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 /** Read-only Filesystem. Allow read access, no locking. */
917 #define MDB_ROFS 0x40000000U
918 /** Some fields are initialized. */
919 #define MDB_ENV_ACTIVE 0x20000000U
920 uint32_t me_flags; /**< @ref mdb_env */
921 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
922 unsigned int me_maxreaders; /**< size of the reader table */
923 unsigned int me_numreaders; /**< max numreaders set by this env */
924 MDB_dbi me_numdbs; /**< number of DBs opened */
925 MDB_dbi me_maxdbs; /**< size of the DB table */
926 pid_t me_pid; /**< process ID of this env */
927 char *me_path; /**< path to the DB files */
928 char *me_map; /**< the memory map of the data file */
929 MDB_txninfo *me_txns; /**< the memory map of the lock file */
930 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
931 MDB_txn *me_txn; /**< current write transaction */
932 size_t me_mapsize; /**< size of the data memory map */
933 off_t me_size; /**< current file size */
934 pgno_t me_maxpg; /**< me_mapsize / me_psize */
935 txnid_t me_pgfirst; /**< ID of first old page record we used */
936 txnid_t me_pglast; /**< ID of last old page record we used */
937 MDB_dbx *me_dbxs; /**< array of static DB info */
938 uint16_t *me_dbflags; /**< array of DB flags */
939 MDB_oldpages *me_pghead; /**< list of old page records */
940 MDB_oldpages *me_pgfree; /**< list of page records to free */
941 pthread_key_t me_txkey; /**< thread-key for readers */
942 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
943 /** IDL of pages that became unused in a write txn */
945 /** ID2L of pages that were written during a write txn */
946 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
948 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
950 #elif defined(MDB_USE_POSIX_SEM)
951 sem_t *me_rmutex; /* Shared mutexes are not supported */
955 /** max number of pages to commit in one writev() call */
956 #define MDB_COMMIT_PAGES 64
957 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
958 #undef MDB_COMMIT_PAGES
959 #define MDB_COMMIT_PAGES IOV_MAX
962 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
963 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
964 static int mdb_page_touch(MDB_cursor *mc);
966 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
967 static int mdb_page_search_root(MDB_cursor *mc,
968 MDB_val *key, int modify);
969 #define MDB_PS_MODIFY 1
970 #define MDB_PS_ROOTONLY 2
971 static int mdb_page_search(MDB_cursor *mc,
972 MDB_val *key, int flags);
973 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
975 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
976 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
977 pgno_t newpgno, unsigned int nflags);
979 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
980 static int mdb_env_pick_meta(const MDB_env *env);
981 static int mdb_env_write_meta(MDB_txn *txn);
982 static void mdb_env_close0(MDB_env *env, int excl);
984 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
985 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
986 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
987 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
988 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
989 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
990 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
991 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
992 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
994 static int mdb_rebalance(MDB_cursor *mc);
995 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
997 static void mdb_cursor_pop(MDB_cursor *mc);
998 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1000 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1001 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1002 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1003 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1004 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1006 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1007 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1009 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1010 static void mdb_xcursor_init0(MDB_cursor *mc);
1011 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1013 static int mdb_drop0(MDB_cursor *mc, int subs);
1014 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1017 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1021 static SECURITY_DESCRIPTOR mdb_null_sd;
1022 static SECURITY_ATTRIBUTES mdb_all_sa;
1023 static int mdb_sec_inited;
1026 /** Return the library version info. */
1028 mdb_version(int *major, int *minor, int *patch)
1030 if (major) *major = MDB_VERSION_MAJOR;
1031 if (minor) *minor = MDB_VERSION_MINOR;
1032 if (patch) *patch = MDB_VERSION_PATCH;
1033 return MDB_VERSION_STRING;
1036 /** Table of descriptions for MDB @ref errors */
1037 static char *const mdb_errstr[] = {
1038 "MDB_KEYEXIST: Key/data pair already exists",
1039 "MDB_NOTFOUND: No matching key/data pair found",
1040 "MDB_PAGE_NOTFOUND: Requested page not found",
1041 "MDB_CORRUPTED: Located page was wrong type",
1042 "MDB_PANIC: Update of meta page failed",
1043 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1044 "MDB_INVALID: File is not an MDB file",
1045 "MDB_MAP_FULL: Environment mapsize limit reached",
1046 "MDB_DBS_FULL: Environment maxdbs limit reached",
1047 "MDB_READERS_FULL: Environment maxreaders limit reached",
1048 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1049 "MDB_TXN_FULL: Nested transaction has too many dirty pages - transaction too big",
1050 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1051 "MDB_PAGE_FULL: Internal error - page has no more space"
1055 mdb_strerror(int err)
1059 return ("Successful return: 0");
1061 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1062 i = err - MDB_KEYEXIST;
1063 return mdb_errstr[i];
1066 return strerror(err);
1070 /** Display a key in hexadecimal and return the address of the result.
1071 * @param[in] key the key to display
1072 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1073 * @return The key in hexadecimal form.
1076 mdb_dkey(MDB_val *key, char *buf)
1079 unsigned char *c = key->mv_data;
1081 if (key->mv_size > MAXKEYSIZE)
1082 return "MAXKEYSIZE";
1083 /* may want to make this a dynamic check: if the key is mostly
1084 * printable characters, print it as-is instead of converting to hex.
1088 for (i=0; i<key->mv_size; i++)
1089 ptr += sprintf(ptr, "%02x", *c++);
1091 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1096 /** Display all the keys in the page. */
1098 mdb_page_list(MDB_page *mp)
1101 unsigned int i, nkeys, nsize;
1105 nkeys = NUMKEYS(mp);
1106 fprintf(stderr, "numkeys %d\n", nkeys);
1107 for (i=0; i<nkeys; i++) {
1108 node = NODEPTR(mp, i);
1109 key.mv_size = node->mn_ksize;
1110 key.mv_data = node->mn_data;
1111 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1112 if (F_ISSET(node->mn_flags, F_BIGDATA))
1113 nsize += sizeof(pgno_t);
1115 nsize += NODEDSZ(node);
1116 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1121 mdb_cursor_chk(MDB_cursor *mc)
1127 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1128 for (i=0; i<mc->mc_top; i++) {
1130 node = NODEPTR(mp, mc->mc_ki[i]);
1131 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1134 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1140 /** Count all the pages in each DB and in the freelist
1141 * and make sure it matches the actual number of pages
1144 static void mdb_audit(MDB_txn *txn)
1148 MDB_ID freecount, count;
1153 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1154 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1155 freecount += *(MDB_ID *)data.mv_data;
1158 for (i = 0; i<txn->mt_numdbs; i++) {
1159 MDB_xcursor mx, *mxp;
1160 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1161 mdb_cursor_init(&mc, txn, i, mxp);
1162 if (txn->mt_dbs[i].md_root == P_INVALID)
1164 count += txn->mt_dbs[i].md_branch_pages +
1165 txn->mt_dbs[i].md_leaf_pages +
1166 txn->mt_dbs[i].md_overflow_pages;
1167 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1168 mdb_page_search(&mc, NULL, 0);
1172 mp = mc.mc_pg[mc.mc_top];
1173 for (j=0; j<NUMKEYS(mp); j++) {
1174 MDB_node *leaf = NODEPTR(mp, j);
1175 if (leaf->mn_flags & F_SUBDATA) {
1177 memcpy(&db, NODEDATA(leaf), sizeof(db));
1178 count += db.md_branch_pages + db.md_leaf_pages +
1179 db.md_overflow_pages;
1183 while (mdb_cursor_sibling(&mc, 1) == 0);
1186 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1187 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1188 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1194 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1196 return txn->mt_dbxs[dbi].md_cmp(a, b);
1200 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1202 if (txn->mt_dbxs[dbi].md_dcmp)
1203 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1205 return EINVAL; /* too bad you can't distinguish this from a valid result */
1208 /** Allocate a single page.
1209 * Re-use old malloc'd pages first, otherwise just malloc.
1212 mdb_page_malloc(MDB_cursor *mc) {
1214 size_t sz = mc->mc_txn->mt_env->me_psize;
1215 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1216 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1217 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1218 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1219 } else if ((ret = malloc(sz)) != NULL) {
1220 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1225 /** Allocate pages for writing.
1226 * If there are free pages available from older transactions, they
1227 * will be re-used first. Otherwise a new page will be allocated.
1228 * @param[in] mc cursor A cursor handle identifying the transaction and
1229 * database for which we are allocating.
1230 * @param[in] num the number of pages to allocate.
1231 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1232 * will always be satisfied by a single contiguous chunk of memory.
1233 * @return 0 on success, non-zero on failure.
1236 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1238 MDB_txn *txn = mc->mc_txn;
1240 pgno_t pgno = P_INVALID;
1245 /* The free list won't have any content at all until txn 2 has
1246 * committed. The pages freed by txn 2 will be unreferenced
1247 * after txn 3 commits, and so will be safe to re-use in txn 4.
1249 if (txn->mt_txnid > 3) {
1251 if (!txn->mt_env->me_pghead &&
1252 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1253 /* See if there's anything in the free DB */
1259 txnid_t *kptr, last;
1261 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1262 if (!txn->mt_env->me_pgfirst) {
1263 mdb_page_search(&m2, NULL, 0);
1264 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1265 kptr = (txnid_t *)NODEKEY(leaf);
1272 last = txn->mt_env->me_pglast + 1;
1274 key.mv_data = &last;
1275 key.mv_size = sizeof(last);
1276 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1279 last = *(txnid_t *)key.mv_data;
1282 /* Unusable if referred by a meta page or reader... */
1284 if (last < txn->mt_txnid-1) {
1285 j = txn->mt_env->me_txns->mti_numreaders;
1286 r = txn->mt_env->me_txns->mti_readers + j;
1287 for (j = -j; j && (last<r[j].mr_txnid || !r[j].mr_pid); j++) ;
1291 /* It's usable, grab it.
