINTF("page %zu not a meta page", p->mp_pgno);
2326 if (m->mm_magic != MDB_MAGIC) {
2327 DPUTS("meta has invalid magic");
2331 if (m->mm_version != MDB_VERSION) {
2332 DPRINTF("database is version %u, expected version %u",
2333 m->mm_version, MDB_VERSION);
2334 return MDB_VERSION_MISMATCH;
2337 memcpy(meta, m, sizeof(*m));
2341 /** Write the environment parameters of a freshly created DB environment.
2342 * @param[in] env the environment handle
2343 * @param[out] meta address of where to store the meta information
2344 * @return 0 on success, non-zero on failure.
2347 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2354 DPUTS("writing new meta page");
2356 GET_PAGESIZE(psize);
2358 meta->mm_magic = MDB_MAGIC;
2359 meta->mm_version = MDB_VERSION;
2360 meta->mm_psize = psize;
2361 meta->mm_last_pg = 1;
2362 meta->mm_flags = env->me_flags & 0xffff;
2363 meta->mm_flags |= MDB_INTEGERKEY;
2364 meta->mm_dbs[0].md_root = P_INVALID;
2365 meta->mm_dbs[1].md_root = P_INVALID;
2367 p = calloc(2, psize);
2369 p->mp_flags = P_META;
2372 memcpy(m, meta, sizeof(*meta));
2374 q = (MDB_page *)((char *)p + psize);
2377 q->mp_flags = P_META;
2380 memcpy(m, meta, sizeof(*meta));
2385 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2386 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2389 rc = write(env->me_fd, p, psize * 2);
2390 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2396 /** Update the environment info to commit a transaction.
2397 * @param[in] txn the transaction that's being committed
2398 * @return 0 on success, non-zero on failure.
2401 mdb_env_write_meta(MDB_txn *txn)
2404 MDB_meta meta, metab;
2406 int rc, len, toggle;
2412 assert(txn != NULL);
2413 assert(txn->mt_env != NULL);
2415 toggle = !txn->mt_toggle;
2416 DPRINTF("writing meta page %d for root page %zu",
2417 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2421 if (env->me_flags & MDB_WRITEMAP) {
2422 MDB_meta *mp = env->me_metas[toggle];
2423 mp->mm_dbs[0] = txn->mt_dbs[0];
2424 mp->mm_dbs[1] = txn->mt_dbs[1];
2425 mp->mm_last_pg = txn->mt_next_pgno - 1;
2426 mp->mm_txnid = txn->mt_txnid;
2427 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2428 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2431 ptr += env->me_psize;
2432 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2439 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2440 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2442 ptr = (char *)&meta;
2443 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2444 len = sizeof(MDB_meta) - off;
2447 meta.mm_dbs[0] = txn->mt_dbs[0];
2448 meta.mm_dbs[1] = txn->mt_dbs[1];
2449 meta.mm_last_pg = txn->mt_next_pgno - 1;
2450 meta.mm_txnid = txn->mt_txnid;
2453 off += env->me_psize;
2456 /* Write to the SYNC fd */
2459 memset(&ov, 0, sizeof(ov));
2461 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2464 rc = pwrite(env->me_mfd, ptr, len, off);
2469 DPUTS("write failed, disk error?");
2470 /* On a failure, the pagecache still contains the new data.
2471 * Write some old data back, to prevent it from being used.
2472 * Use the non-SYNC fd; we know it will fail anyway.
2474 meta.mm_last_pg = metab.mm_last_pg;
2475 meta.mm_txnid = metab.mm_txnid;
2477 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2479 r2 = pwrite(env->me_fd, ptr, len, off);
2482 env->me_flags |= MDB_FATAL_ERROR;
2486 /* Memory ordering issues are irrelevant; since the entire writer
2487 * is wrapped by wmutex, all of these changes will become visible
2488 * after the wmutex is unlocked. Since the DB is multi-version,
2489 * readers will get consistent data regardless of how fresh or
2490 * how stale their view of these values is.
2492 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2497 /** Check both meta pages to see which one is newer.
2498 * @param[in] env the environment handle
2499 * @return meta toggle (0 or 1).
2502 mdb_env_pick_meta(const MDB_env *env)
2504 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2508 mdb_env_create(MDB_env **env)
2512 e = calloc(1, sizeof(MDB_env));
2516 e->me_free_pgs = mdb_midl_alloc();
2517 if (!e->me_free_pgs) {
2521 e->me_maxreaders = DEFAULT_READERS;
2523 e->me_fd = INVALID_HANDLE_VALUE;
2524 e->me_lfd = INVALID_HANDLE_VALUE;
2525 e->me_mfd = INVALID_HANDLE_VALUE;
2526 #ifdef MDB_USE_POSIX_SEM
2527 e->me_rmutex = SEM_FAILED;
2528 e->me_wmutex = SEM_FAILED;
2530 e->me_pid = getpid();
2531 VGMEMP_CREATE(e,0,0);
2537 mdb_env_set_mapsize(MDB_env *env, size_t size)
2541 env->me_mapsize = size;
2543 env->me_maxpg = env->me_mapsize / env->me_psize;
2548 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2552 env->me_maxdbs = dbs;
2557 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2559 if (env->me_map || readers < 1)
2561 env->me_maxreaders = readers;
2566 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2568 if (!env || !readers)
2570 *readers = env->me_maxreaders;
2574 /** Further setup required for opening an MDB environment
2577 mdb_env_open2(MDB_env *env, unsigned int flags)
2579 int i, newenv = 0, prot;
2583 env->me_flags = flags;
2585 memset(&meta, 0, sizeof(meta));
2587 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2590 DPUTS("new mdbenv");
2594 if (!env->me_mapsize) {
2595 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2601 LONG sizelo, sizehi;
2602 sizelo = env->me_mapsize & 0xffffffff;
2603 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2605 /* Windows won't create mappings for zero length files.
2606 * Just allocate the maxsize right now.
2609 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2610 if (!SetEndOfFile(env->me_fd))
2612 SetFilePointer(env->me_fd, 0, NULL, 0);
2614 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2615 PAGE_READWRITE : PAGE_READONLY,
2616 sizehi, sizelo, NULL);
2619 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2620 FILE_MAP_WRITE : FILE_MAP_READ,
2621 0, 0, env->me_mapsize, meta.mm_address);
2628 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2631 if (flags & MDB_WRITEMAP) {
2633 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2636 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2638 if (env->me_map == MAP_FAILED) {
2645 meta.mm_mapsize = env->me_mapsize;
2646 if (flags & MDB_FIXEDMAP)
2647 meta.mm_address = env->me_map;
2648 i = mdb_env_init_meta(env, &meta);
2649 if (i != MDB_SUCCESS) {
2653 env->me_psize = meta.mm_psize;
2655 env->me_maxpg = env->me_mapsize / env->me_psize;
2657 p = (MDB_page *)env->me_map;
2658 env->me_metas[0] = METADATA(p);
2659 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2663 int toggle = mdb_env_pick_meta(env);
2664 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2666 DPRINTF("opened database version %u, pagesize %u",
2667 env->me_metas[0]->mm_version, env->me_psize);
2668 DPRINTF("using meta page %d", toggle);
2669 DPRINTF("depth: %u", db->md_depth);
2670 DPRINTF("entries: %zu", db->md_entries);
2671 DPRINTF("branch pages: %zu", db->md_branch_pages);
2672 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2673 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2674 DPRINTF("root: %zu", db->md_root);
2682 /** Release a reader thread's slot in the reader lock table.
2683 * This function is called automatically when a thread exits.
2684 * @param[in] ptr This points to the slot in the reader lock table.
2687 mdb_env_reader_dest(void *ptr)
2689 MDB_reader *reader = ptr;
2695 /** Junk for arranging thread-specific callbacks on Windows. This is
2696 * necessarily platform and compiler-specific. Windows supports up
2697 * to 1088 keys. Let's assume nobody opens more than 64 environments
2698 * in a single process, for now. They can override this if needed.
2700 #ifndef MAX_TLS_KEYS
2701 #define MAX_TLS_KEYS 64
2703 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2704 static int mdb_tls_nkeys;
2706 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2710 case DLL_PROCESS_ATTACH: break;
2711 case DLL_THREAD_ATTACH: break;
2712 case DLL_THREAD_DETACH:
2713 for (i=0; i<mdb_tls_nkeys; i++) {
2714 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2715 mdb_env_reader_dest(r);
2718 case DLL_PROCESS_DETACH: break;
2723 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2725 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2729 /* Force some symbol references.
2730 * _tls_used forces the linker to create the TLS directory if not already done
2731 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2733 #pragma comment(linker, "/INCLUDE:_tls_used")
2734 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2735 #pragma const_seg(".CRT$XLB")
2736 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2737 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2740 #pragma comment(linker, "/INCLUDE:__tls_used")
2741 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2742 #pragma data_seg(".CRT$XLB")
2743 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2745 #endif /* WIN 32/64 */
2746 #endif /* !__GNUC__ */
2749 /** Downgrade the exclusive lock on the region back to shared */
2751 mdb_env_share_locks(MDB_env *env, int *excl)
2753 int rc = 0, toggle = mdb_env_pick_meta(env);
2755 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2760 /* First acquire a shared lock. The Unlock will
2761 * then release the existing exclusive lock.
2763 memset(&ov, 0, sizeof(ov));
2764 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2765 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2770 struct flock lock_info;
2771 /* The shared lock replaces the existing lock */
2772 memset((void *)&lock_info, 0, sizeof(lock_info));
2773 lock_info.l_type = F_RDLCK;
2774 lock_info.l_whence = SEEK_SET;
2775 lock_info.l_start = 0;
2776 lock_info.l_len = 1;
2777 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2778 (rc = ErrCode()) == EINTR) ;
2779 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
2786 /** Try to get exlusive lock, otherwise shared.
2787 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
2790 mdb_env_excl_lock(MDB_env *env, int *excl)
2794 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2798 memset(&ov, 0, sizeof(ov));
2799 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2804 struct flock lock_info;
2805 memset((void *)&lock_info, 0, sizeof(lock_info));
2806 lock_info.l_type = F_WRLCK;
2807 lock_info.l_whence = SEEK_SET;
2808 lock_info.l_start = 0;
2809 lock_info.l_len = 1;
2810 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2811 (rc = ErrCode()) == EINTR) ;
2815 # ifdef MDB_USE_POSIX_SEM
2816 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
2819 lock_info.l_type = F_RDLCK;
2820 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
2821 (rc = ErrCode()) == EINTR) ;
2829 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
2831 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2833 * @(#) $Revision: 5.1 $
2834 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2835 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2837 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2841 * Please do not copyright this code. This code is in the public domain.
2843 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2844 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2845 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2846 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2847 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2848 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2849 * PERFORMANCE OF THIS SOFTWARE.
2852 * chongo <Landon Curt Noll> /\oo/\
2853 * http://www.isthe.com/chongo/
2855 * Share and Enjoy! :-)
2858 typedef unsigned long long mdb_hash_t;
2859 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2861 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2862 * @param[in] str string to hash
2863 * @param[in] hval initial value for hash
2864 * @return 64 bit hash
2866 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2867 * hval arg on the first call.
2870 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
2872 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
2873 unsigned char *end = s + val->mv_size;
2875 * FNV-1a hash each octet of the string
2878 /* xor the bottom with the current octet */
2879 hval ^= (mdb_hash_t)*s++;
2881 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2882 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2883 (hval << 7) + (hval << 8) + (hval << 40);
2885 /* return our new hash value */
2889 /** Hash the string and output the hash in hex.