1296 if (!txn->mt_env->me_pgfirst) {
1297 mdb_node_read(txn, leaf, &data);
1299 txn->mt_env->me_pglast = last;
1300 if (!txn->mt_env->me_pgfirst)
1301 txn->mt_env->me_pgfirst = last;
1302 idl = (MDB_ID *) data.mv_data;
1303 /* We might have a zero-length IDL due to freelist growth
1304 * during a prior commit
1306 if (!idl[0]) goto again;
1307 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1310 mop->mo_next = txn->mt_env->me_pghead;
1311 mop->mo_txnid = last;
1312 txn->mt_env->me_pghead = mop;
1313 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1318 DPRINTF("IDL read txn %zu root %zu num %zu",
1319 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1320 for (i=0; i<idl[0]; i++) {
1321 DPRINTF("IDL %zu", idl[i+1]);
1328 if (txn->mt_env->me_pghead) {
1329 MDB_oldpages *mop = txn->mt_env->me_pghead;
1331 /* FIXME: For now, always use fresh pages. We
1332 * really ought to search the free list for a
1337 /* peel pages off tail, so we only have to truncate the list */
1338 pgno = MDB_IDL_LAST(mop->mo_pages);
1339 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1341 if (mop->mo_pages[2] > mop->mo_pages[1])
1342 mop->mo_pages[0] = 0;
1346 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1347 txn->mt_env->me_pghead = mop->mo_next;
1348 if (mc->mc_dbi == FREE_DBI) {
1349 mop->mo_next = txn->mt_env->me_pgfree;
1350 txn->mt_env->me_pgfree = mop;
1359 if (pgno == P_INVALID) {
1360 /* DB size is maxed out */
1361 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1362 DPUTS("DB size maxed out");
1363 return MDB_MAP_FULL;
1366 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1367 if (pgno == P_INVALID) {
1368 pgno = txn->mt_next_pgno;
1369 txn->mt_next_pgno += num;
1371 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1374 if (txn->mt_env->me_dpages && num == 1) {
1375 np = txn->mt_env->me_dpages;
1376 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1377 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1378 txn->mt_env->me_dpages = np->mp_next;
1380 size_t sz = txn->mt_env->me_psize * num;
1381 if ((np = malloc(sz)) == NULL)
1383 VGMEMP_ALLOC(txn->mt_env, np, sz);
1385 if (pgno == P_INVALID) {
1386 np->mp_pgno = txn->mt_next_pgno;
1387 txn->mt_next_pgno += num;
1392 mid.mid = np->mp_pgno;
1394 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1395 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1397 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1404 /** Copy a page: avoid copying unused portions of the page.
1405 * @param[in] dst page to copy into
1406 * @param[in] src page to copy from
1409 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1411 dst->mp_flags = src->mp_flags | P_DIRTY;
1412 dst->mp_pages = src->mp_pages;
1414 if (IS_LEAF2(src)) {
1415 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1417 unsigned int i, nkeys = NUMKEYS(src);
1418 for (i=0; i<nkeys; i++)
1419 dst->mp_ptrs[i] = src->mp_ptrs[i];
1420 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1421 psize - src->mp_upper);
1425 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1426 * @param[in] mc cursor pointing to the page to be touched
1427 * @return 0 on success, non-zero on failure.
1430 mdb_page_touch(MDB_cursor *mc)
1432 MDB_page *mp = mc->mc_pg[mc->mc_top];
1436 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1438 if ((rc = mdb_page_alloc(mc, 1, &np)))
1440 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1441 assert(mp->mp_pgno != np->mp_pgno);
1442 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1444 /* If page isn't full, just copy the used portion */
1445 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1448 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1450 np->mp_flags |= P_DIRTY;
1455 /* Adjust other cursors pointing to mp */
1456 if (mc->mc_flags & C_SUB) {
1457 MDB_cursor *m2, *m3;
1458 MDB_dbi dbi = mc->mc_dbi-1;
1460 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1461 if (m2 == mc) continue;
1462 m3 = &m2->mc_xcursor->mx_cursor;
1463 if (m3->mc_snum < mc->mc_snum) continue;
1464 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1465 m3->mc_pg[mc->mc_top] = mp;
1471 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1472 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1473 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1474 m2->mc_pg[mc->mc_top] = mp;
1478 mc->mc_pg[mc->mc_top] = mp;
1479 /** If this page has a parent, update the parent to point to
1483 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1485 mc->mc_db->md_root = mp->mp_pgno;
1486 } else if (mc->mc_txn->mt_parent) {
1489 /* If txn has a parent, make sure the page is in our
1492 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1493 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1494 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1495 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1496 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1497 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1498 mc->mc_pg[mc->mc_top] = mp;
1504 np = mdb_page_malloc(mc);
1507 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1508 mid.mid = np->mp_pgno;
1510 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1518 mdb_env_sync(MDB_env *env, int force)
1521 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1522 if (env->me_flags & MDB_WRITEMAP) {
1523 int flags = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
1524 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1527 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1531 if (MDB_FDATASYNC(env->me_fd))
1538 /** Make shadow copies of all of parent txn's cursors */
1540 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1542 MDB_cursor *mc, *m2;
1543 unsigned int i, j, size;
1545 for (i=0;i<src->mt_numdbs; i++) {
1546 if (src->mt_cursors[i]) {
1547 size = sizeof(MDB_cursor);
1548 if (src->mt_cursors[i]->mc_xcursor)
1549 size += sizeof(MDB_xcursor);
1550 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1557 mc->mc_db = &dst->mt_dbs[i];
1558 mc->mc_dbx = m2->mc_dbx;
1559 mc->mc_dbflag = &dst->mt_dbflags[i];
1560 mc->mc_snum = m2->mc_snum;
1561 mc->mc_top = m2->mc_top;
1562 mc->mc_flags = m2->mc_flags | C_SHADOW;
1563 for (j=0; j<mc->mc_snum; j++) {
1564 mc->mc_pg[j] = m2->mc_pg[j];
1565 mc->mc_ki[j] = m2->mc_ki[j];
1567 if (m2->mc_xcursor) {
1568 MDB_xcursor *mx, *mx2;
1569 mx = (MDB_xcursor *)(mc+1);
1570 mc->mc_xcursor = mx;
1571 mx2 = m2->mc_xcursor;
1572 mx->mx_db = mx2->mx_db;
1573 mx->mx_dbx = mx2->mx_dbx;
1574 mx->mx_dbflag = mx2->mx_dbflag;
1575 mx->mx_cursor.mc_txn = dst;
1576 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1577 mx->mx_cursor.mc_db = &mx->mx_db;
1578 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1579 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1580 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1581 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1582 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1583 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1584 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1585 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1588 mc->mc_xcursor = NULL;
1590 mc->mc_next = dst->mt_cursors[i];
1591 dst->mt_cursors[i] = mc;
1598 /** Merge shadow cursors back into parent's */
1600 mdb_cursor_merge(MDB_txn *txn)
1603 for (i=0; i<txn->mt_numdbs; i++) {
1604 if (txn->mt_cursors[i]) {
1606 while ((mc = txn->mt_cursors[i])) {
1607 txn->mt_cursors[i] = mc->mc_next;
1608 if (mc->mc_flags & C_SHADOW) {
1609 MDB_cursor *m2 = mc->mc_orig;
1611 m2->mc_snum = mc->mc_snum;
1612 m2->mc_top = mc->mc_top;
1613 for (j=0; j<mc->mc_snum; j++) {
1614 m2->mc_pg[j] = mc->mc_pg[j];
1615 m2->mc_ki[j] = mc->mc_ki[j];
1618 if (mc->mc_flags & C_ALLOCD)
1626 mdb_txn_reset0(MDB_txn *txn);
1628 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1629 * @param[in] txn the transaction handle to initialize
1630 * @return 0 on success, non-zero on failure. This can only
1631 * fail for read-only transactions, and then only if the
1632 * reader table is full.
1635 mdb_txn_renew0(MDB_txn *txn)
1637 MDB_env *env = txn->mt_env;
1641 txn->mt_numdbs = env->me_numdbs;
1642 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1644 if (txn->mt_flags & MDB_TXN_RDONLY) {
1645 if (env->me_flags & MDB_ROFS) {
1646 i = mdb_env_pick_meta(env);
1647 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1648 txn->mt_u.reader = NULL;
1650 MDB_reader *r = pthread_getspecific(env->me_txkey);
1652 pid_t pid = env->me_pid;
1653 pthread_t tid = pthread_self();
1656 for (i=0; i<env->me_txns->mti_numreaders; i++)
1657 if (env->me_txns->mti_readers[i].mr_pid == 0)
1659 if (i == env->me_maxreaders) {
1660 UNLOCK_MUTEX_R(env);
1661 return MDB_READERS_FULL;
1663 env->me_txns->mti_readers[i].mr_pid = pid;
1664 env->me_txns->mti_readers[i].mr_tid = tid;
1665 if (i >= env->me_txns->mti_numreaders)
1666 env->me_txns->mti_numreaders = i+1;
1667 /* Save numreaders for un-mutexed mdb_env_close() */
1668 env->me_numreaders = env->me_txns->mti_numreaders;
1669 UNLOCK_MUTEX_R(env);
1670 r = &env->me_txns->mti_readers[i];
1671 pthread_setspecific(env->me_txkey, r);
1673 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1674 txn->mt_u.reader = r;
1676 txn->mt_toggle = txn->mt_txnid & 1;
1677 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1681 txn->mt_txnid = env->me_txns->mti_txnid;
1682 txn->mt_toggle = txn->mt_txnid & 1;
1683 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1686 if (txn->mt_txnid == mdb_debug_start)
1689 txn->mt_u.dirty_list = env->me_dirty_list;
1690 txn->mt_u.dirty_list[0].mid = 0;
1691 txn->mt_free_pgs = env->me_free_pgs;
1692 txn->mt_free_pgs[0] = 0;
1696 /* Copy the DB info and flags */
1697 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1698 for (i=2; i<txn->mt_numdbs; i++)
1699 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1700 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1701 if (txn->mt_numdbs > 2)
1702 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1708 mdb_txn_renew(MDB_txn *txn)
1712 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1715 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1716 DPUTS("environment had fatal error, must shutdown!");
1720 rc = mdb_txn_renew0(txn);
1721 if (rc == MDB_SUCCESS) {
1722 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1723 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1724 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1730 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1735 if (env->me_flags & MDB_FATAL_ERROR) {
1736 DPUTS("environment had fatal error, must shutdown!");
1739 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1742 /* Nested transactions: Max 1 child, write txns only, no writemap */
1743 if (parent->mt_child ||
1744 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1745 (env->me_flags & MDB_WRITEMAP))
1750 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1751 if (!(flags & MDB_RDONLY))
1752 size += env->me_maxdbs * sizeof(MDB_cursor *);
1754 if ((txn = calloc(1, size)) == NULL) {
1755 DPRINTF("calloc: %s", strerror(ErrCode()));
1758 txn->mt_dbs = (MDB_db *)(txn+1);
1759 if (flags & MDB_RDONLY) {
1760 txn->mt_flags |= MDB_TXN_RDONLY;
1761 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1763 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1764 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1769 txn->mt_free_pgs = mdb_midl_alloc();
1770 if (!txn->mt_free_pgs) {
1774 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1775 if (!txn->mt_u.dirty_list) {
1776 free(txn->mt_free_pgs);
1780 txn->mt_txnid = parent->mt_txnid;
1781 txn->mt_toggle = parent->mt_toggle;
1782 txn->mt_u.dirty_list[0].mid = 0;
1783 txn->mt_free_pgs[0] = 0;
1784 txn->mt_next_pgno = parent->mt_next_pgno;
1785 parent->mt_child = txn;
1786 txn->mt_parent = parent;
1787 txn->mt_numdbs = parent->mt_numdbs;
1788 txn->mt_dbxs = parent->mt_dbxs;
1789 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1790 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1791 mdb_cursor_shadow(parent, txn);
1794 rc = mdb_txn_renew0(txn);
1800 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1801 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1802 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1808 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1809 * @param[in] txn the transaction handle to reset
1812 mdb_txn_reset0(MDB_txn *txn)
1814 MDB_env *env = txn->mt_env;
1816 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1817 if (!(env->me_flags & MDB_ROFS))
1818 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1824 /* close(free) all cursors */
1825 for (i=0; i<txn->mt_numdbs; i++) {
1826 if (txn->mt_cursors[i]) {
1828 while ((mc = txn->mt_cursors[i])) {
1829 txn->mt_cursors[i] = mc->mc_next;
1830 if (mc->mc_flags & C_ALLOCD)
1836 if (!(env->me_flags & MDB_WRITEMAP)) {
1837 /* return all dirty pages to dpage list */
1838 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1839 dp = txn->mt_u.dirty_list[i].mptr;
1840 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1841 dp->mp_next = txn->mt_env->me_dpages;
1842 VGMEMP_FREE(txn->mt_env, dp);
1843 txn->mt_env->me_dpages = dp;
1845 /* large pages just get freed directly */
1846 VGMEMP_FREE(txn->mt_env, dp);
1852 if (txn->mt_parent) {
1853 txn->mt_parent->mt_child = NULL;
1854 mdb_midl_free(txn->mt_free_pgs);
1855 free(txn->mt_u.dirty_list);
1858 if (mdb_midl_shrink(&txn->mt_free_pgs))
1859 env->me_free_pgs = txn->mt_free_pgs;
1862 while ((mop = txn->mt_env->me_pghead)) {
1863 txn->mt_env->me_pghead = mop->mo_next;
1866 txn->mt_env->me_pgfirst = 0;
1867 txn->mt_env->me_pglast = 0;
1870 /* The writer mutex was locked in mdb_txn_begin. */
1871 UNLOCK_MUTEX_W(env);
1876 mdb_txn_reset(MDB_txn *txn)
1881 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1882 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1883 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1885 mdb_txn_reset0(txn);
1889 mdb_txn_abort(MDB_txn *txn)
1894 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1895 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1896 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1899 mdb_txn_abort(txn->mt_child);
1901 mdb_txn_reset0(txn);
1906 mdb_txn_commit(MDB_txn *txn)
1914 pgno_t next, freecnt;
1917 assert(txn != NULL);
1918 assert(txn->mt_env != NULL);
1920 if (txn->mt_child) {
1921 mdb_txn_commit(txn->mt_child);
1922 txn->mt_child = NULL;
1927 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1928 if (txn->mt_numdbs > env->me_numdbs) {
1929 /* update the DB flags */
1931 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
1932 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
1939 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1940 DPUTS("error flag is set, can't commit");
1942 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1947 /* Merge (and close) our cursors with parent's */
1948 mdb_cursor_merge(txn);
1950 if (txn->mt_parent) {
1956 /* Update parent's DB table */
1957 ip = &txn->mt_parent->mt_dbs[2];
1958 jp = &txn->mt_dbs[2];
1959 for (i = 2; i < txn->mt_numdbs; i++) {
1960 if (ip->md_root != jp->md_root)
1964 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1966 /* Append our free list to parent's */
1967 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1969 mdb_midl_free(txn->mt_free_pgs);
1971 /* Merge our dirty list with parent's */
1972 dst = txn->mt_parent->mt_u.dirty_list;
1973 src = txn->mt_u.dirty_list;
1974 x = mdb_mid2l_search(dst, src[1].mid);
1975 for (y=1; y<=src[0].mid; y++) {
1976 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1980 dst[x].mptr = src[y].mptr;
1983 for (; y<=src[0].mid; y++) {
1984 if (++x >= MDB_IDL_UM_MAX) {
1986 return MDB_TXN_FULL;
1991 free(txn->mt_u.dirty_list);
1992 txn->mt_parent->mt_child = NULL;
1997 if (txn != env->me_txn) {
1998 DPUTS("attempt to commit unknown transaction");
2003 if (!txn->mt_u.dirty_list[0].mid)
2006 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2007 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2009 /* Update DB root pointers. Their pages have already been
2010 * touched so this is all in-place and cannot fail.
2012 if (txn->mt_numdbs > 2) {
2015 data.mv_size = sizeof(MDB_db);
2017 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2018 for (i = 2; i < txn->mt_numdbs; i++) {
2019 if (txn->mt_dbflags[i] & DB_DIRTY) {
2020 data.mv_data = &txn->mt_dbs[i];
2021 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2026 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2028 /* should only be one record now */
2029 if (env->me_pghead) {
2030 /* make sure first page of freeDB is touched and on freelist */
2031 mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2034 /* Delete IDLs we used from the free list */
2035 if (env->me_pgfirst) {
2040 key.mv_size = sizeof(cur);
2041 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
2044 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
2045 rc = mdb_cursor_del(&mc, 0);
2051 env->me_pgfirst = 0;
2055 /* save to free list */
2057 freecnt = txn->mt_free_pgs[0];
2058 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2061 /* make sure last page of freeDB is touched and on freelist */
2062 key.mv_size = MAXKEYSIZE+1;
2064 mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2066 mdb_midl_sort(txn->mt_free_pgs);
2070 MDB_IDL idl = txn->mt_free_pgs;
2071 DPRINTF("IDL write txn %zu root %zu num %zu",
2072 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2073 for (i=0; i<idl[0]; i++) {
2074 DPRINTF("IDL %zu", idl[i+1]);
2078 /* write to last page of freeDB */
2079 key.mv_size = sizeof(pgno_t);
2080 key.mv_data = &txn->mt_txnid;
2081 data.mv_data = txn->mt_free_pgs;
2082 /* The free list can still grow during this call,
2083 * despite the pre-emptive touches above. So check
2084 * and make sure the entire thing got written.
2087 freecnt = txn->mt_free_pgs[0];
2088 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2089 rc = mdb_cursor_put(&mc, &key, &data, 0);
2094 } while (freecnt != txn->mt_free_pgs[0]);
2096 /* should only be one record now */
2098 if (env->me_pghead) {
2104 mop = env->me_pghead;
2106 key.mv_size = sizeof(id);
2108 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2109 data.mv_data = mop->mo_pages;
2110 orig = mop->mo_pages[0];
2111 /* These steps may grow the freelist again
2112 * due to freed overflow pages...