2890 * @param[in] str string to hash
2891 * @param[out] hexbuf an array of 17 chars to hold the hash
2894 mdb_hash_hex(MDB_val *val, char *hexbuf)
2897 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
2898 for (i=0; i<8; i++) {
2899 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2905 /** Open and/or initialize the lock region for the environment.
2906 * @param[in] env The MDB environment.
2907 * @param[in] lpath The pathname of the file used for the lock region.
2908 * @param[in] mode The Unix permissions for the file, if we create it.
2909 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
2910 * @return 0 on success, non-zero on failure.
2913 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2921 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2922 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2923 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2926 /* Try to get exclusive lock. If we succeed, then
2927 * nobody is using the lock region and we should initialize it.
2929 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2930 size = GetFileSize(env->me_lfd, NULL);
2936 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1)
2938 /* Lose record locks when exec*() */
2939 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
2940 fcntl(env->me_lfd, F_SETFD, fdflags);
2942 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
2943 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1)
2947 /* Try to get exclusive lock. If we succeed, then
2948 * nobody is using the lock region and we should initialize it.
2950 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2952 size = lseek(env->me_lfd, 0, SEEK_END);
2954 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2955 if (size < rsize && *excl > 0) {
2957 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2958 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
2960 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
2964 size = rsize - sizeof(MDB_txninfo);
2965 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2970 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2972 if (!mh) goto fail_errno;
2973 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2975 if (!env->me_txns) goto fail_errno;
2977 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2979 if (m == MAP_FAILED) goto fail_errno;
2985 BY_HANDLE_FILE_INFORMATION stbuf;
2994 if (!mdb_sec_inited) {
2995 InitializeSecurityDescriptor(&mdb_null_sd,
2996 SECURITY_DESCRIPTOR_REVISION);
2997 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2998 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2999 mdb_all_sa.bInheritHandle = FALSE;
3000 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3003 GetFileInformationByHandle(env->me_lfd, &stbuf);
3004 idbuf.volume = stbuf.dwVolumeSerialNumber;
3005 idbuf.nhigh = stbuf.nFileIndexHigh;
3006 idbuf.nlow = stbuf.nFileIndexLow;
3007 val.mv_data = &idbuf;
3008 val.mv_size = sizeof(idbuf);
3009 mdb_hash_hex(&val, hexbuf);
3010 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3011 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3012 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3013 if (!env->me_rmutex) goto fail_errno;
3014 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3015 if (!env->me_wmutex) goto fail_errno;
3016 #elif defined(MDB_USE_POSIX_SEM)
3025 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3026 idbuf.dev = stbuf.st_dev;
3027 idbuf.ino = stbuf.st_ino;
3028 val.mv_data = &idbuf;
3029 val.mv_size = sizeof(idbuf);
3030 mdb_hash_hex(&val, hexbuf);
3031 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3032 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3033 /* Clean up after a previous run, if needed: Try to
3034 * remove both semaphores before doing anything else.
3036 sem_unlink(env->me_txns->mti_rmname);
3037 sem_unlink(env->me_txns->mti_wmname);
3038 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3039 O_CREAT|O_EXCL, mode, 1);
3040 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3041 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3042 O_CREAT|O_EXCL, mode, 1);
3043 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3044 #else /* MDB_USE_POSIX_SEM */
3045 pthread_mutexattr_t mattr;
3047 if ((rc = pthread_mutexattr_init(&mattr))
3048 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3049 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3050 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3052 pthread_mutexattr_destroy(&mattr);
3053 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3055 env->me_txns->mti_version = MDB_VERSION;
3056 env->me_txns->mti_magic = MDB_MAGIC;
3057 env->me_txns->mti_txnid = 0;
3058 env->me_txns->mti_numreaders = 0;
3061 if (env->me_txns->mti_magic != MDB_MAGIC) {
3062 DPUTS("lock region has invalid magic");
3066 if (env->me_txns->mti_version != MDB_VERSION) {
3067 DPRINTF("lock region is version %u, expected version %u",
3068 env->me_txns->mti_version, MDB_VERSION);
3069 rc = MDB_VERSION_MISMATCH;
3073 if (rc != EACCES && rc != EAGAIN) {
3077 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3078 if (!env->me_rmutex) goto fail_errno;
3079 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3080 if (!env->me_wmutex) goto fail_errno;
3081 #elif defined(MDB_USE_POSIX_SEM)
3082 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3083 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3084 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3085 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3096 /** The name of the lock file in the DB environment */
3097 #define LOCKNAME "/lock.mdb"
3098 /** The name of the data file in the DB environment */
3099 #define DATANAME "/data.mdb"
3100 /** The suffix of the lock file when no subdir is used */
3101 #define LOCKSUFF "-lock"
3104 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3106 int oflags, rc, len, excl;
3107 char *lpath, *dpath;
3109 if (env->me_fd != INVALID_HANDLE_VALUE)
3113 if (flags & MDB_NOSUBDIR) {
3114 rc = len + sizeof(LOCKSUFF) + len + 1;
3116 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3121 if (flags & MDB_NOSUBDIR) {
3122 dpath = lpath + len + sizeof(LOCKSUFF);
3123 sprintf(lpath, "%s" LOCKSUFF, path);
3124 strcpy(dpath, path);
3126 dpath = lpath + len + sizeof(LOCKNAME);
3127 sprintf(lpath, "%s" LOCKNAME, path);
3128 sprintf(dpath, "%s" DATANAME, path);
3131 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3135 /* silently ignore WRITEMAP if we're only getting read access */
3136 if (F_ISSET(flags, MDB_RDONLY) && F_ISSET(flags, MDB_WRITEMAP))
3137 flags ^= MDB_WRITEMAP;
3140 if (F_ISSET(flags, MDB_RDONLY)) {
3141 oflags = GENERIC_READ;
3142 len = OPEN_EXISTING;
3144 oflags = GENERIC_READ|GENERIC_WRITE;
3147 mode = FILE_ATTRIBUTE_NORMAL;
3148 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3149 NULL, len, mode, NULL);
3151 if (F_ISSET(flags, MDB_RDONLY))
3154 oflags = O_RDWR | O_CREAT;
3156 env->me_fd = open(dpath, oflags, mode);
3158 if (env->me_fd == INVALID_HANDLE_VALUE) {
3163 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
3164 if (flags & (MDB_RDONLY|MDB_NOSYNC|MDB_NOMETASYNC|MDB_WRITEMAP)) {
3165 env->me_mfd = env->me_fd;
3167 /* synchronous fd for meta writes */
3169 env->me_mfd = CreateFile(dpath, oflags,
3170 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3171 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3173 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3175 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3180 DPRINTF("opened dbenv %p", (void *) env);
3181 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3182 env->me_numdbs = 2; /* this notes that me_txkey was set */
3184 /* Windows TLS callbacks need help finding their TLS info. */
3185 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3186 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3193 rc = mdb_env_share_locks(env, &excl);
3197 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3198 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3199 env->me_path = strdup(path);
3200 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3206 mdb_env_close0(env, excl);
3212 /** Destroy resources from mdb_env_open() and clear our readers */
3214 mdb_env_close0(MDB_env *env, int excl)
3218 if (env->me_lfd == INVALID_HANDLE_VALUE) /* 1st field to get inited */
3221 free(env->me_dbflags);
3225 if (env->me_numdbs) {
3226 pthread_key_delete(env->me_txkey);
3228 /* Delete our key from the global list */
3229 for (i=0; i<mdb_tls_nkeys; i++)
3230 if (mdb_tls_keys[i] == env->me_txkey) {
3231 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3239 munmap(env->me_map, env->me_mapsize);
3241 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3243 if (env->me_fd != INVALID_HANDLE_VALUE)
3246 pid_t pid = env->me_pid;
3247 /* Clearing readers is done in this function because
3248 * me_txkey with its destructor must be disabled first.
3250 for (i = env->me_numreaders; --i >= 0; )
3251 if (env->me_txns->mti_readers[i].mr_pid == pid)
3252 env->me_txns->mti_readers[i].mr_pid = 0;
3254 if (env->me_rmutex) {
3255 CloseHandle(env->me_rmutex);
3256 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3258 /* Windows automatically destroys the mutexes when
3259 * the last handle closes.
3261 #elif defined(MDB_USE_POSIX_SEM)
3262 if (env->me_rmutex != SEM_FAILED) {
3263 sem_close(env->me_rmutex);
3264 if (env->me_wmutex != SEM_FAILED)
3265 sem_close(env->me_wmutex);
3266 /* If we have the filelock: If we are the
3267 * only remaining user, clean up semaphores.
3270 mdb_env_excl_lock(env, &excl);
3272 sem_unlink(env->me_txns->mti_rmname);
3273 sem_unlink(env->me_txns->mti_wmname);
3277 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3281 env->me_lfd = INVALID_HANDLE_VALUE; /* Mark env as reset */
3285 mdb_env_copy(MDB_env *env, const char *path)
3287 MDB_txn *txn = NULL;
3288 int rc, len, oflags;
3291 HANDLE newfd = INVALID_HANDLE_VALUE;
3293 if (env->me_flags & MDB_NOSUBDIR) {
3297 len += sizeof(DATANAME);
3298 lpath = malloc(len);
3301 sprintf(lpath, "%s" DATANAME, path);
3304 /* The destination path must exist, but the destination file must not.
3305 * We don't want the OS to cache the writes, since the source data is
3306 * already in the OS cache.
3309 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3310 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3312 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3318 if (!(env->me_flags & MDB_NOSUBDIR))
3320 if (newfd == INVALID_HANDLE_VALUE) {
3325 #ifdef F_NOCACHE /* __APPLE__ */
3326 rc = fcntl(newfd, F_NOCACHE, 1);
3333 /* Temporarily block writers until we snapshot the meta pages */
3336 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3338 UNLOCK_MUTEX_W(env);
3342 wsize = env->me_psize * 2;
3346 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3347 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3350 rc = write(newfd, env->me_map, wsize);
3351 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3353 UNLOCK_MUTEX_W(env);
3358 ptr = env->me_map + wsize;
3359 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3363 rc = WriteFile(newfd, ptr, wsize, &len, NULL);
3364 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3367 rc = write(newfd, ptr, wsize);
3368 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3373 if (newfd != INVALID_HANDLE_VALUE)
3380 mdb_env_close(MDB_env *env)
3387 VGMEMP_DESTROY(env);
3388 while ((dp = env->me_dpages) != NULL) {
3389 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3390 env->me_dpages = dp->mp_next;
3394 mdb_env_close0(env, 0);
3395 mdb_midl_free(env->me_free_pgs);
3399 /** Compare two items pointing at aligned size_t's */
3401 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3403 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3404 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3407 /** Compare two items pointing at aligned int's */
3409 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3411 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3412 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3415 /** Compare two items pointing at ints of unknown alignment.
3416 * Nodes and keys are guaranteed to be 2-byte aligned.
3419 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3421 #if BYTE_ORDER == LITTLE_ENDIAN
3422 unsigned short *u, *c;
3425 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3426 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3429 } while(!x && u > (unsigned short *)a->mv_data);
3432 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3436 /** Compare two items lexically */
3438 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3445 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3451 diff = memcmp(a->mv_data, b->mv_data, len);
3452 return diff ? diff : len_diff<0 ? -1 : len_diff;
3455 /** Compare two items in reverse byte order */
3457 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3459 const unsigned char *p1, *p2, *p1_lim;
3463 p1_lim = (const unsigned char *)a->mv_data;
3464 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3465 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3467 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3473 while (p1 > p1_lim) {
3474 diff = *--p1 - *--p2;
3478 return len_diff<0 ? -1 : len_diff;
3481 /** Search for key within a page, using binary search.
3482 * Returns the smallest entry larger or equal to the key.
3483 * If exactp is non-null, stores whether the found entry was an exact match
3484 * in *exactp (1 or 0).
3485 * Updates the cursor index with the index of the found entry.
3486 * If no entry larger or equal to the key is found, returns NULL.
3489 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3491 unsigned int i = 0, nkeys;
3494 MDB_page *mp = mc->mc_pg[mc->mc_top];
3495 MDB_node *node = NULL;
3500 nkeys = NUMKEYS(mp);
3505 COPY_PGNO(pgno, mp->mp_pgno);
3506 DPRINTF("searching %u keys in %s %spage %zu",
3507 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3514 low = IS_LEAF(mp) ? 0 : 1;
3516 cmp = mc->mc_dbx->md_cmp;
3518 /* Branch pages have no data, so if using integer keys,
3519 * alignment is guaranteed. Use faster mdb_cmp_int.
3521 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3522 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3529 nodekey.mv_size = mc->mc_db->md_pad;
3530 node = NODEPTR(mp, 0); /* fake */
3531 while (low <= high) {
3532 i = (low + high) >> 1;
3533 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3534 rc = cmp(key, &nodekey);
3535 DPRINTF("found leaf index %u [%s], rc = %i",
3536 i, DKEY(&nodekey), rc);
3545 while (low <= high) {
3546 i = (low + high) >> 1;
3548 node = NODEPTR(mp, i);
3549 nodekey.mv_size = NODEKSZ(node);
3550 nodekey.mv_data = NODEKEY(node);
3552 rc = cmp(key, &nodekey);
3555 DPRINTF("found leaf index %u [%s], rc = %i",
3556 i, DKEY(&nodekey), rc);
3558 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3559 i, DKEY(&nodekey), NODEPGNO(node), rc);
3570 if (rc > 0) { /* Found entry is less than the key. */
3571 i++; /* Skip to get the smallest entry larger than key. */
3573 node = NODEPTR(mp, i);
3576 *exactp = (rc == 0);
3577 /* store the key index */
3578 mc->mc_ki[mc->mc_top] = i;
3580 /* There is no entry larger or equal to the key. */
3583 /* nodeptr is fake for LEAF2 */
3589 mdb_cursor_adjust(MDB_cursor *mc, func)
3593 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3594 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3601 /** Pop a page off the top of the cursor's stack. */
3603 mdb_cursor_pop(MDB_cursor *mc)
3607 MDB_page *top = mc->mc_pg[mc->mc_top];
3613 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3614 mc->mc_dbi, (void *) mc);
3618 /** Push a page onto the top of the cursor's stack. */
3620 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3622 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3623 mc->mc_dbi, (void *) mc);
3625 if (mc->mc_snum >= CURSOR_STACK) {
3626 assert(mc->mc_snum < CURSOR_STACK);
3627 return MDB_CURSOR_FULL;
3630 mc->mc_top = mc->mc_snum++;
3631 mc->mc_pg[mc->mc_top] = mp;
3632 mc->mc_ki[mc->mc_top] = 0;
3637 /** Find the address of the page corresponding to a given page number.
3638 * @param[in] txn the transaction for this access.
3639 * @param[in] pgno the page number for the page to retrieve.
3640 * @param[out] ret address of a pointer where the page's address will be stored.
3641 * @return 0 on success, non-zero on failure.
3644 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3648 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
3649 if (pgno < txn->mt_next_pgno)
3650 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3653 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3655 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3656 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3657 p = txn->mt_u.dirty_list[x].mptr;
3661 if (pgno < txn->mt_next_pgno)
3662 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3667 DPRINTF("page %zu not found", pgno);
3670 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3673 /** Search for the page a given key should be in.
3674 * Pushes parent pages on the cursor stack. This function continues a
3675 * search on a cursor that has already been initialized. (Usually by
3676 * #mdb_page_search() but also by #mdb_node_move().)
3677 * @param[in,out] mc the cursor for this operation.
3678 * @param[in] key the key to search for. If NULL, search for the lowest
3679 * page. (This is used by #mdb_cursor_first().)
3680 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3681 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3682 * @return 0 on success, non-zero on failure.
3685 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3687 MDB_page *mp = mc->mc_pg[mc->mc_top];
3692 while (IS_BRANCH(mp)) {
3696 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3697 assert(NUMKEYS(mp) > 1);
3698 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3700 if (key == NULL) /* Initialize cursor to first page. */
3702 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3703 /* cursor to last page */
3707 node = mdb_node_search(mc, key, &exact);
3709 i = NUMKEYS(mp) - 1;
3711 i = mc->mc_ki[mc->mc_top];
3720 DPRINTF("following index %u for key [%s]",
3722 assert(i < NUMKEYS(mp));
3723 node = NODEPTR(mp, i);
3725 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3728 mc->mc_ki[mc->mc_top] = i;
3729 if ((rc = mdb_cursor_push(mc, mp)))
3733 if ((rc = mdb_page_touch(mc)) != 0)
3735 mp = mc->mc_pg[mc->mc_top];
3740 DPRINTF("internal error, index points to a %02X page!?",
3742 return MDB_CORRUPTED;
3745 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3746 key ? DKEY(key) : NULL);
3751 /** Search for the page a given key should be in.
3752 * Pushes parent pages on the cursor stack. This function just sets up
3753 * the search; it finds the root page for \b mc's database and sets this
3754 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3755 * called to complete the search.
3756 * @param[in,out] mc the cursor for this operation.
3757 * @param[in] key the key to search for. If NULL, search for the lowest
3758 * page. (This is used by #mdb_cursor_first().)
3759 * @param[in] modify If true, visited pages are updated with new page numbers.
3760 * @return 0 on success, non-zero on failure.
3763 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
3768 /* Make sure the txn is still viable, then find the root from
3769 * the txn's db table.
3771 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3772 DPUTS("transaction has failed, must abort");
3775 /* Make sure we're using an up-to-date root */
3776 if (mc->mc_dbi > MAIN_DBI) {
3777 if ((*mc->mc_dbflag & DB_STALE) ||
3778 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
3780 unsigned char dbflag = 0;
3781 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3782 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
3785 if (*mc->mc_dbflag & DB_STALE) {
3788 MDB_node *leaf = mdb_node_search(&mc2,
3789 &mc->mc_dbx->md_name, &exact);
3791 return MDB_NOTFOUND;
3792 mdb_node_read(mc->mc_txn, leaf, &data);
3793 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3795 if (flags & MDB_PS_MODIFY)
3797 *mc->mc_dbflag = dbflag;
3800 root = mc->mc_db->md_root;
3802 if (root == P_INVALID) { /* Tree is empty. */
3803 DPUTS("tree is empty");
3804 return MDB_NOTFOUND;
3809 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
3810 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3816 DPRINTF("db %u root page %zu has flags 0x%X",
3817 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3819 if (flags & MDB_PS_MODIFY) {
3820 if ((rc = mdb_page_touch(mc)))
3824 if (flags & MDB_PS_ROOTONLY)
3827 return mdb_page_search_root(mc, key, flags);
3830 /** Return the data associated with a given node.
3831 * @param[in] txn The transaction for this operation.
3832 * @param[in] leaf The node being read.
3833 * @param[out] data Updated to point to the node's data.
3834 * @return 0 on success, non-zero on failure.
3837 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3839 MDB_page *omp; /* overflow page */
3843 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3844 data->mv_size = NODEDSZ(leaf);
3845 data->mv_data = NODEDATA(leaf);
3849 /* Read overflow data.
3851 data->mv_size = NODEDSZ(leaf);
3852 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3853 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3854 DPRINTF("read overflow page %zu failed", pgno);
3857 data->mv_data = METADATA(omp);
3863 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3864 MDB_val *key, MDB_val *data)
3873 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3875 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3878 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3882 mdb_cursor_init(&mc, txn, dbi, &mx);
3883 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3886 /** Find a sibling for a page.
3887 * Replaces the page at the top of the cursor's stack with the
3888 * specified sibling, if one exists.
3889 * @param[in] mc The cursor for this operation.
3890 * @param[in] move_right Non-zero if the right sibling is requested,
3891 * otherwise the left sibling.
3892 * @return 0 on success, non-zero on failure.
3895 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3901 if (mc->mc_snum < 2) {
3902 return MDB_NOTFOUND; /* root has no siblings */
3906 DPRINTF("parent page is page %zu, index %u",
3907 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3909 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3910 : (mc->mc_ki[mc->mc_top] == 0)) {
3911 DPRINTF("no more keys left, moving to %s sibling",
3912 move_right ? "right" : "left");
3913 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3917 mc->mc_ki[mc->mc_top]++;
3919 mc->mc_ki[mc->mc_top]--;
3920 DPRINTF("just moving to %s index key %u",
3921 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3923 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3925 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3926 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3929 mdb_cursor_push(mc, mp);
3934 /** Move the cursor to the next data item. */
3936 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3942 if (mc->mc_flags & C_EOF) {
3943 return MDB_NOTFOUND;
3946 assert(mc->mc_flags & C_INITIALIZED);
3948 mp = mc->mc_pg[mc->mc_top];
3950 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3951 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3952 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3953 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3954 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3955 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3959 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3960 if (op == MDB_NEXT_DUP)
3961 return MDB_NOTFOUND;
3965 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3967 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3968 DPUTS("=====> move to next sibling page");
3969 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3970 mc->mc_flags |= C_EOF;
3971 mc->mc_flags &= ~C_INITIALIZED;
3972 return MDB_NOTFOUND;
3974 mp = mc->mc_pg[mc->mc_top];
3975 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3977 mc->mc_ki[mc->mc_top]++;
3979 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3980 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3983 key->mv_size = mc->mc_db->md_pad;
3984 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3988 assert(IS_LEAF(mp));
3989 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3991 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3992 mdb_xcursor_init1(mc, leaf);
3995 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3998 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3999 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4000 if (rc != MDB_SUCCESS)
4005 MDB_GET_KEY(leaf, key);
4009 /** Move the cursor to the previous data item. */
4011 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4017 assert(mc->mc_flags & C_INITIALIZED);
4019 mp = mc->mc_pg[mc->mc_top];
4021 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4022 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4023 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4024 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4025 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4026 if (op != MDB_PREV || rc == MDB_SUCCESS)
4029 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4030 if (op == MDB_PREV_DUP)
4031 return MDB_NOTFOUND;
4036 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4038 if (mc->mc_ki[mc->mc_top] == 0) {
4039 DPUTS("=====> move to prev sibling page");
4040 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4041 mc->mc_flags &= ~C_INITIALIZED;
4042 return MDB_NOTFOUND;
4044 mp = mc->mc_pg[mc->mc_top];
4045 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4046 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4048 mc->mc_ki[mc->mc_top]--;
4050 mc->mc_flags &= ~C_EOF;
4052 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4053 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4056 key->mv_size = mc->mc_db->md_pad;
4057 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4061 assert(IS_LEAF(mp));
4062 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4064 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4065 mdb_xcursor_init1(mc, leaf);
4068 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4071 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4072 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4073 if (rc != MDB_SUCCESS)
4078 MDB_GET_KEY(leaf, key);
4082 /** Set the cursor on a specific data item. */
4084 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4085 MDB_cursor_op op, int *exactp)
4089 MDB_node *leaf = NULL;
4094 assert(key->mv_size > 0);
4096 /* See if we're already on the right page */
4097 if (mc->mc_flags & C_INITIALIZED) {
4100 mp = mc->mc_pg[mc->mc_top];
4102 mc->mc_ki[mc->mc_top] = 0;
4103 return MDB_NOTFOUND;
4105 if (mp->mp_flags & P_LEAF2) {
4106 nodekey.mv_size = mc->mc_db->md_pad;
4107 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4109 leaf = NODEPTR(mp, 0);
4110 MDB_GET_KEY(leaf, &nodekey);
4112 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4114 /* Probably happens rarely, but first node on the page
4115 * was the one we wanted.