2114 mdb_cursor_put(&mc, &key, &data, 0);
2115 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2116 /* could have been used again here */
2117 if (mop->mo_pages[0] != orig) {
2118 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2119 data.mv_data = mop->mo_pages;
2121 mdb_cursor_put(&mc, &key, &data, 0);
2123 env->me_pghead = NULL;
2126 /* was completely used up */
2127 mdb_cursor_del(&mc, 0);
2131 env->me_pgfirst = 0;
2135 while (env->me_pgfree) {
2136 MDB_oldpages *mop = env->me_pgfree;
2137 env->me_pgfree = mop->mo_next;
2141 /* Check for growth of freelist again */
2142 if (freecnt != txn->mt_free_pgs[0])
2145 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2146 if (mdb_midl_shrink(&txn->mt_free_pgs))
2147 env->me_free_pgs = txn->mt_free_pgs;
2154 if (env->me_flags & MDB_WRITEMAP) {
2155 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2156 dp = txn->mt_u.dirty_list[i].mptr;
2157 /* clear dirty flag */
2158 dp->mp_flags &= ~P_DIRTY;
2159 txn->mt_u.dirty_list[i].mid = 0;
2161 txn->mt_u.dirty_list[0].mid = 0;
2165 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2171 /* Windows actually supports scatter/gather I/O, but only on
2172 * unbuffered file handles. Since we're relying on the OS page
2173 * cache for all our data, that's self-defeating. So we just
2174 * write pages one at a time. We use the ov structure to set
2175 * the write offset, to at least save the overhead of a Seek
2179 memset(&ov, 0, sizeof(ov));
2180 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2182 dp = txn->mt_u.dirty_list[i].mptr;
2183 DPRINTF("committing page %zu", dp->mp_pgno);
2184 size = dp->mp_pgno * env->me_psize;
2185 ov.Offset = size & 0xffffffff;
2186 ov.OffsetHigh = size >> 16;
2187 ov.OffsetHigh >>= 16;
2188 /* clear dirty flag */
2189 dp->mp_flags &= ~P_DIRTY;
2190 wsize = env->me_psize;
2191 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2192 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2195 DPRINTF("WriteFile: %d", n);
2202 struct iovec iov[MDB_COMMIT_PAGES];
2206 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2207 dp = txn->mt_u.dirty_list[i].mptr;
2208 if (dp->mp_pgno != next) {
2210 rc = writev(env->me_fd, iov, n);
2214 DPUTS("short write, filesystem full?");
2216 DPRINTF("writev: %s", strerror(n));
2223 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2226 DPRINTF("committing page %zu", dp->mp_pgno);
2227 iov[n].iov_len = env->me_psize;
2228 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2229 iov[n].iov_base = (char *)dp;
2230 size += iov[n].iov_len;
2231 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2232 /* clear dirty flag */
2233 dp->mp_flags &= ~P_DIRTY;
2234 if (++n >= MDB_COMMIT_PAGES) {
2244 rc = writev(env->me_fd, iov, n);
2248 DPUTS("short write, filesystem full?");
2250 DPRINTF("writev: %s", strerror(n));
2257 /* Drop the dirty pages.
2259 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2260 dp = txn->mt_u.dirty_list[i].mptr;
2261 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2262 dp->mp_next = txn->mt_env->me_dpages;
2263 VGMEMP_FREE(txn->mt_env, dp);
2264 txn->mt_env->me_dpages = dp;
2266 VGMEMP_FREE(txn->mt_env, dp);
2269 txn->mt_u.dirty_list[i].mid = 0;
2271 txn->mt_u.dirty_list[0].mid = 0;
2274 if ((n = mdb_env_sync(env, 0)) != 0 ||
2275 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2282 if (txn->mt_numdbs > env->me_numdbs) {
2283 /* update the DB flags */
2285 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2286 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2290 UNLOCK_MUTEX_W(env);
2296 /** Read the environment parameters of a DB environment before
2297 * mapping it into memory.
2298 * @param[in] env the environment handle
2299 * @param[out] meta address of where to store the meta information
2300 * @return 0 on success, non-zero on failure.
2303 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2310 /* We don't know the page size yet, so use a minimum value.
2314 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2316 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2321 else if (rc != MDB_PAGESIZE) {
2325 DPRINTF("read: %s", strerror(err));
2329 p = (MDB_page *)&pbuf;
2331 if (!F_ISSET(p->mp_flags, P_META)) {
2332 DPRINTF("page %zu not a meta page", p->mp_pgno);
2337 if (m->mm_magic != MDB_MAGIC) {
2338 DPUTS("meta has invalid magic");
2342 if (m->mm_version != MDB_VERSION) {
2343 DPRINTF("database is version %u, expected version %u",
2344 m->mm_version, MDB_VERSION);
2345 return MDB_VERSION_MISMATCH;
2348 memcpy(meta, m, sizeof(*m));
2352 /** Write the environment parameters of a freshly created DB environment.
2353 * @param[in] env the environment handle
2354 * @param[out] meta address of where to store the meta information
2355 * @return 0 on success, non-zero on failure.
2358 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2365 DPUTS("writing new meta page");
2367 GET_PAGESIZE(psize);
2369 meta->mm_magic = MDB_MAGIC;
2370 meta->mm_version = MDB_VERSION;
2371 meta->mm_psize = psize;
2372 meta->mm_last_pg = 1;
2373 meta->mm_flags = env->me_flags & 0xffff;
2374 meta->mm_flags |= MDB_INTEGERKEY;
2375 meta->mm_dbs[0].md_root = P_INVALID;
2376 meta->mm_dbs[1].md_root = P_INVALID;
2378 p = calloc(2, psize);
2380 p->mp_flags = P_META;
2383 memcpy(m, meta, sizeof(*meta));
2385 q = (MDB_page *)((char *)p + psize);
2388 q->mp_flags = P_META;
2391 memcpy(m, meta, sizeof(*meta));
2396 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2397 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2400 rc = write(env->me_fd, p, psize * 2);
2401 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2407 /** Update the environment info to commit a transaction.
2408 * @param[in] txn the transaction that's being committed
2409 * @return 0 on success, non-zero on failure.
2412 mdb_env_write_meta(MDB_txn *txn)
2415 MDB_meta meta, metab;
2417 int rc, len, toggle;
2423 assert(txn != NULL);
2424 assert(txn->mt_env != NULL);
2426 toggle = !txn->mt_toggle;
2427 DPRINTF("writing meta page %d for root page %zu",
2428 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2432 if (env->me_flags & MDB_WRITEMAP) {
2433 MDB_meta *mp = env->me_metas[toggle];
2434 mp->mm_dbs[0] = txn->mt_dbs[0];
2435 mp->mm_dbs[1] = txn->mt_dbs[1];
2436 mp->mm_last_pg = txn->mt_next_pgno - 1;
2437 mp->mm_txnid = txn->mt_txnid;
2438 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2439 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2442 ptr += env->me_psize;
2443 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2450 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2451 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2453 ptr = (char *)&meta;
2454 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2455 len = sizeof(MDB_meta) - off;
2458 meta.mm_dbs[0] = txn->mt_dbs[0];
2459 meta.mm_dbs[1] = txn->mt_dbs[1];
2460 meta.mm_last_pg = txn->mt_next_pgno - 1;
2461 meta.mm_txnid = txn->mt_txnid;
2464 off += env->me_psize;
2467 /* Write to the SYNC fd */
2470 memset(&ov, 0, sizeof(ov));
2472 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2475 rc = pwrite(env->me_mfd, ptr, len, off);
2480 DPUTS("write failed, disk error?");
2481 /* On a failure, the pagecache still contains the new data.
2482 * Write some old data back, to prevent it from being used.
2483 * Use the non-SYNC fd; we know it will fail anyway.
2485 meta.mm_last_pg = metab.mm_last_pg;
2486 meta.mm_txnid = metab.mm_txnid;
2488 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2490 r2 = pwrite(env->me_fd, ptr, len, off);
2493 env->me_flags |= MDB_FATAL_ERROR;
2497 /* Memory ordering issues are irrelevant; since the entire writer
2498 * is wrapped by wmutex, all of these changes will become visible
2499 * after the wmutex is unlocked. Since the DB is multi-version,
2500 * readers will get consistent data regardless of how fresh or
2501 * how stale their view of these values is.
2503 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2508 /** Check both meta pages to see which one is newer.
2509 * @param[in] env the environment handle
2510 * @return meta toggle (0 or 1).
2513 mdb_env_pick_meta(const MDB_env *env)
2515 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2519 mdb_env_create(MDB_env **env)
2523 e = calloc(1, sizeof(MDB_env));
2527 e->me_free_pgs = mdb_midl_alloc();
2528 if (!e->me_free_pgs) {
2532 e->me_maxreaders = DEFAULT_READERS;
2534 e->me_fd = INVALID_HANDLE_VALUE;
2535 e->me_lfd = INVALID_HANDLE_VALUE;
2536 e->me_mfd = INVALID_HANDLE_VALUE;
2537 #ifdef MDB_USE_POSIX_SEM
2538 e->me_rmutex = SEM_FAILED;
2539 e->me_wmutex = SEM_FAILED;
2541 e->me_pid = getpid();
2542 VGMEMP_CREATE(e,0,0);
2548 mdb_env_set_mapsize(MDB_env *env, size_t size)
2552 env->me_mapsize = size;
2554 env->me_maxpg = env->me_mapsize / env->me_psize;
2559 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2563 env->me_maxdbs = dbs;
2568 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2570 if (env->me_map || readers < 1)
2572 env->me_maxreaders = readers;
2577 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2579 if (!env || !readers)
2581 *readers = env->me_maxreaders;
2585 /** Further setup required for opening an MDB environment
2588 mdb_env_open2(MDB_env *env)
2590 unsigned int flags = env->me_flags;
2591 int i, newenv = 0, prot;
2595 memset(&meta, 0, sizeof(meta));
2597 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2600 DPUTS("new mdbenv");
2604 if (!env->me_mapsize) {
2605 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2611 LONG sizelo, sizehi;
2612 sizelo = env->me_mapsize & 0xffffffff;
2613 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2615 /* Windows won't create mappings for zero length files.
2616 * Just allocate the maxsize right now.