4117 mc->mc_ki[mc->mc_top] = 0;
4124 unsigned int nkeys = NUMKEYS(mp);
4126 if (mp->mp_flags & P_LEAF2) {
4127 nodekey.mv_data = LEAF2KEY(mp,
4128 nkeys-1, nodekey.mv_size);
4130 leaf = NODEPTR(mp, nkeys-1);
4131 MDB_GET_KEY(leaf, &nodekey);
4133 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4135 /* last node was the one we wanted */
4136 mc->mc_ki[mc->mc_top] = nkeys-1;
4142 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4143 /* This is definitely the right page, skip search_page */
4144 if (mp->mp_flags & P_LEAF2) {
4145 nodekey.mv_data = LEAF2KEY(mp,
4146 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4148 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4149 MDB_GET_KEY(leaf, &nodekey);
4151 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4153 /* current node was the one we wanted */
4163 /* If any parents have right-sibs, search.
4164 * Otherwise, there's nothing further.
4166 for (i=0; i<mc->mc_top; i++)
4168 NUMKEYS(mc->mc_pg[i])-1)
4170 if (i == mc->mc_top) {
4171 /* There are no other pages */
4172 mc->mc_ki[mc->mc_top] = nkeys;
4173 return MDB_NOTFOUND;
4177 /* There are no other pages */
4178 mc->mc_ki[mc->mc_top] = 0;
4179 return MDB_NOTFOUND;
4183 rc = mdb_page_search(mc, key, 0);
4184 if (rc != MDB_SUCCESS)
4187 mp = mc->mc_pg[mc->mc_top];
4188 assert(IS_LEAF(mp));
4191 leaf = mdb_node_search(mc, key, exactp);
4192 if (exactp != NULL && !*exactp) {
4193 /* MDB_SET specified and not an exact match. */
4194 return MDB_NOTFOUND;
4198 DPUTS("===> inexact leaf not found, goto sibling");
4199 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4200 return rc; /* no entries matched */
4201 mp = mc->mc_pg[mc->mc_top];
4202 assert(IS_LEAF(mp));
4203 leaf = NODEPTR(mp, 0);
4207 mc->mc_flags |= C_INITIALIZED;
4208 mc->mc_flags &= ~C_EOF;
4211 key->mv_size = mc->mc_db->md_pad;
4212 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4216 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4217 mdb_xcursor_init1(mc, leaf);
4220 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4221 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4222 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4225 if (op == MDB_GET_BOTH) {
4231 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4232 if (rc != MDB_SUCCESS)
4235 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4237 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4239 rc = mc->mc_dbx->md_dcmp(data, &d2);
4241 if (op == MDB_GET_BOTH || rc > 0)
4242 return MDB_NOTFOUND;
4247 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4248 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4253 /* The key already matches in all other cases */
4254 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4255 MDB_GET_KEY(leaf, key);
4256 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4261 /** Move the cursor to the first item in the database. */
4263 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4268 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4269 rc = mdb_page_search(mc, NULL, 0);
4270 if (rc != MDB_SUCCESS)
4273 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4275 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4276 mc->mc_flags |= C_INITIALIZED;
4277 mc->mc_flags &= ~C_EOF;
4279 mc->mc_ki[mc->mc_top] = 0;
4281 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4282 key->mv_size = mc->mc_db->md_pad;
4283 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4288 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4289 mdb_xcursor_init1(mc, leaf);
4290 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4295 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4296 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4300 MDB_GET_KEY(leaf, key);
4304 /** Move the cursor to the last item in the database. */
4306 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4311 if (!(mc->mc_flags & C_EOF)) {
4313 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4316 lkey.mv_size = MAXKEYSIZE+1;
4317 lkey.mv_data = NULL;
4318 rc = mdb_page_search(mc, &lkey, 0);
4319 if (rc != MDB_SUCCESS)
4322 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4324 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4325 mc->mc_flags |= C_INITIALIZED|C_EOF;
4327 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4329 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4330 key->mv_size = mc->mc_db->md_pad;
4331 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4336 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4337 mdb_xcursor_init1(mc, leaf);
4338 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4343 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4344 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4349 MDB_GET_KEY(leaf, key);
4354 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4363 case MDB_GET_CURRENT:
4364 if (!mc->mc_flags & C_INITIALIZED) {
4367 MDB_page *mp = mc->mc_pg[mc->mc_top];
4369 mc->mc_ki[mc->mc_top] = 0;
4375 key->mv_size = mc->mc_db->md_pad;
4376 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4378 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4379 MDB_GET_KEY(leaf, key);
4381 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4382 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4384 rc = mdb_node_read(mc->mc_txn, leaf, data);
4391 case MDB_GET_BOTH_RANGE:
4392 if (data == NULL || mc->mc_xcursor == NULL) {
4400 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4402 } else if (op == MDB_SET_RANGE)
4403 rc = mdb_cursor_set(mc, key, data, op, NULL);
4405 rc = mdb_cursor_set(mc, key, data, op, &exact);
4407 case MDB_GET_MULTIPLE:
4409 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4410 !(mc->mc_flags & C_INITIALIZED)) {
4415 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4416 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4419 case MDB_NEXT_MULTIPLE:
4421 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4425 if (!(mc->mc_flags & C_INITIALIZED))
4426 rc = mdb_cursor_first(mc, key, data);
4428 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4429 if (rc == MDB_SUCCESS) {
4430 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4433 mx = &mc->mc_xcursor->mx_cursor;
4434 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4436 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4437 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4445 case MDB_NEXT_NODUP:
4446 if (!(mc->mc_flags & C_INITIALIZED))
4447 rc = mdb_cursor_first(mc, key, data);
4449 rc = mdb_cursor_next(mc, key, data, op);
4453 case MDB_PREV_NODUP:
4454 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4455 rc = mdb_cursor_last(mc, key, data);
4456 mc->mc_flags &= ~C_EOF;
4458 rc = mdb_cursor_prev(mc, key, data, op);
4461 rc = mdb_cursor_first(mc, key, data);
4465 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4466 !(mc->mc_flags & C_INITIALIZED) ||
4467 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4471 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4474 rc = mdb_cursor_last(mc, key, data);
4478 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4479 !(mc->mc_flags & C_INITIALIZED) ||
4480 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4484 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4487 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4495 /** Touch all the pages in the cursor stack.
4496 * Makes sure all the pages are writable, before attempting a write operation.
4497 * @param[in] mc The cursor to operate on.
4500 mdb_cursor_touch(MDB_cursor *mc)
4504 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4506 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4507 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4510 *mc->mc_dbflag = DB_DIRTY;
4512 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4513 rc = mdb_page_touch(mc);
4517 mc->mc_top = mc->mc_snum-1;
4522 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4525 MDB_node *leaf = NULL;
4526 MDB_val xdata, *rdata, dkey;
4529 int do_sub = 0, insert = 0;
4530 unsigned int mcount = 0;
4534 char dbuf[MAXKEYSIZE+1];
4535 unsigned int nflags;
4538 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4541 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4542 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4546 if (flags == MDB_CURRENT) {
4547 if (!(mc->mc_flags & C_INITIALIZED))
4550 } else if (mc->mc_db->md_root == P_INVALID) {
4552 /* new database, write a root leaf page */
4553 DPUTS("allocating new root leaf page");
4554 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4558 mdb_cursor_push(mc, np);
4559 mc->mc_db->md_root = np->mp_pgno;
4560 mc->mc_db->md_depth++;
4561 *mc->mc_dbflag = DB_DIRTY;
4562 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4564 np->mp_flags |= P_LEAF2;
4565 mc->mc_flags |= C_INITIALIZED;
4571 if (flags & MDB_APPEND) {
4573 rc = mdb_cursor_last(mc, &k2, &d2);
4575 rc = mc->mc_dbx->md_cmp(key, &k2);
4578 mc->mc_ki[mc->mc_top]++;
4584 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4586 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4587 DPRINTF("duplicate key [%s]", DKEY(key));
4589 return MDB_KEYEXIST;
4591 if (rc && rc != MDB_NOTFOUND)
4595 /* Cursor is positioned, now make sure all pages are writable */
4596 rc2 = mdb_cursor_touch(mc);
4601 /* The key already exists */
4602 if (rc == MDB_SUCCESS) {
4603 /* there's only a key anyway, so this is a no-op */
4604 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4605 unsigned int ksize = mc->mc_db->md_pad;
4606 if (key->mv_size != ksize)
4608 if (flags == MDB_CURRENT) {
4609 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4610 memcpy(ptr, key->mv_data, ksize);
4615 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4618 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4619 /* Was a single item before, must convert now */
4621 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4622 /* Just overwrite the current item */
4623 if (flags == MDB_CURRENT)
4626 dkey.mv_size = NODEDSZ(leaf);
4627 dkey.mv_data = NODEDATA(leaf);
4628 #if UINT_MAX < SIZE_MAX
4629 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4630 #ifdef MISALIGNED_OK
4631 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4633 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4636 /* if data matches, ignore it */
4637 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4638 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4640 /* create a fake page for the dup items */
4641 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4642 dkey.mv_data = dbuf;
4643 fp = (MDB_page *)&pbuf;
4644 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4645 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4646 fp->mp_lower = PAGEHDRSZ;
4647 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4648 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4649 fp->mp_flags |= P_LEAF2;
4650 fp->mp_pad = data->mv_size;
4651 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4653 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4654 (dkey.mv_size & 1) + (data->mv_size & 1);
4656 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4659 xdata.mv_size = fp->mp_upper;
4664 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4665 /* See if we need to convert from fake page to subDB */
4667 unsigned int offset;
4670 fp = NODEDATA(leaf);
4671 if (flags == MDB_CURRENT) {
4673 fp->mp_flags |= P_DIRTY;
4674 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4675 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4679 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4680 offset = fp->mp_pad;
4681 if (SIZELEFT(fp) >= offset)
4683 offset *= 4; /* space for 4 more */
4685 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4687 offset += offset & 1;
4688 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4689 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4691 /* yes, convert it */
4693 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4694 dummy.md_pad = fp->mp_pad;
4695 dummy.md_flags = MDB_DUPFIXED;
4696 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4697 dummy.md_flags |= MDB_INTEGERKEY;
4700 dummy.md_branch_pages = 0;
4701 dummy.md_leaf_pages = 1;
4702 dummy.md_overflow_pages = 0;
4703 dummy.md_entries = NUMKEYS(fp);
4705 xdata.mv_size = sizeof(MDB_db);
4706 xdata.mv_data = &dummy;
4707 if ((rc = mdb_page_alloc(mc, 1, &mp)))
4709 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4710 flags |= F_DUPDATA|F_SUBDATA;
4711 dummy.md_root = mp->mp_pgno;
4713 /* no, just grow it */
4715 xdata.mv_size = NODEDSZ(leaf) + offset;
4716 xdata.mv_data = &pbuf;
4717 mp = (MDB_page *)&pbuf;
4718 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4721 mp->mp_flags = fp->mp_flags | P_DIRTY;
4722 mp->mp_pad = fp->mp_pad;
4723 mp->mp_lower = fp->mp_lower;
4724 mp->mp_upper = fp->mp_upper + offset;
4726 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4728 nsize = NODEDSZ(leaf) - fp->mp_upper;
4729 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4730 for (i=0; i<NUMKEYS(fp); i++)
4731 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4733 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4737 /* data is on sub-DB, just store it */
4738 flags |= F_DUPDATA|F_SUBDATA;
4742 /* overflow page overwrites need special handling */
4743 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4746 int ovpages, dpages;
4748 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4749 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4750 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4751 mdb_page_get(mc->mc_txn, pg, &omp);
4752 /* Is the ov page writable and large enough? */
4753 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4754 /* yes, overwrite it. Note in this case we don't
4755 * bother to try shrinking the node if the new data
4756 * is smaller than the overflow threshold.