2619 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2620 if (!SetEndOfFile(env->me_fd))
2622 SetFilePointer(env->me_fd, 0, NULL, 0);
2624 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2625 PAGE_READWRITE : PAGE_READONLY,
2626 sizehi, sizelo, NULL);
2629 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2630 FILE_MAP_WRITE : FILE_MAP_READ,
2631 0, 0, env->me_mapsize, meta.mm_address);
2638 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2641 if (flags & MDB_WRITEMAP) {
2643 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2646 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2648 if (env->me_map == MAP_FAILED) {
2655 meta.mm_mapsize = env->me_mapsize;
2656 if (flags & MDB_FIXEDMAP)
2657 meta.mm_address = env->me_map;
2658 i = mdb_env_init_meta(env, &meta);
2659 if (i != MDB_SUCCESS) {
2663 env->me_psize = meta.mm_psize;
2665 env->me_maxpg = env->me_mapsize / env->me_psize;
2667 p = (MDB_page *)env->me_map;
2668 env->me_metas[0] = METADATA(p);
2669 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2673 int toggle = mdb_env_pick_meta(env);
2674 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2676 DPRINTF("opened database version %u, pagesize %u",
2677 env->me_metas[0]->mm_version, env->me_psize);
2678 DPRINTF("using meta page %d", toggle);
2679 DPRINTF("depth: %u", db->md_depth);
2680 DPRINTF("entries: %zu", db->md_entries);
2681 DPRINTF("branch pages: %zu", db->md_branch_pages);
2682 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2683 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2684 DPRINTF("root: %zu", db->md_root);
2692 /** Release a reader thread's slot in the reader lock table.
2693 * This function is called automatically when a thread exits.
2694 * @param[in] ptr This points to the slot in the reader lock table.
2697 mdb_env_reader_dest(void *ptr)
2699 MDB_reader *reader = ptr;
2705 /** Junk for arranging thread-specific callbacks on Windows. This is
2706 * necessarily platform and compiler-specific. Windows supports up
2707 * to 1088 keys. Let's assume nobody opens more than 64 environments
2708 * in a single process, for now. They can override this if needed.
2710 #ifndef MAX_TLS_KEYS
2711 #define MAX_TLS_KEYS 64
2713 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2714 static int mdb_tls_nkeys;
2716 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2720 case DLL_PROCESS_ATTACH: break;
2721 case DLL_THREAD_ATTACH: break;
2722 case DLL_THREAD_DETACH:
2723 for (i=0; i<mdb_tls_nkeys; i++) {
2724 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2725 mdb_env_reader_dest(r);
2728 case DLL_PROCESS_DETACH: break;
2733 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2735 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2739 /* Force some symbol references.
2740 * _tls_used forces the linker to create the TLS directory if not already done
2741 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2743 #pragma comment(linker, "/INCLUDE:_tls_used")
2744 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2745 #pragma const_seg(".CRT$XLB")
2746 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2747 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2750 #pragma comment(linker, "/INCLUDE:__tls_used")
2751 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2752 #pragma data_seg(".CRT$XLB")
2753 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2755 #endif /* WIN 32/64 */
2756 #endif /* !__GNUC__ */
2759 /** Downgrade the exclusive lock on the region back to shared */
2761 mdb_env_share_locks(MDB_env *env, int *excl)
2763 int rc = 0, toggle = mdb_env_pick_meta(env);
2765 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2770 /* First acquire a shared lock. The Unlock will
2771 * then release the existing exclusive lock.
2773 memset(&ov, 0, sizeof(ov));
2774 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2775 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2780 struct flock lock_info;
2781 /* The shared lock replaces the existing lock */
2782 memset((void *)&lock_info, 0, sizeof(lock_info));
2783 lock_info.l_type = F_RDLCK;
2784 lock_info.l_whence = SEEK_SET;
2785 lock_info.l_start = 0;
2786 lock_info.l_len = 1;
2787 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2788 (rc = ErrCode()) == EINTR) ;
2789 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
2796 /** Try to get exlusive lock, otherwise shared.
2797 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
2800 mdb_env_excl_lock(MDB_env *env, int *excl)
2804 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2808 memset(&ov, 0, sizeof(ov));
2809 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2814 struct flock lock_info;
2815 memset((void *)&lock_info, 0, sizeof(lock_info));
2816 lock_info.l_type = F_WRLCK;
2817 lock_info.l_whence = SEEK_SET;
2818 lock_info.l_start = 0;
2819 lock_info.l_len = 1;
2820 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2821 (rc = ErrCode()) == EINTR) ;
2825 # ifdef MDB_USE_POSIX_SEM
2826 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
2829 lock_info.l_type = F_RDLCK;
2830 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
2831 (rc = ErrCode()) == EINTR) ;
2839 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
2841 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2843 * @(#) $Revision: 5.1 $
2844 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2845 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2847 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2851 * Please do not copyright this code. This code is in the public domain.
2853 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2854 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2855 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2856 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2857 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2858 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2859 * PERFORMANCE OF THIS SOFTWARE.
2862 * chongo <Landon Curt Noll> /\oo/\
2863 * http://www.isthe.com/chongo/
2865 * Share and Enjoy! :-)
2868 typedef unsigned long long mdb_hash_t;
2869 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2871 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2872 * @param[in] str string to hash
2873 * @param[in] hval initial value for hash
2874 * @return 64 bit hash
2876 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2877 * hval arg on the first call.
2880 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
2882 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
2883 unsigned char *end = s + val->mv_size;
2885 * FNV-1a hash each octet of the string
2888 /* xor the bottom with the current octet */
2889 hval ^= (mdb_hash_t)*s++;
2891 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2892 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2893 (hval << 7) + (hval << 8) + (hval << 40);
2895 /* return our new hash value */
2899 /** Hash the string and output the hash in hex.
2900 * @param[in] str string to hash
2901 * @param[out] hexbuf an array of 17 chars to hold the hash
2904 mdb_hash_hex(MDB_val *val, char *hexbuf)
2907 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
2908 for (i=0; i<8; i++) {
2909 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2915 /** Open and/or initialize the lock region for the environment.
2916 * @param[in] env The MDB environment.
2917 * @param[in] lpath The pathname of the file used for the lock region.
2918 * @param[in] mode The Unix permissions for the file, if we create it.
2919 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
2920 * @return 0 on success, non-zero on failure.
2923 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2931 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2932 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2933 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2935 if (rc == ERROR_WRITE_PROTECT && (env->me_flags & MDB_RDONLY)) {
2936 env->me_flags |= MDB_ROFS;
2941 /* Try to get exclusive lock. If we succeed, then
2942 * nobody is using the lock region and we should initialize it.
2944 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2945 size = GetFileSize(env->me_lfd, NULL);
2951 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2953 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
2954 env->me_flags |= MDB_ROFS;
2959 /* Lose record locks when exec*() */
2960 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
2961 fcntl(env->me_lfd, F_SETFD, fdflags);
2963 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
2964 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1) {
2966 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
2967 env->me_flags |= MDB_ROFS;
2974 /* Try to get exclusive lock. If we succeed, then
2975 * nobody is using the lock region and we should initialize it.
2977 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2979 size = lseek(env->me_lfd, 0, SEEK_END);
2981 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2982 if (size < rsize && *excl > 0) {
2984 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2985 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
2987 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
2991 size = rsize - sizeof(MDB_txninfo);
2992 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2997 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2999 if (!mh) goto fail_errno;
3000 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3002 if (!env->me_txns) goto fail_errno;
3004 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3006 if (m == MAP_FAILED) goto fail_errno;
3012 BY_HANDLE_FILE_INFORMATION stbuf;
3021 if (!mdb_sec_inited) {
3022 InitializeSecurityDescriptor(&mdb_null_sd,
3023 SECURITY_DESCRIPTOR_REVISION);
3024 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3025 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3026 mdb_all_sa.bInheritHandle = FALSE;
3027 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3030 GetFileInformationByHandle(env->me_lfd, &stbuf);
3031 idbuf.volume = stbuf.dwVolumeSerialNumber;
3032 idbuf.nhigh = stbuf.nFileIndexHigh;
3033 idbuf.nlow = stbuf.nFileIndexLow;
3034 val.mv_data = &idbuf;
3035 val.mv_size = sizeof(idbuf);
3036 mdb_hash_hex(&val, hexbuf);
3037 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3038 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3039 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3040 if (!env->me_rmutex) goto fail_errno;
3041 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3042 if (!env->me_wmutex) goto fail_errno;
3043 #elif defined(MDB_USE_POSIX_SEM)
3052 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3053 idbuf.dev = stbuf.st_dev;
3054 idbuf.ino = stbuf.st_ino;
3055 val.mv_data = &idbuf;
3056 val.mv_size = sizeof(idbuf);
3057 mdb_hash_hex(&val, hexbuf);
3058 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3059 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3060 /* Clean up after a previous run, if needed: Try to
3061 * remove both semaphores before doing anything else.