4758 if (F_ISSET(flags, MDB_RESERVE))
4759 data->mv_data = METADATA(omp);
4761 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4764 /* no, free ovpages */
4766 mc->mc_db->md_overflow_pages -= ovpages;
4767 for (i=0; i<ovpages; i++) {
4768 DPRINTF("freed ov page %zu", pg);
4769 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4773 } else if (NODEDSZ(leaf) == data->mv_size) {
4774 /* same size, just replace it. Note that we could
4775 * also reuse this node if the new data is smaller,
4776 * but instead we opt to shrink the node in that case.
4778 if (F_ISSET(flags, MDB_RESERVE))
4779 data->mv_data = NODEDATA(leaf);
4781 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4784 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4785 mc->mc_db->md_entries--;
4787 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4794 nflags = flags & NODE_ADD_FLAGS;
4795 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4796 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4797 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4798 nflags &= ~MDB_APPEND;
4800 nflags |= MDB_SPLIT_REPLACE;
4801 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4803 /* There is room already in this leaf page. */
4804 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4805 if (rc == 0 && !do_sub && insert) {
4806 /* Adjust other cursors pointing to mp */
4807 MDB_cursor *m2, *m3;
4808 MDB_dbi dbi = mc->mc_dbi;
4809 unsigned i = mc->mc_top;
4810 MDB_page *mp = mc->mc_pg[i];
4812 if (mc->mc_flags & C_SUB)
4815 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4816 if (mc->mc_flags & C_SUB)
4817 m3 = &m2->mc_xcursor->mx_cursor;
4820 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4821 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4828 if (rc != MDB_SUCCESS)
4829 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4831 /* Now store the actual data in the child DB. Note that we're
4832 * storing the user data in the keys field, so there are strict
4833 * size limits on dupdata. The actual data fields of the child
4834 * DB are all zero size.
4841 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4842 if (flags & MDB_CURRENT) {
4843 xflags = MDB_CURRENT;
4845 mdb_xcursor_init1(mc, leaf);
4846 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4848 /* converted, write the original data first */
4850 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4854 /* Adjust other cursors pointing to mp */
4856 unsigned i = mc->mc_top;
4857 MDB_page *mp = mc->mc_pg[i];
4859 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4860 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4861 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4862 mdb_xcursor_init1(m2, leaf);
4867 if (flags & MDB_APPENDDUP)
4868 xflags |= MDB_APPEND;
4869 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4870 if (flags & F_SUBDATA) {
4871 void *db = NODEDATA(leaf);
4872 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4875 /* sub-writes might have failed so check rc again.
4876 * Don't increment count if we just replaced an existing item.
4878 if (!rc && !(flags & MDB_CURRENT))
4879 mc->mc_db->md_entries++;
4880 if (flags & MDB_MULTIPLE) {
4882 if (mcount < data[1].mv_size) {
4883 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4884 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4894 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4899 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4902 if (!mc->mc_flags & C_INITIALIZED)
4905 rc = mdb_cursor_touch(mc);
4909 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4911 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4912 if (flags != MDB_NODUPDATA) {
4913 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4914 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4916 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4917 /* If sub-DB still has entries, we're done */
4918 if (mc->mc_xcursor->mx_db.md_entries) {
4919 if (leaf->mn_flags & F_SUBDATA) {
4920 /* update subDB info */
4921 void *db = NODEDATA(leaf);
4922 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4924 /* shrink fake page */
4925 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4927 mc->mc_db->md_entries--;
4930 /* otherwise fall thru and delete the sub-DB */
4933 if (leaf->mn_flags & F_SUBDATA) {
4934 /* add all the child DB's pages to the free list */
4935 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4936 if (rc == MDB_SUCCESS) {
4937 mc->mc_db->md_entries -=
4938 mc->mc_xcursor->mx_db.md_entries;
4943 return mdb_cursor_del0(mc, leaf);
4946 /** Allocate and initialize new pages for a database.
4947 * @param[in] mc a cursor on the database being added to.
4948 * @param[in] flags flags defining what type of page is being allocated.
4949 * @param[in] num the number of pages to allocate. This is usually 1,
4950 * unless allocating overflow pages for a large record.
4951 * @param[out] mp Address of a page, or NULL on failure.
4952 * @return 0 on success, non-zero on failure.
4955 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
4960 if ((rc = mdb_page_alloc(mc, num, &np)))
4962 DPRINTF("allocated new mpage %zu, page size %u",
4963 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4964 np->mp_flags = flags | P_DIRTY;
4965 np->mp_lower = PAGEHDRSZ;
4966 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4969 mc->mc_db->md_branch_pages++;
4970 else if (IS_LEAF(np))
4971 mc->mc_db->md_leaf_pages++;
4972 else if (IS_OVERFLOW(np)) {
4973 mc->mc_db->md_overflow_pages += num;
4981 /** Calculate the size of a leaf node.
4982 * The size depends on the environment's page size; if a data item
4983 * is too large it will be put onto an overflow page and the node
4984 * size will only include the key and not the data. Sizes are always
4985 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4986 * of the #MDB_node headers.
4987 * @param[in] env The environment handle.
4988 * @param[in] key The key for the node.
4989 * @param[in] data The data for the node.
4990 * @return The number of bytes needed to store the node.
4993 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4997 sz = LEAFSIZE(key, data);
4998 if (sz >= env->me_psize / MDB_MINKEYS) {
4999 /* put on overflow page */
5000 sz -= data->mv_size - sizeof(pgno_t);
5004 return sz + sizeof(indx_t);
5007 /** Calculate the size of a branch node.
5008 * The size should depend on the environment's page size but since
5009 * we currently don't support spilling large keys onto overflow
5010 * pages, it's simply the size of the #MDB_node header plus the
5011 * size of the key. Sizes are always rounded up to an even number
5012 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5013 * @param[in] env The environment handle.
5014 * @param[in] key The key for the node.
5015 * @return The number of bytes needed to store the node.
5018 mdb_branch_size(MDB_env *env, MDB_val *key)
5023 if (sz >= env->me_psize / MDB_MINKEYS) {
5024 /* put on overflow page */
5025 /* not implemented */
5026 /* sz -= key->size - sizeof(pgno_t); */
5029 return sz + sizeof(indx_t);
5032 /** Add a node to the page pointed to by the cursor.
5033 * @param[in] mc The cursor for this operation.
5034 * @param[in] indx The index on the page where the new node should be added.
5035 * @param[in] key The key for the new node.
5036 * @param[in] data The data for the new node, if any.
5037 * @param[in] pgno The page number, if adding a branch node.
5038 * @param[in] flags Flags for the node.
5039 * @return 0 on success, non-zero on failure. Possible errors are:
5041 * <li>ENOMEM - failed to allocate overflow pages for the node.
5042 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5043 * should never happen since all callers already calculate the
5044 * page's free space before calling this function.
5048 mdb_node_add(MDB_cursor *mc, indx_t indx,
5049 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5052 size_t node_size = NODESIZE;
5055 MDB_page *mp = mc->mc_pg[mc->mc_top];
5056 MDB_page *ofp = NULL; /* overflow page */
5059 assert(mp->mp_upper >= mp->mp_lower);
5061 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5062 IS_LEAF(mp) ? "leaf" : "branch",
5063 IS_SUBP(mp) ? "sub-" : "",
5064 mp->mp_pgno, indx, data ? data->mv_size : 0,
5065 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5068 /* Move higher keys up one slot. */
5069 int ksize = mc->mc_db->md_pad, dif;
5070 char *ptr = LEAF2KEY(mp, indx, ksize);
5071 dif = NUMKEYS(mp) - indx;
5073 memmove(ptr+ksize, ptr, dif*ksize);
5074 /* insert new key */
5075 memcpy(ptr, key->mv_data, ksize);
5077 /* Just using these for counting */
5078 mp->mp_lower += sizeof(indx_t);
5079 mp->mp_upper -= ksize - sizeof(indx_t);
5084 node_size += key->mv_size;
5088 if (F_ISSET(flags, F_BIGDATA)) {
5089 /* Data already on overflow page. */
5090 node_size += sizeof(pgno_t);
5091 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5092 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5094 /* Put data on overflow page. */
5095 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5096 data->mv_size, node_size+data->mv_size);
5097 node_size += sizeof(pgno_t);
5098 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5100 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5103 node_size += data->mv_size;
5106 node_size += node_size & 1;
5108 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5109 DPRINTF("not enough room in page %zu, got %u ptrs",
5110 mp->mp_pgno, NUMKEYS(mp));
5111 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5112 mp->mp_upper - mp->mp_lower);
5113 DPRINTF("node size = %zu", node_size);
5114 return MDB_PAGE_FULL;
5117 /* Move higher pointers up one slot. */
5118 for (i = NUMKEYS(mp); i > indx; i--)
5119 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5121 /* Adjust free space offsets. */
5122 ofs = mp->mp_upper - node_size;
5123 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5124 mp->mp_ptrs[indx] = ofs;
5126 mp->mp_lower += sizeof(indx_t);
5128 /* Write the node data. */
5129 node = NODEPTR(mp, indx);
5130 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5131 node->mn_flags = flags;
5133 SETDSZ(node,data->mv_size);
5138 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5143 if (F_ISSET(flags, F_BIGDATA))
5144 memcpy(node->mn_data + key->mv_size, data->mv_data,
5146 else if (F_ISSET(flags, MDB_RESERVE))
5147 data->mv_data = node->mn_data + key->mv_size;
5149 memcpy(node->mn_data + key->mv_size, data->mv_data,
5152 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5154 if (F_ISSET(flags, MDB_RESERVE))
5155 data->mv_data = METADATA(ofp);
5157 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5164 /** Delete the specified node from a page.
5165 * @param[in] mp The page to operate on.
5166 * @param[in] indx The index of the node to delete.
5167 * @param[in] ksize The size of a node. Only used if the page is
5168 * part of a #MDB_DUPFIXED database.
5171 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5174 indx_t i, j, numkeys, ptr;
5181 COPY_PGNO(pgno, mp->mp_pgno);
5182 DPRINTF("delete node %u on %s page %zu", indx,
5183 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5186 assert(indx < NUMKEYS(mp));
5189 int x = NUMKEYS(mp) - 1 - indx;
5190 base = LEAF2KEY(mp, indx, ksize);
5192 memmove(base, base + ksize, x * ksize);
5193 mp->mp_lower -= sizeof(indx_t);
5194 mp->mp_upper += ksize - sizeof(indx_t);
5198 node = NODEPTR(mp, indx);
5199 sz = NODESIZE + node->mn_ksize;
5201 if (F_ISSET(node->mn_flags, F_BIGDATA))
5202 sz += sizeof(pgno_t);
5204 sz += NODEDSZ(node);
5208 ptr = mp->mp_ptrs[indx];
5209 numkeys = NUMKEYS(mp);
5210 for (i = j = 0; i < numkeys; i++) {
5212 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5213 if (mp->mp_ptrs[i] < ptr)
5214 mp->mp_ptrs[j] += sz;
5219 base = (char *)mp + mp->mp_upper;
5220 memmove(base + sz, base, ptr - mp->mp_upper);
5222 mp->mp_lower -= sizeof(indx_t);
5226 /** Compact the main page after deleting a node on a subpage.
5227 * @param[in] mp The main page to operate on.
5228 * @param[in] indx The index of the subpage on the main page.