3063 sem_unlink(env->me_txns->mti_rmname);
3064 sem_unlink(env->me_txns->mti_wmname);
3065 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3066 O_CREAT|O_EXCL, mode, 1);
3067 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3068 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3069 O_CREAT|O_EXCL, mode, 1);
3070 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3071 #else /* MDB_USE_POSIX_SEM */
3072 pthread_mutexattr_t mattr;
3074 if ((rc = pthread_mutexattr_init(&mattr))
3075 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3076 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3077 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3079 pthread_mutexattr_destroy(&mattr);
3080 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3082 env->me_txns->mti_version = MDB_VERSION;
3083 env->me_txns->mti_magic = MDB_MAGIC;
3084 env->me_txns->mti_txnid = 0;
3085 env->me_txns->mti_numreaders = 0;
3088 if (env->me_txns->mti_magic != MDB_MAGIC) {
3089 DPUTS("lock region has invalid magic");
3093 if (env->me_txns->mti_version != MDB_VERSION) {
3094 DPRINTF("lock region is version %u, expected version %u",
3095 env->me_txns->mti_version, MDB_VERSION);
3096 rc = MDB_VERSION_MISMATCH;
3100 if (rc != EACCES && rc != EAGAIN) {
3104 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3105 if (!env->me_rmutex) goto fail_errno;
3106 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3107 if (!env->me_wmutex) goto fail_errno;
3108 #elif defined(MDB_USE_POSIX_SEM)
3109 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3110 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3111 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3112 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3123 /** The name of the lock file in the DB environment */
3124 #define LOCKNAME "/lock.mdb"
3125 /** The name of the data file in the DB environment */
3126 #define DATANAME "/data.mdb"
3127 /** The suffix of the lock file when no subdir is used */
3128 #define LOCKSUFF "-lock"
3131 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3133 int oflags, rc, len, excl;
3134 char *lpath, *dpath;
3136 if (env->me_fd != INVALID_HANDLE_VALUE)
3140 if (flags & MDB_NOSUBDIR) {
3141 rc = len + sizeof(LOCKSUFF) + len + 1;
3143 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3148 if (flags & MDB_NOSUBDIR) {
3149 dpath = lpath + len + sizeof(LOCKSUFF);
3150 sprintf(lpath, "%s" LOCKSUFF, path);
3151 strcpy(dpath, path);
3153 dpath = lpath + len + sizeof(LOCKNAME);
3154 sprintf(lpath, "%s" LOCKNAME, path);
3155 sprintf(dpath, "%s" DATANAME, path);
3158 /* silently ignore WRITEMAP if we're only getting read access */
3159 if (F_ISSET(flags, MDB_RDONLY|MDB_WRITEMAP))
3160 flags ^= MDB_WRITEMAP;
3161 env->me_flags = flags |= MDB_ENV_ACTIVE;
3163 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3168 if (F_ISSET(flags, MDB_RDONLY)) {
3169 oflags = GENERIC_READ;
3170 len = OPEN_EXISTING;
3172 oflags = GENERIC_READ|GENERIC_WRITE;
3175 mode = FILE_ATTRIBUTE_NORMAL;
3176 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3177 NULL, len, mode, NULL);
3179 if (F_ISSET(flags, MDB_RDONLY))
3182 oflags = O_RDWR | O_CREAT;
3184 env->me_fd = open(dpath, oflags, mode);
3186 if (env->me_fd == INVALID_HANDLE_VALUE) {
3191 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3192 if (flags & (MDB_RDONLY|MDB_NOSYNC|MDB_NOMETASYNC|MDB_WRITEMAP)) {
3193 env->me_mfd = env->me_fd;
3195 /* synchronous fd for meta writes */
3197 env->me_mfd = CreateFile(dpath, oflags,
3198 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3199 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3201 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3203 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3208 DPRINTF("opened dbenv %p", (void *) env);
3209 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3210 env->me_numdbs = 2; /* this notes that me_txkey was set */
3212 /* Windows TLS callbacks need help finding their TLS info. */
3213 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3214 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3221 rc = mdb_env_share_locks(env, &excl);
3225 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3226 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3227 env->me_path = strdup(path);
3228 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3234 mdb_env_close0(env, excl);
3240 /** Destroy resources from mdb_env_open() and clear our readers */
3242 mdb_env_close0(MDB_env *env, int excl)
3246 if (!(env->me_flags & MDB_ENV_ACTIVE))
3249 free(env->me_dbflags);
3253 if (env->me_numdbs) {
3254 pthread_key_delete(env->me_txkey);
3256 /* Delete our key from the global list */
3257 for (i=0; i<mdb_tls_nkeys; i++)
3258 if (mdb_tls_keys[i] == env->me_txkey) {
3259 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3267 munmap(env->me_map, env->me_mapsize);
3269 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3271 if (env->me_fd != INVALID_HANDLE_VALUE)
3274 pid_t pid = env->me_pid;
3275 /* Clearing readers is done in this function because
3276 * me_txkey with its destructor must be disabled first.
3278 for (i = env->me_numreaders; --i >= 0; )
3279 if (env->me_txns->mti_readers[i].mr_pid == pid)
3280 env->me_txns->mti_readers[i].mr_pid = 0;
3282 if (env->me_rmutex) {
3283 CloseHandle(env->me_rmutex);
3284 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3286 /* Windows automatically destroys the mutexes when
3287 * the last handle closes.
3289 #elif defined(MDB_USE_POSIX_SEM)
3290 if (env->me_rmutex != SEM_FAILED) {
3291 sem_close(env->me_rmutex);
3292 if (env->me_wmutex != SEM_FAILED)
3293 sem_close(env->me_wmutex);
3294 /* If we have the filelock: If we are the
3295 * only remaining user, clean up semaphores.
3298 mdb_env_excl_lock(env, &excl);
3300 sem_unlink(env->me_txns->mti_rmname);
3301 sem_unlink(env->me_txns->mti_wmname);
3305 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3307 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3311 env->me_flags &= ~MDB_ENV_ACTIVE;
3315 mdb_env_copy(MDB_env *env, const char *path)
3317 MDB_txn *txn = NULL;
3318 int rc, len, oflags;
3321 HANDLE newfd = INVALID_HANDLE_VALUE;
3323 if (env->me_flags & MDB_NOSUBDIR) {
3324 lpath = (char *)path;
3327 len += sizeof(DATANAME);
3328 lpath = malloc(len);
3331 sprintf(lpath, "%s" DATANAME, path);
3334 /* The destination path must exist, but the destination file must not.
3335 * We don't want the OS to cache the writes, since the source data is
3336 * already in the OS cache.
3339 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3340 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3342 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3348 if (!(env->me_flags & MDB_NOSUBDIR))
3350 if (newfd == INVALID_HANDLE_VALUE) {
3355 #ifdef F_NOCACHE /* __APPLE__ */
3356 rc = fcntl(newfd, F_NOCACHE, 1);
3363 /* Do the lock/unlock of the reader mutex before starting the
3364 * write txn. Otherwise other read txns could block writers.
3366 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3370 if (!(env->me_flags & MDB_ROFS)) {
3371 /* We must start the actual read txn after blocking writers */
3372 mdb_txn_reset0(txn);
3374 /* Temporarily block writers until we snapshot the meta pages */
3377 rc = mdb_txn_renew0(txn);
3379 UNLOCK_MUTEX_W(env);
3384 wsize = env->me_psize * 2;
3388 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3389 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3392 rc = write(newfd, env->me_map, wsize);
3393 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3395 if (! (env->me_flags & MDB_ROFS))
3396 UNLOCK_MUTEX_W(env);
3401 ptr = env->me_map + wsize;
3402 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3406 rc = WriteFile(newfd, ptr, wsize, &len, NULL);
3407 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3410 rc = write(newfd, ptr, wsize);
3411 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3416 if (newfd != INVALID_HANDLE_VALUE)
3423 mdb_env_close(MDB_env *env)
3430 VGMEMP_DESTROY(env);
3431 while ((dp = env->me_dpages) != NULL) {
3432 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3433 env->me_dpages = dp->mp_next;
3437 mdb_env_close0(env, 0);
3438 mdb_midl_free(env->me_free_pgs);
3442 /** Compare two items pointing at aligned size_t's */
3444 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3446 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3447 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3450 /** Compare two items pointing at aligned int's */
3452 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3454 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3455 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3458 /** Compare two items pointing at ints of unknown alignment.
3459 * Nodes and keys are guaranteed to be 2-byte aligned.
3462 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3464 #if BYTE_ORDER == LITTLE_ENDIAN
3465 unsigned short *u, *c;
3468 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3469 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3472 } while(!x && u > (unsigned short *)a->mv_data);
3475 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3479 /** Compare two items lexically */
3481 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3488 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3494 diff = memcmp(a->mv_data, b->mv_data, len);
3495 return diff ? diff : len_diff<0 ? -1 : len_diff;
3498 /** Compare two items in reverse byte order */
3500 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3502 const unsigned char *p1, *p2, *p1_lim;
3506 p1_lim = (const unsigned char *)a->mv_data;
3507 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3508 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3510 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3516 while (p1 > p1_lim) {
3517 diff = *--p1 - *--p2;
3521 return len_diff<0 ? -1 : len_diff;
3524 /** Search for key within a page, using binary search.
3525 * Returns the smallest entry larger or equal to the key.
3526 * If exactp is non-null, stores whether the found entry was an exact match
3527 * in *exactp (1 or 0).
3528 * Updates the cursor index with the index of the found entry.
3529 * If no entry larger or equal to the key is found, returns NULL.
3532 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3534 unsigned int i = 0, nkeys;
3537 MDB_page *mp = mc->mc_pg[mc->mc_top];
3538 MDB_node *node = NULL;
3543 nkeys = NUMKEYS(mp);
3548 COPY_PGNO(pgno, mp->mp_pgno);
3549 DPRINTF("searching %u keys in %s %spage %zu",
3550 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3557 low = IS_LEAF(mp) ? 0 : 1;
3559 cmp = mc->mc_dbx->md_cmp;
3561 /* Branch pages have no data, so if using integer keys,
3562 * alignment is guaranteed. Use faster mdb_cmp_int.
3564 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3565 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3572 nodekey.mv_size = mc->mc_db->md_pad;
3573 node = NODEPTR(mp, 0); /* fake */
3574 while (low <= high) {
3575 i = (low + high) >> 1;
3576 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3577 rc = cmp(key, &nodekey);
3578 DPRINTF("found leaf index %u [%s], rc = %i",
3579 i, DKEY(&nodekey), rc);
3588 while (low <= high) {
3589 i = (low + high) >> 1;
3591 node = NODEPTR(mp, i);
3592 nodekey.mv_size = NODEKSZ(node);
3593 nodekey.mv_data = NODEKEY(node);
3595 rc = cmp(key, &nodekey);
3598 DPRINTF("found leaf index %u [%s], rc = %i",
3599 i, DKEY(&nodekey), rc);
3601 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3602 i, DKEY(&nodekey), NODEPGNO(node), rc);
3613 if (rc > 0) { /* Found entry is less than the key. */
3614 i++; /* Skip to get the smallest entry larger than key. */
3616 node = NODEPTR(mp, i);
3619 *exactp = (rc == 0);
3620 /* store the key index */
3621 mc->mc_ki[mc->mc_top] = i;
3623 /* There is no entry larger or equal to the key. */
3626 /* nodeptr is fake for LEAF2 */
3632 mdb_cursor_adjust(MDB_cursor *mc, func)
3636 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3637 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3644 /** Pop a page off the top of the cursor's stack. */
3646 mdb_cursor_pop(MDB_cursor *mc)
3650 MDB_page *top = mc->mc_pg[mc->mc_top];
3656 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3657 mc->mc_dbi, (void *) mc);
3661 /** Push a page onto the top of the cursor's stack. */
3663 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3665 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3666 mc->mc_dbi, (void *) mc);
3668 if (mc->mc_snum >= CURSOR_STACK) {
3669 assert(mc->mc_snum < CURSOR_STACK);
3670 return MDB_CURSOR_FULL;
3673 mc->mc_top = mc->mc_snum++;
3674 mc->mc_pg[mc->mc_top] = mp;
3675 mc->mc_ki[mc->mc_top] = 0;
3680 /** Find the address of the page corresponding to a given page number.
3681 * @param[in] txn the transaction for this access.
3682 * @param[in] pgno the page number for the page to retrieve.
3683 * @param[out] ret address of a pointer where the page's address will be stored.
3684 * @return 0 on success, non-zero on failure.
3687 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3691 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
3692 if (pgno < txn->mt_next_pgno)
3693 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3696 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3698 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3699 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3700 p = txn->mt_u.dirty_list[x].mptr;
3704 if (pgno < txn->mt_next_pgno)
3705 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3710 DPRINTF("page %zu not found", pgno);
3713 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3716 /** Search for the page a given key should be in.
3717 * Pushes parent pages on the cursor stack. This function continues a
3718 * search on a cursor that has already been initialized. (Usually by
3719 * #mdb_page_search() but also by #mdb_node_move().)
3720 * @param[in,out] mc the cursor for this operation.
3721 * @param[in] key the key to search for. If NULL, search for the lowest
3722 * page. (This is used by #mdb_cursor_first().)
3723 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3724 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3725 * @return 0 on success, non-zero on failure.
3728 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3730 MDB_page *mp = mc->mc_pg[mc->mc_top];
3735 while (IS_BRANCH(mp)) {
3739 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3740 assert(NUMKEYS(mp) > 1);
3741 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3743 if (key == NULL) /* Initialize cursor to first page. */
3745 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3746 /* cursor to last page */
3750 node = mdb_node_search(mc, key, &exact);
3752 i = NUMKEYS(mp) - 1;
3754 i = mc->mc_ki[mc->mc_top];
3763 DPRINTF("following index %u for key [%s]",
3765 assert(i < NUMKEYS(mp));
3766 node = NODEPTR(mp, i);
3768 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3771 mc->mc_ki[mc->mc_top] = i;
3772 if ((rc = mdb_cursor_push(mc, mp)))
3776 if ((rc = mdb_page_touch(mc)) != 0)
3778 mp = mc->mc_pg[mc->mc_top];
3783 DPRINTF("internal error, index points to a %02X page!?",
3785 return MDB_CORRUPTED;
3788 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3789 key ? DKEY(key) : NULL);
3794 /** Search for the page a given key should be in.
3795 * Pushes parent pages on the cursor stack. This function just sets up
3796 * the search; it finds the root page for \b mc's database and sets this
3797 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3798 * called to complete the search.
3799 * @param[in,out] mc the cursor for this operation.
3800 * @param[in] key the key to search for. If NULL, search for the lowest
3801 * page. (This is used by #mdb_cursor_first().)
3802 * @param[in] modify If true, visited pages are updated with new page numbers.
3803 * @return 0 on success, non-zero on failure.
3806 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
3811 /* Make sure the txn is still viable, then find the root from
3812 * the txn's db table.
3814 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3815 DPUTS("transaction has failed, must abort");
3818 /* Make sure we're using an up-to-date root */
3819 if (mc->mc_dbi > MAIN_DBI) {
3820 if ((*mc->mc_dbflag & DB_STALE) ||
3821 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
3823 unsigned char dbflag = 0;
3824 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3825 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
3828 if (*mc->mc_dbflag & DB_STALE) {
3831 MDB_node *leaf = mdb_node_search(&mc2,
3832 &mc->mc_dbx->md_name, &exact);
3834 return MDB_NOTFOUND;
3835 mdb_node_read(mc->mc_txn, leaf, &data);
3836 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3838 if (flags & MDB_PS_MODIFY)
3840 *mc->mc_dbflag = dbflag;
3843 root = mc->mc_db->md_root;
3845 if (root == P_INVALID) { /* Tree is empty. */
3846 DPUTS("tree is empty");
3847 return MDB_NOTFOUND;
3852 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
3853 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3859 DPRINTF("db %u root page %zu has flags 0x%X",
3860 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3862 if (flags & MDB_PS_MODIFY) {
3863 if ((rc = mdb_page_touch(mc)))
3867 if (flags & MDB_PS_ROOTONLY)
3870 return mdb_page_search_root(mc, key, flags);
3873 /** Return the data associated with a given node.
3874 * @param[in] txn The transaction for this operation.
3875 * @param[in] leaf The node being read.
3876 * @param[out] data Updated to point to the node's data.
3877 * @return 0 on success, non-zero on failure.
3880 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3882 MDB_page *omp; /* overflow page */
3886 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3887 data->mv_size = NODEDSZ(leaf);
3888 data->mv_data = NODEDATA(leaf);
3892 /* Read overflow data.
3894 data->mv_size = NODEDSZ(leaf);
3895 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3896 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3897 DPRINTF("read overflow page %zu failed", pgno);
3900 data->mv_data = METADATA(omp);
3906 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3907 MDB_val *key, MDB_val *data)
3916 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3918 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3921 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3925 mdb_cursor_init(&mc, txn, dbi, &mx);
3926 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3929 /** Find a sibling for a page.
3930 * Replaces the page at the top of the cursor's stack with the
3931 * specified sibling, if one exists.
3932 * @param[in] mc The cursor for this operation.
3933 * @param[in] move_right Non-zero if the right sibling is requested,
3934 * otherwise the left sibling.
3935 * @return 0 on success, non-zero on failure.
3938 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3944 if (mc->mc_snum < 2) {
3945 return MDB_NOTFOUND; /* root has no siblings */
3949 DPRINTF("parent page is page %zu, index %u",
3950 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3952 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3953 : (mc->mc_ki[mc->mc_top] == 0)) {
3954 DPRINTF("no more keys left, moving to %s sibling",
3955 move_right ? "right" : "left");
3956 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3960 mc->mc_ki[mc->mc_top]++;
3962 mc->mc_ki[mc->mc_top]--;
3963 DPRINTF("just moving to %s index key %u",
3964 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3966 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3968 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3969 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3972 mdb_cursor_push(mc, mp);
3977 /** Move the cursor to the next data item. */
3979 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3985 if (mc->mc_flags & C_EOF) {
3986 return MDB_NOTFOUND;
3989 assert(mc->mc_flags & C_INITIALIZED);
3991 mp = mc->mc_pg[mc->mc_top];
3993 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3994 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3995 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3996 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3997 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3998 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4002 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4003 if (op == MDB_NEXT_DUP)
4004 return MDB_NOTFOUND;
4008 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4010 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4011 DPUTS("=====> move to next sibling page");
4012 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4013 mc->mc_flags |= C_EOF;
4014 mc->mc_flags &= ~C_INITIALIZED;
4015 return MDB_NOTFOUND;
4017 mp = mc->mc_pg[mc->mc_top];
4018 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4020 mc->mc_ki[mc->mc_top]++;
4022 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4023 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4026 key->mv_size = mc->mc_db->md_pad;
4027 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4031 assert(IS_LEAF(mp));
4032 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4034 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4035 mdb_xcursor_init1(mc, leaf);
4038 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4041 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4042 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4043 if (rc != MDB_SUCCESS)
4048 MDB_GET_KEY(leaf, key);
4052 /** Move the cursor to the previous data item. */
4054 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4060 assert(mc->mc_flags & C_INITIALIZED);
4062 mp = mc->mc_pg[mc->mc_top];
4064 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4065 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4066 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4067 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4068 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4069 if (op != MDB_PREV || rc == MDB_SUCCESS)
4072 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4073 if (op == MDB_PREV_DUP)
4074 return MDB_NOTFOUND;
4079 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4081 if (mc->mc_ki[mc->mc_top] == 0) {
4082 DPUTS("=====> move to prev sibling page");
4083 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4084 mc->mc_flags &= ~C_INITIALIZED;
4085 return MDB_NOTFOUND;
4087 mp = mc->mc_pg[mc->mc_top];
4088 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4089 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4091 mc->mc_ki[mc->mc_top]--;
4093 mc->mc_flags &= ~C_EOF;
4095 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4096 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4099 key->mv_size = mc->mc_db->md_pad;
4100 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4104 assert(IS_LEAF(mp));
4105 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4107 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4108 mdb_xcursor_init1(mc, leaf);
4111 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4114 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4115 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4116 if (rc != MDB_SUCCESS)
4121 MDB_GET_KEY(leaf, key);
4125 /** Set the cursor on a specific data item. */
4127 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4128 MDB_cursor_op op, int *exactp)
4132 MDB_node *leaf = NULL;
4137 assert(key->mv_size > 0);
4139 /* See if we're already on the right page */
4140 if (mc->mc_flags & C_INITIALIZED) {
4143 mp = mc->mc_pg[mc->mc_top];
4145 mc->mc_ki[mc->mc_top] = 0;
4146 return MDB_NOTFOUND;
4148 if (mp->mp_flags & P_LEAF2) {
4149 nodekey.mv_size = mc->mc_db->md_pad;
4150 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4152 leaf = NODEPTR(mp, 0);
4153 MDB_GET_KEY(leaf, &nodekey);
4155 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4157 /* Probably happens rarely, but first node on the page
4158 * was the one we wanted.