5231 mdb_node_shrink(MDB_page *mp, indx_t indx)
5238 indx_t i, numkeys, ptr;
5240 node = NODEPTR(mp, indx);
5241 sp = (MDB_page *)NODEDATA(node);
5242 osize = NODEDSZ(node);
5244 delta = sp->mp_upper - sp->mp_lower;
5245 SETDSZ(node, osize - delta);
5246 xp = (MDB_page *)((char *)sp + delta);
5248 /* shift subpage upward */
5250 nsize = NUMKEYS(sp) * sp->mp_pad;
5251 memmove(METADATA(xp), METADATA(sp), nsize);
5254 nsize = osize - sp->mp_upper;
5255 numkeys = NUMKEYS(sp);
5256 for (i=numkeys-1; i>=0; i--)
5257 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5259 xp->mp_upper = sp->mp_lower;
5260 xp->mp_lower = sp->mp_lower;
5261 xp->mp_flags = sp->mp_flags;
5262 xp->mp_pad = sp->mp_pad;
5263 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5265 /* shift lower nodes upward */
5266 ptr = mp->mp_ptrs[indx];
5267 numkeys = NUMKEYS(mp);
5268 for (i = 0; i < numkeys; i++) {
5269 if (mp->mp_ptrs[i] <= ptr)
5270 mp->mp_ptrs[i] += delta;
5273 base = (char *)mp + mp->mp_upper;
5274 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5275 mp->mp_upper += delta;
5278 /** Initial setup of a sorted-dups cursor.
5279 * Sorted duplicates are implemented as a sub-database for the given key.
5280 * The duplicate data items are actually keys of the sub-database.
5281 * Operations on the duplicate data items are performed using a sub-cursor
5282 * initialized when the sub-database is first accessed. This function does
5283 * the preliminary setup of the sub-cursor, filling in the fields that
5284 * depend only on the parent DB.
5285 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5288 mdb_xcursor_init0(MDB_cursor *mc)
5290 MDB_xcursor *mx = mc->mc_xcursor;
5292 mx->mx_cursor.mc_xcursor = NULL;
5293 mx->mx_cursor.mc_txn = mc->mc_txn;
5294 mx->mx_cursor.mc_db = &mx->mx_db;
5295 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5296 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5297 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5298 mx->mx_cursor.mc_snum = 0;
5299 mx->mx_cursor.mc_top = 0;
5300 mx->mx_cursor.mc_flags = C_SUB;
5301 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5302 mx->mx_dbx.md_dcmp = NULL;
5303 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5306 /** Final setup of a sorted-dups cursor.
5307 * Sets up the fields that depend on the data from the main cursor.
5308 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5309 * @param[in] node The data containing the #MDB_db record for the
5310 * sorted-dup database.
5313 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5315 MDB_xcursor *mx = mc->mc_xcursor;
5317 if (node->mn_flags & F_SUBDATA) {
5318 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5319 mx->mx_cursor.mc_pg[0] = 0;
5320 mx->mx_cursor.mc_snum = 0;
5321 mx->mx_cursor.mc_flags = C_SUB;
5323 MDB_page *fp = NODEDATA(node);
5324 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5325 mx->mx_db.md_flags = 0;
5326 mx->mx_db.md_depth = 1;
5327 mx->mx_db.md_branch_pages = 0;
5328 mx->mx_db.md_leaf_pages = 1;
5329 mx->mx_db.md_overflow_pages = 0;
5330 mx->mx_db.md_entries = NUMKEYS(fp);
5331 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5332 mx->mx_cursor.mc_snum = 1;
5333 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5334 mx->mx_cursor.mc_top = 0;
5335 mx->mx_cursor.mc_pg[0] = fp;
5336 mx->mx_cursor.mc_ki[0] = 0;
5337 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5338 mx->mx_db.md_flags = MDB_DUPFIXED;
5339 mx->mx_db.md_pad = fp->mp_pad;
5340 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5341 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5344 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5346 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5348 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5349 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5350 #if UINT_MAX < SIZE_MAX
5351 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5352 #ifdef MISALIGNED_OK
5353 mx->mx_dbx.md_cmp = mdb_cmp_long;
5355 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5360 /** Initialize a cursor for a given transaction and database. */
5362 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5367 mc->mc_db = &txn->mt_dbs[dbi];
5368 mc->mc_dbx = &txn->mt_dbxs[dbi];
5369 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5374 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5376 mc->mc_xcursor = mx;
5377 mdb_xcursor_init0(mc);
5379 mc->mc_xcursor = NULL;
5381 if (*mc->mc_dbflag & DB_STALE) {
5382 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5387 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5390 MDB_xcursor *mx = NULL;
5391 size_t size = sizeof(MDB_cursor);
5393 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5396 /* Allow read access to the freelist */
5397 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5400 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5401 size += sizeof(MDB_xcursor);
5403 if ((mc = malloc(size)) != NULL) {
5404 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5405 mx = (MDB_xcursor *)(mc + 1);
5407 mdb_cursor_init(mc, txn, dbi, mx);
5408 if (txn->mt_cursors) {
5409 mc->mc_next = txn->mt_cursors[dbi];
5410 txn->mt_cursors[dbi] = mc;
5412 mc->mc_flags |= C_ALLOCD;
5423 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5425 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5428 if (txn->mt_cursors)
5431 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5435 /* Return the count of duplicate data items for the current key */
5437 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5441 if (mc == NULL || countp == NULL)
5444 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5447 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5448 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5451 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5454 *countp = mc->mc_xcursor->mx_db.md_entries;
5460 mdb_cursor_close(MDB_cursor *mc)
5463 /* remove from txn, if tracked */
5464 if (mc->mc_txn->mt_cursors) {
5465 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5466 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5468 *prev = mc->mc_next;
5470 if (mc->mc_flags & C_ALLOCD)
5476 mdb_cursor_txn(MDB_cursor *mc)
5478 if (!mc) return NULL;
5483 mdb_cursor_dbi(MDB_cursor *mc)
5489 /** Replace the key for a node with a new key.
5490 * @param[in] mp The page containing the node to operate on.
5491 * @param[in] indx The index of the node to operate on.
5492 * @param[in] key The new key to use.
5493 * @return 0 on success, non-zero on failure.
5496 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5502 indx_t ptr, i, numkeys;
5505 node = NODEPTR(mp, indx);
5506 ptr = mp->mp_ptrs[indx];
5510 char kbuf2[(MAXKEYSIZE*2+1)];
5511 k2.mv_data = NODEKEY(node);
5512 k2.mv_size = node->mn_ksize;
5513 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5515 mdb_dkey(&k2, kbuf2),
5521 delta0 = delta = key->mv_size - node->mn_ksize;
5523 /* Must be 2-byte aligned. If new key is
5524 * shorter by 1, the shift will be skipped.
5526 delta += (delta & 1);
5528 if (delta > 0 && SIZELEFT(mp) < delta) {
5529 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5530 return MDB_PAGE_FULL;
5533 numkeys = NUMKEYS(mp);
5534 for (i = 0; i < numkeys; i++) {
5535 if (mp->mp_ptrs[i] <= ptr)
5536 mp->mp_ptrs[i] -= delta;
5539 base = (char *)mp + mp->mp_upper;
5540 len = ptr - mp->mp_upper + NODESIZE;
5541 memmove(base - delta, base, len);
5542 mp->mp_upper -= delta;
5544 node = NODEPTR(mp, indx);
5547 /* But even if no shift was needed, update ksize */
5549 node->mn_ksize = key->mv_size;
5552 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5557 /** Move a node from csrc to cdst.
5560 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5566 unsigned short flags;
5570 /* Mark src and dst as dirty. */
5571 if ((rc = mdb_page_touch(csrc)) ||
5572 (rc = mdb_page_touch(cdst)))
5575 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5576 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5577 key.mv_size = csrc->mc_db->md_pad;
5578 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5580 data.mv_data = NULL;
5584 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5585 assert(!((long)srcnode&1));
5586 srcpg = NODEPGNO(srcnode);
5587 flags = srcnode->mn_flags;
5588 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5589 unsigned int snum = csrc->mc_snum;
5591 /* must find the lowest key below src */
5592 mdb_page_search_root(csrc, NULL, 0);
5593 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5594 key.mv_size = csrc->mc_db->md_pad;
5595 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5597 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5598 key.mv_size = NODEKSZ(s2);
5599 key.mv_data = NODEKEY(s2);
5601 csrc->mc_snum = snum--;
5602 csrc->mc_top = snum;
5604 key.mv_size = NODEKSZ(srcnode);
5605 key.mv_data = NODEKEY(srcnode);
5607 data.mv_size = NODEDSZ(srcnode);
5608 data.mv_data = NODEDATA(srcnode);
5610 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5611 unsigned int snum = cdst->mc_snum;
5614 /* must find the lowest key below dst */
5615 mdb_page_search_root(cdst, NULL, 0);
5616 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5617 bkey.mv_size = cdst->mc_db->md_pad;
5618 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5620 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5621 bkey.mv_size = NODEKSZ(s2);
5622 bkey.mv_data = NODEKEY(s2);
5624 cdst->mc_snum = snum--;
5625 cdst->mc_top = snum;
5626 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5629 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5630 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5631 csrc->mc_ki[csrc->mc_top],
5633 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5634 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5636 /* Add the node to the destination page.
5638 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5639 if (rc != MDB_SUCCESS)
5642 /* Delete the node from the source page.
5644 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5647 /* Adjust other cursors pointing to mp */
5648 MDB_cursor *m2, *m3;
5649 MDB_dbi dbi = csrc->mc_dbi;
5650 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5652 if (csrc->mc_flags & C_SUB)
5655 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5656 if (m2 == csrc) continue;
5657 if (csrc->mc_flags & C_SUB)
5658 m3 = &m2->mc_xcursor->mx_cursor;
5661 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5662 csrc->mc_ki[csrc->mc_top]) {
5663 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5664 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5669 /* Update the parent separators.
5671 if (csrc->mc_ki[csrc->mc_top] == 0) {
5672 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5673 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5674 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5676 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5677 key.mv_size = NODEKSZ(srcnode);
5678 key.mv_data = NODEKEY(srcnode);
5680 DPRINTF("update separator for source page %zu to [%s]",
5681 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5682 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5683 &key)) != MDB_SUCCESS)
5686 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5688 nullkey.mv_size = 0;
5689 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5690 assert(rc == MDB_SUCCESS);
5694 if (cdst->mc_ki[cdst->mc_top] == 0) {
5695 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5696 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5697 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5699 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5700 key.mv_size = NODEKSZ(srcnode);
5701 key.mv_data = NODEKEY(srcnode);
5703 DPRINTF("update separator for destination page %zu to [%s]",
5704 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5705 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5706 &key)) != MDB_SUCCESS)
5709 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5711 nullkey.mv_size = 0;
5712 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5713 assert(rc == MDB_SUCCESS);
5720 /** Merge one page into another.
5721 * The nodes from the page pointed to by \b csrc will
5722 * be copied to the page pointed to by \b cdst and then
5723 * the \b csrc page will be freed.
5724 * @param[in] csrc Cursor pointing to the source page.
5725 * @param[in] cdst Cursor pointing to the destination page.
5728 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5736 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5737 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5739 assert(csrc->mc_snum > 1); /* can't merge root page */
5740 assert(cdst->mc_snum > 1);
5742 /* Mark dst as dirty. */
5743 if ((rc = mdb_page_touch(cdst)))
5746 /* Move all nodes from src to dst.
5748 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5749 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5750 key.mv_size = csrc->mc_db->md_pad;
5751 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5752 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5753 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5754 if (rc != MDB_SUCCESS)
5756 key.mv_data = (char *)key.mv_data + key.mv_size;
5759 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5760 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5761 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5762 unsigned int snum = csrc->mc_snum;
5764 /* must find the lowest key below src */
5765 mdb_page_search_root(csrc, NULL, 0);
5766 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5767 key.mv_size = csrc->mc_db->md_pad;
5768 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5770 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5771 key.mv_size = NODEKSZ(s2);
5772 key.mv_data = NODEKEY(s2);
5774 csrc->mc_snum = snum--;
5775 csrc->mc_top = snum;
5777 key.mv_size = srcnode->mn_ksize;
5778 key.mv_data = NODEKEY(srcnode);
5781 data.mv_size = NODEDSZ(srcnode);
5782 data.mv_data = NODEDATA(srcnode);
5783 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5784 if (rc != MDB_SUCCESS)
5789 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5790 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);
5792 /* Unlink the src page from parent and add to free list.
5794 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5795 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5797 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5801 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5802 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5803 csrc->mc_db->md_leaf_pages--;
5805 csrc->mc_db->md_branch_pages--;
5807 /* Adjust other cursors pointing to mp */
5808 MDB_cursor *m2, *m3;
5809 MDB_dbi dbi = csrc->mc_dbi;
5810 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5812 if (csrc->mc_flags & C_SUB)
5815 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5816 if (csrc->mc_flags & C_SUB)
5817 m3 = &m2->mc_xcursor->mx_cursor;
5820 if (m3 == csrc) continue;
5821 if (m3->mc_snum < csrc->mc_snum) continue;
5822 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5823 m3->mc_pg[csrc->mc_top] = mp;
5824 m3->mc_ki[csrc->mc_top] += nkeys;
5828 mdb_cursor_pop(csrc);
5830 return mdb_rebalance(csrc);
5833 /** Copy the contents of a cursor.
5834 * @param[in] csrc The cursor to copy from.
5835 * @param[out] cdst The cursor to copy to.
5838 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5842 cdst->mc_txn = csrc->mc_txn;
5843 cdst->mc_dbi = csrc->mc_dbi;
5844 cdst->mc_db = csrc->mc_db;
5845 cdst->mc_dbx = csrc->mc_dbx;
5846 cdst->mc_snum = csrc->mc_snum;
5847 cdst->mc_top = csrc->mc_top;
5848 cdst->mc_flags = csrc->mc_flags;
5850 for (i=0; i<csrc->mc_snum; i++) {
5851 cdst->mc_pg[i] = csrc->mc_pg[i];
5852 cdst->mc_ki[i] = csrc->mc_ki[i];
5856 /** Rebalance the tree after a delete operation.
5857 * @param[in] mc Cursor pointing to the page where rebalancing
5859 * @return 0 on success, non-zero on failure.
5862 mdb_rebalance(MDB_cursor *mc)
5872 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5873 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5874 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5875 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5879 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5882 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5883 DPRINTF("no need to rebalance page %zu, above fill threshold",
5889 if (mc->mc_snum < 2) {
5890 MDB_page *mp = mc->mc_pg[0];
5891 if (NUMKEYS(mp) == 0) {
5892 DPUTS("tree is completely empty");
5893 mc->mc_db->md_root = P_INVALID;
5894 mc->mc_db->md_depth = 0;
5895 mc->mc_db->md_leaf_pages = 0;
5896 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5900 /* Adjust other cursors pointing to mp */
5901 MDB_cursor *m2, *m3;
5902 MDB_dbi dbi = mc->mc_dbi;
5904 if (mc->mc_flags & C_SUB)
5907 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5908 if (m2 == mc) continue;
5909 if (mc->mc_flags & C_SUB)
5910 m3 = &m2->mc_xcursor->mx_cursor;
5913 if (m3->mc_snum < mc->mc_snum) continue;
5914 if (m3->mc_pg[0] == mp) {
5920 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5921 DPUTS("collapsing root page!");
5922 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5923 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5924 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5927 mc->mc_db->md_depth--;
5928 mc->mc_db->md_branch_pages--;
5930 /* Adjust other cursors pointing to mp */
5931 MDB_cursor *m2, *m3;
5932 MDB_dbi dbi = mc->mc_dbi;
5934 if (mc->mc_flags & C_SUB)
5937 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5938 if (m2 == mc) continue;
5939 if (mc->mc_flags & C_SUB)
5940 m3 = &m2->mc_xcursor->mx_cursor;
5943 if (m3->mc_snum < mc->mc_snum) continue;
5944 if (m3->mc_pg[0] == mp) {
5945 m3->mc_pg[0] = mc->mc_pg[0];
5950 DPUTS("root page doesn't need rebalancing");
5954 /* The parent (branch page) must have at least 2 pointers,
5955 * otherwise the tree is invalid.
5957 ptop = mc->mc_top-1;
5958 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5960 /* Leaf page fill factor is below the threshold.
5961 * Try to move keys from left or right neighbor, or
5962 * merge with a neighbor page.
5967 mdb_cursor_copy(mc, &mn);
5968 mn.mc_xcursor = NULL;
5970 if (mc->mc_ki[ptop] == 0) {
5971 /* We're the leftmost leaf in our parent.
5973 DPUTS("reading right neighbor");
5975 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5976 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5978 mn.mc_ki[mn.mc_top] = 0;
5979 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5981 /* There is at least one neighbor to the left.
5983 DPUTS("reading left neighbor");
5985 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5986 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5988 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5989 mc->mc_ki[mc->mc_top] = 0;
5992 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5993 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);
5995 /* If the neighbor page is above threshold and has at least two
5996 * keys, move one key from it.
5998 * Otherwise we should try to merge them.
6000 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6001 return mdb_node_move(&mn, mc);
6002 else { /* FIXME: if (has_enough_room()) */
6003 mc->mc_flags &= ~C_INITIALIZED;
6004 if (mc->mc_ki[ptop] == 0)
6005 return mdb_page_merge(&mn, mc);
6007 return mdb_page_merge(mc, &mn);
6011 /** Complete a delete operation started by #mdb_cursor_del(). */
6013 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6017 /* add overflow pages to free list */
6018 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6022 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6023 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6024 mc->mc_db->md_overflow_pages -= ovpages;
6025 for (i=0; i<ovpages; i++) {
6026 DPRINTF("freed ov page %zu", pg);
6027 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6031 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6032 mc->mc_db->md_entries--;
6033 rc = mdb_rebalance(mc);
6034 if (rc != MDB_SUCCESS)
6035 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6041 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6042 MDB_val *key, MDB_val *data)
6047 MDB_val rdata, *xdata;
6051 assert(key != NULL);
6053 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6055 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6058 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6062 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6066 mdb_cursor_init(&mc, txn, dbi, &mx);
6077 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6079 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6083 /** Split a page and insert a new node.
6084 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6085 * The cursor will be updated to point to the actual page and index where
6086 * the node got inserted after the split.
6087 * @param[in] newkey The key for the newly inserted node.
6088 * @param[in] newdata The data for the newly inserted node.
6089 * @param[in] newpgno The page number, if the new node is a branch node.
6090 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6091 * @return 0 on success, non-zero on failure.
6094 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6095 unsigned int nflags)
6098 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6101 unsigned int i, j, split_indx, nkeys, pmax;
6103 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6105 MDB_page *mp, *rp, *pp;
6110 mp = mc->mc_pg[mc->mc_top];
6111 newindx = mc->mc_ki[mc->mc_top];
6113 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6114 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6115 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6117 /* Create a right sibling. */
6118 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6120 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6122 if (mc->mc_snum < 2) {
6123 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6125 /* shift current top to make room for new parent */
6126 mc->mc_pg[1] = mc->mc_pg[0];
6127 mc->mc_ki[1] = mc->mc_ki[0];
6130 mc->mc_db->md_root = pp->mp_pgno;
6131 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6132 mc->mc_db->md_depth++;
6135 /* Add left (implicit) pointer. */
6136 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6137 /* undo the pre-push */
6138 mc->mc_pg[0] = mc->mc_pg[1];
6139 mc->mc_ki[0] = mc->mc_ki[1];
6140 mc->mc_db->md_root = mp->mp_pgno;
6141 mc->mc_db->md_depth--;
6148 ptop = mc->mc_top-1;
6149 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6152 mc->mc_flags |= C_SPLITTING;
6153 mdb_cursor_copy(mc, &mn);
6154 mn.mc_pg[mn.mc_top] = rp;
6155 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6157 if (nflags & MDB_APPEND) {
6158 mn.mc_ki[mn.mc_top] = 0;
6160 split_indx = newindx;
6165 nkeys = NUMKEYS(mp);
6166 split_indx = nkeys / 2;
6167 if (newindx < split_indx)
6173 unsigned int lsize, rsize, ksize;
6174 /* Move half of the keys to the right sibling */
6176 x = mc->mc_ki[mc->mc_top] - split_indx;
6177 ksize = mc->mc_db->md_pad;
6178 split = LEAF2KEY(mp, split_indx, ksize);
6179 rsize = (nkeys - split_indx) * ksize;
6180 lsize = (nkeys - split_indx) * sizeof(indx_t);
6181 mp->mp_lower -= lsize;
6182 rp->mp_lower += lsize;
6183 mp->mp_upper += rsize - lsize;
6184 rp->mp_upper -= rsize - lsize;
6185 sepkey.mv_size = ksize;
6186 if (newindx == split_indx) {
6187 sepkey.mv_data = newkey->mv_data;
6189 sepkey.mv_data = split;
6192 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6193 memcpy(rp->mp_ptrs, split, rsize);
6194 sepkey.mv_data = rp->mp_ptrs;
6195 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6196 memcpy(ins, newkey->mv_data, ksize);
6197 mp->mp_lower += sizeof(indx_t);
6198 mp->mp_upper -= ksize - sizeof(indx_t);
6201 memcpy(rp->mp_ptrs, split, x * ksize);
6202 ins = LEAF2KEY(rp, x, ksize);
6203 memcpy(ins, newkey->mv_data, ksize);
6204 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6205 rp->mp_lower += sizeof(indx_t);
6206 rp->mp_upper -= ksize - sizeof(indx_t);
6207 mc->mc_ki[mc->mc_top] = x;
6208 mc->mc_pg[mc->mc_top] = rp;
6213 /* For leaf pages, check the split point based on what
6214 * fits where, since otherwise mdb_node_add can fail.
6216 * This check is only needed when the data items are
6217 * relatively large, such that being off by one will
6218 * make the difference between success or failure.
6220 * It's also relevant if a page happens to be laid out
6221 * such that one half of its nodes are all "small" and
6222 * the other half of its nodes are "large." If the new
6223 * item is also "large" and falls on the half with
6224 * "large" nodes, it also may not fit.
6227 unsigned int psize, nsize;
6228 /* Maximum free space in an empty page */
6229 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6230 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6231 if ((nkeys < 20) || (nsize > pmax/16)) {
6232 if (newindx <= split_indx) {
6235 for (i=0; i<split_indx; i++) {
6236 node = NODEPTR(mp, i);
6237 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6238 if (F_ISSET(node->mn_flags, F_BIGDATA))
6239 psize += sizeof(pgno_t);
6241 psize += NODEDSZ(node);
6245 split_indx = newindx;
6256 for (i=nkeys-1; i>=split_indx; i--) {
6257 node = NODEPTR(mp, i);
6258 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6259 if (F_ISSET(node->mn_flags, F_BIGDATA))
6260 psize += sizeof(pgno_t);
6262 psize += NODEDSZ(node);
6266 split_indx = newindx;
6277 /* First find the separating key between the split pages.
6278 * The case where newindx == split_indx is ambiguous; the
6279 * new item could go to the new page or stay on the original
6280 * page. If newpos == 1 it goes to the new page.
6282 if (newindx == split_indx && newpos) {
6283 sepkey.mv_size = newkey->mv_size;
6284 sepkey.mv_data = newkey->mv_data;
6286 node = NODEPTR(mp, split_indx);
6287 sepkey.mv_size = node->mn_ksize;
6288 sepkey.mv_data = NODEKEY(node);
6292 DPRINTF("separator is [%s]", DKEY(&sepkey));
6294 /* Copy separator key to the parent.
6296 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6300 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6303 if (mn.mc_snum == mc->mc_snum) {
6304 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6305 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6306 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6307 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6312 /* Right page might now have changed parent.
6313 * Check if left page also changed parent.
6315 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6316 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6317 for (i=0; i<ptop; i++) {
6318 mc->mc_pg[i] = mn.mc_pg[i];
6319 mc->mc_ki[i] = mn.mc_ki[i];
6321 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6322 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6326 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6329 mc->mc_flags ^= C_SPLITTING;
6330 if (rc != MDB_SUCCESS) {
6333 if (nflags & MDB_APPEND) {
6334 mc->mc_pg[mc->mc_top] = rp;
6335 mc->mc_ki[mc->mc_top] = 0;
6336 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6339 for (i=0; i<mc->mc_top; i++)
6340 mc->mc_ki[i] = mn.mc_ki[i];
6347 /* Move half of the keys to the right sibling. */
6349 /* grab a page to hold a temporary copy */
6350 copy = mdb_page_malloc(mc);
6354 copy->mp_pgno = mp->mp_pgno;
6355 copy->mp_flags = mp->mp_flags;
6356 copy->mp_lower = PAGEHDRSZ;
6357 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6358 mc->mc_pg[mc->mc_top] = copy;
6359 for (i = j = 0; i <= nkeys; j++) {
6360 if (i == split_indx) {
6361 /* Insert in right sibling. */
6362 /* Reset insert index for right sibling. */
6363 if (i != newindx || (newpos ^ ins_new)) {
6365 mc->mc_pg[mc->mc_top] = rp;
6369 if (i == newindx && !ins_new) {
6370 /* Insert the original entry that caused the split. */
6371 rkey.mv_data = newkey->mv_data;
6372 rkey.mv_size = newkey->mv_size;
6381 /* Update index for the new key. */
6382 mc->mc_ki[mc->mc_top] = j;
6383 } else if (i == nkeys) {
6386 node = NODEPTR(mp, i);
6387 rkey.mv_data = NODEKEY(node);
6388 rkey.mv_size = node->mn_ksize;
6390 xdata.mv_data = NODEDATA(node);
6391 xdata.mv_size = NODEDSZ(node);
6394 pgno = NODEPGNO(node);
6395 flags = node->mn_flags;
6400 if (!IS_LEAF(mp) && j == 0) {
6401 /* First branch index doesn't need key data. */
6405 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6409 nkeys = NUMKEYS(copy);
6410 for (i=0; i<nkeys; i++)
6411 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6412 mp->mp_lower = copy->mp_lower;
6413 mp->mp_upper = copy->mp_upper;
6414 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6415 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6417 /* reset back to original page */
6418 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6419 mc->mc_pg[mc->mc_top] = mp;
6420 if (nflags & MDB_RESERVE) {
6421 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6422 if (!(node->mn_flags & F_BIGDATA))
6423 newdata->mv_data = NODEDATA(node);
6429 /* return tmp page to freelist */
6430 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
6431 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
6432 mc->mc_txn->mt_env->me_dpages = copy;
6435 /* Adjust other cursors pointing to mp */
6436 MDB_cursor *m2, *m3;
6437 MDB_dbi dbi = mc->mc_dbi;
6438 int fixup = NUMKEYS(mp);
6440 if (mc->mc_flags & C_SUB)
6443 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6444 if (m2 == mc) continue;
6445 if (mc->mc_flags & C_SUB)
6446 m3 = &m2->mc_xcursor->mx_cursor;
6449 if (!(m3->mc_flags & C_INITIALIZED))
6451 if (m3->mc_flags & C_SPLITTING)
6456 for (k=m3->mc_top; k>=0; k--) {
6457 m3->mc_ki[k+1] = m3->mc_ki[k];
6458 m3->mc_pg[k+1] = m3->mc_pg[k];
6460 if (m3->mc_ki[0] >= split_indx) {
6465 m3->mc_pg[0] = mc->mc_pg[0];
6469 if (m3->mc_pg[mc->mc_top] == mp) {
6470 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6471 m3->mc_ki[mc->mc_top]++;
6472 if (m3->mc_ki[mc->mc_top] >= fixup) {
6473 m3->mc_pg[mc->mc_top] = rp;
6474 m3->mc_ki[mc->mc_top] -= fixup;
6475 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6477 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6478 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6487 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6488 MDB_val *key, MDB_val *data, unsigned int flags)
6493 assert(key != NULL);
6494 assert(data != NULL);
6496 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6499 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6503 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6507 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6510 mdb_cursor_init(&mc, txn, dbi, &mx);
6511 return mdb_cursor_put(&mc, key, data, flags);
6514 /** Only a subset of the @ref mdb_env flags can be changed
6515 * at runtime. Changing other flags requires closing the environment
6516 * and re-opening it with the new flags.
6518 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
6520 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6522 if ((flag & CHANGEABLE) != flag)
6525 env->me_flags |= flag;
6527 env->me_flags &= ~flag;
6532 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6537 *arg = env->me_flags;
6542 mdb_env_get_path(MDB_env *env, const char **arg)
6547 *arg = env->me_path;
6551 /** Common code for #mdb_stat() and #mdb_env_stat().
6552 * @param[in] env the environment to operate in.
6553 * @param[in] db the #MDB_db record containing the stats to return.
6554 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6555 * @return 0, this function always succeeds.
6558 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6560 arg->ms_psize = env->me_psize;
6561 arg->ms_depth = db->md_depth;
6562 arg->ms_branch_pages = db->md_branch_pages;
6563 arg->ms_leaf_pages = db->md_leaf_pages;
6564 arg->ms_overflow_pages = db->md_overflow_pages;
6565 arg->ms_entries = db->md_entries;
6570 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6574 if (env == NULL || arg == NULL)
6577 toggle = mdb_env_pick_meta(env);
6579 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6582 /** Set the default comparison functions for a database.
6583 * Called immediately after a database is opened to set the defaults.
6584 * The user can then override them with #mdb_set_compare() or
6585 * #mdb_set_dupsort().
6586 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6587 * @param[in] dbi A database handle returned by #mdb_open()
6590 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6592 uint16_t f = txn->mt_dbs[dbi].md_flags;
6594 txn->mt_dbxs[dbi].md_cmp =
6595 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6596 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6598 txn->mt_dbxs[dbi].md_dcmp =
6599 !(f & MDB_DUPSORT) ? 0 :
6600 ((f & MDB_INTEGERDUP)
6601 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6602 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6605 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6610 int rc, dbflag, exact;
6611 unsigned int unused = 0;
6614 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6615 mdb_default_cmp(txn, FREE_DBI);
6621 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
6622 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
6623 mdb_default_cmp(txn, MAIN_DBI);
6627 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6628 mdb_default_cmp(txn, MAIN_DBI);
6631 /* Is the DB already open? */
6633 for (i=2; i<txn->mt_numdbs; i++) {
6634 if (!txn->mt_dbxs[i].md_name.mv_size) {
6635 /* Remember this free slot */
6636 if (!unused) unused = i;
6639 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6640 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6646 /* If no free slot and max hit, fail */
6647 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
6648 return MDB_DBS_FULL;
6650 /* Find the DB info */
6654 key.mv_data = (void *)name;
6655 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6656 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6657 if (rc == MDB_SUCCESS) {
6658 /* make sure this is actually a DB */
6659 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6660 if (!(node->mn_flags & F_SUBDATA))
6662 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6663 /* Create if requested */
6665 data.mv_size = sizeof(MDB_db);
6666 data.mv_data = &dummy;
6667 memset(&dummy, 0, sizeof(dummy));
6668 dummy.md_root = P_INVALID;
6669 dummy.md_flags = flags & 0xffff;
6670 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6674 /* OK, got info, add to table */
6675 if (rc == MDB_SUCCESS) {
6676 unsigned int slot = unused ? unused : txn->mt_numdbs;
6677 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
6678 txn->mt_dbxs[slot].md_name.mv_size = len;
6679 txn->mt_dbxs[slot].md_rel = NULL;
6680 txn->mt_dbflags[slot] = dbflag;
6681 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
6683 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
6684 mdb_default_cmp(txn, slot);
6687 txn->mt_env->me_numdbs++;
6694 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6696 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6699 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6702 void mdb_close(MDB_env *env, MDB_dbi dbi)
6705 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6707 ptr = env->me_dbxs[dbi].md_name.mv_data;
6708 env->me_dbxs[dbi].md_name.mv_data = NULL;
6709 env->me_dbxs[dbi].md_name.mv_size = 0;
6713 /** Add all the DB's pages to the free list.
6714 * @param[in] mc Cursor on the DB to free.
6715 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6716 * @return 0 on success, non-zero on failure.
6719 mdb_drop0(MDB_cursor *mc, int subs)
6723 rc = mdb_page_search(mc, NULL, 0);
6724 if (rc == MDB_SUCCESS) {
6729 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6730 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6733 mdb_cursor_copy(mc, &mx);
6734 while (mc->mc_snum > 0) {
6735 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6736 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6737 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6738 if (ni->mn_flags & F_SUBDATA) {
6739 mdb_xcursor_init1(mc, ni);
6740 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6746 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6748 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6751 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6756 rc = mdb_cursor_sibling(mc, 1);
6758 /* no more siblings, go back to beginning
6759 * of previous level. (stack was already popped
6760 * by mdb_cursor_sibling)
6762 for (i=1; i<mc->mc_top; i++)
6763 mc->mc_pg[i] = mx.mc_pg[i];
6767 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6768 mc->mc_db->md_root);
6773 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6778 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6781 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6784 rc = mdb_cursor_open(txn, dbi, &mc);
6788 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6792 /* Can't delete the main DB */
6793 if (del && dbi > MAIN_DBI) {
6794 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6796 mdb_close(txn->mt_env, dbi);
6798 /* reset the DB record, mark it dirty */
6799 txn->mt_dbflags[dbi] |= DB_DIRTY;
6800 txn->mt_dbs[dbi].md_depth = 0;
6801 txn->mt_dbs[dbi].md_branch_pages = 0;
6802 txn->mt_dbs[dbi].md_leaf_pages = 0;
6803 txn->mt_dbs[dbi].md_overflow_pages = 0;
6804 txn->mt_dbs[dbi].md_entries = 0;
6805 txn->mt_dbs[dbi].md_root = P_INVALID;
6807 if (!txn->mt_u.dirty_list[0].mid) {
6810 /* make sure we have at least one dirty page in this txn
6811 * otherwise these changes will be ignored.
6813 key.mv_size = sizeof(txnid_t);
6814 key.mv_data = &txn->mt_txnid;
6815 data.mv_size = sizeof(MDB_ID);
6816 data.mv_data = txn->mt_free_pgs;
6817 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
6818 rc = mdb_cursor_put(&m2, &key, &data, 0);
6822 mdb_cursor_close(mc);
6826 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6828 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6831 txn->mt_dbxs[dbi].md_cmp = cmp;
6835 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6837 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6840 txn->mt_dbxs[dbi].md_dcmp = cmp;
6844 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6846 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6849 txn->mt_dbxs[dbi].md_rel = rel;
6853 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6855 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6858 txn->mt_dbxs[dbi].md_relctx = ctx;