4160 mc->mc_ki[mc->mc_top] = 0;
4167 unsigned int nkeys = NUMKEYS(mp);
4169 if (mp->mp_flags & P_LEAF2) {
4170 nodekey.mv_data = LEAF2KEY(mp,
4171 nkeys-1, nodekey.mv_size);
4173 leaf = NODEPTR(mp, nkeys-1);
4174 MDB_GET_KEY(leaf, &nodekey);
4176 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4178 /* last node was the one we wanted */
4179 mc->mc_ki[mc->mc_top] = nkeys-1;
4185 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4186 /* This is definitely the right page, skip search_page */
4187 if (mp->mp_flags & P_LEAF2) {
4188 nodekey.mv_data = LEAF2KEY(mp,
4189 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4191 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4192 MDB_GET_KEY(leaf, &nodekey);
4194 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4196 /* current node was the one we wanted */
4206 /* If any parents have right-sibs, search.
4207 * Otherwise, there's nothing further.
4209 for (i=0; i<mc->mc_top; i++)
4211 NUMKEYS(mc->mc_pg[i])-1)
4213 if (i == mc->mc_top) {
4214 /* There are no other pages */
4215 mc->mc_ki[mc->mc_top] = nkeys;
4216 return MDB_NOTFOUND;
4220 /* There are no other pages */
4221 mc->mc_ki[mc->mc_top] = 0;
4222 return MDB_NOTFOUND;
4226 rc = mdb_page_search(mc, key, 0);
4227 if (rc != MDB_SUCCESS)
4230 mp = mc->mc_pg[mc->mc_top];
4231 assert(IS_LEAF(mp));
4234 leaf = mdb_node_search(mc, key, exactp);
4235 if (exactp != NULL && !*exactp) {
4236 /* MDB_SET specified and not an exact match. */
4237 return MDB_NOTFOUND;
4241 DPUTS("===> inexact leaf not found, goto sibling");
4242 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4243 return rc; /* no entries matched */
4244 mp = mc->mc_pg[mc->mc_top];
4245 assert(IS_LEAF(mp));
4246 leaf = NODEPTR(mp, 0);
4250 mc->mc_flags |= C_INITIALIZED;
4251 mc->mc_flags &= ~C_EOF;
4254 key->mv_size = mc->mc_db->md_pad;
4255 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4259 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4260 mdb_xcursor_init1(mc, leaf);
4263 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4264 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4265 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4268 if (op == MDB_GET_BOTH) {
4274 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4275 if (rc != MDB_SUCCESS)
4278 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4280 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4282 rc = mc->mc_dbx->md_dcmp(data, &d2);
4284 if (op == MDB_GET_BOTH || rc > 0)
4285 return MDB_NOTFOUND;
4290 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4291 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4296 /* The key already matches in all other cases */
4297 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4298 MDB_GET_KEY(leaf, key);
4299 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4304 /** Move the cursor to the first item in the database. */
4306 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4311 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4312 rc = mdb_page_search(mc, NULL, 0);
4313 if (rc != MDB_SUCCESS)
4316 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4318 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4319 mc->mc_flags |= C_INITIALIZED;
4320 mc->mc_flags &= ~C_EOF;
4322 mc->mc_ki[mc->mc_top] = 0;
4324 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4325 key->mv_size = mc->mc_db->md_pad;
4326 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4331 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4332 mdb_xcursor_init1(mc, leaf);
4333 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4338 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4339 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4343 MDB_GET_KEY(leaf, key);
4347 /** Move the cursor to the last item in the database. */
4349 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4354 if (!(mc->mc_flags & C_EOF)) {
4356 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4359 lkey.mv_size = MAXKEYSIZE+1;
4360 lkey.mv_data = NULL;
4361 rc = mdb_page_search(mc, &lkey, 0);
4362 if (rc != MDB_SUCCESS)
4365 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4367 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4368 mc->mc_flags |= C_INITIALIZED|C_EOF;
4370 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4372 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4373 key->mv_size = mc->mc_db->md_pad;
4374 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4379 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4380 mdb_xcursor_init1(mc, leaf);
4381 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4386 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4387 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4392 MDB_GET_KEY(leaf, key);
4397 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4406 case MDB_GET_CURRENT:
4407 if (!mc->mc_flags & C_INITIALIZED) {
4410 MDB_page *mp = mc->mc_pg[mc->mc_top];
4412 mc->mc_ki[mc->mc_top] = 0;
4418 key->mv_size = mc->mc_db->md_pad;
4419 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4421 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4422 MDB_GET_KEY(leaf, key);
4424 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4425 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4427 rc = mdb_node_read(mc->mc_txn, leaf, data);
4434 case MDB_GET_BOTH_RANGE:
4435 if (data == NULL || mc->mc_xcursor == NULL) {
4443 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4445 } else if (op == MDB_SET_RANGE)
4446 rc = mdb_cursor_set(mc, key, data, op, NULL);
4448 rc = mdb_cursor_set(mc, key, data, op, &exact);
4450 case MDB_GET_MULTIPLE:
4452 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4453 !(mc->mc_flags & C_INITIALIZED)) {
4458 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4459 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4462 case MDB_NEXT_MULTIPLE:
4464 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4468 if (!(mc->mc_flags & C_INITIALIZED))
4469 rc = mdb_cursor_first(mc, key, data);
4471 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4472 if (rc == MDB_SUCCESS) {
4473 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4476 mx = &mc->mc_xcursor->mx_cursor;
4477 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4479 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4480 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4488 case MDB_NEXT_NODUP:
4489 if (!(mc->mc_flags & C_INITIALIZED))
4490 rc = mdb_cursor_first(mc, key, data);
4492 rc = mdb_cursor_next(mc, key, data, op);
4496 case MDB_PREV_NODUP:
4497 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4498 rc = mdb_cursor_last(mc, key, data);
4499 mc->mc_flags &= ~C_EOF;
4501 rc = mdb_cursor_prev(mc, key, data, op);
4504 rc = mdb_cursor_first(mc, key, data);
4508 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4509 !(mc->mc_flags & C_INITIALIZED) ||
4510 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4514 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4517 rc = mdb_cursor_last(mc, key, data);
4521 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4522 !(mc->mc_flags & C_INITIALIZED) ||
4523 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4527 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4530 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4538 /** Touch all the pages in the cursor stack.
4539 * Makes sure all the pages are writable, before attempting a write operation.
4540 * @param[in] mc The cursor to operate on.
4543 mdb_cursor_touch(MDB_cursor *mc)
4547 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4549 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4550 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4553 *mc->mc_dbflag = DB_DIRTY;
4555 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4556 rc = mdb_page_touch(mc);
4560 mc->mc_top = mc->mc_snum-1;
4565 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4568 MDB_node *leaf = NULL;
4569 MDB_val xdata, *rdata, dkey;
4572 int do_sub = 0, insert = 0;
4573 unsigned int mcount = 0;
4577 char dbuf[MAXKEYSIZE+1];
4578 unsigned int nflags;
4581 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4584 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4585 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4589 if (flags == MDB_CURRENT) {
4590 if (!(mc->mc_flags & C_INITIALIZED))
4593 } else if (mc->mc_db->md_root == P_INVALID) {
4595 /* new database, write a root leaf page */
4596 DPUTS("allocating new root leaf page");
4597 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4601 mdb_cursor_push(mc, np);
4602 mc->mc_db->md_root = np->mp_pgno;
4603 mc->mc_db->md_depth++;
4604 *mc->mc_dbflag = DB_DIRTY;
4605 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4607 np->mp_flags |= P_LEAF2;
4608 mc->mc_flags |= C_INITIALIZED;
4614 if (flags & MDB_APPEND) {
4616 rc = mdb_cursor_last(mc, &k2, &d2);
4618 rc = mc->mc_dbx->md_cmp(key, &k2);
4621 mc->mc_ki[mc->mc_top]++;
4627 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4629 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4630 DPRINTF("duplicate key [%s]", DKEY(key));
4632 return MDB_KEYEXIST;
4634 if (rc && rc != MDB_NOTFOUND)
4638 /* Cursor is positioned, now make sure all pages are writable */
4639 rc2 = mdb_cursor_touch(mc);
4644 /* The key already exists */
4645 if (rc == MDB_SUCCESS) {
4646 /* there's only a key anyway, so this is a no-op */
4647 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4648 unsigned int ksize = mc->mc_db->md_pad;
4649 if (key->mv_size != ksize)
4651 if (flags == MDB_CURRENT) {
4652 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4653 memcpy(ptr, key->mv_data, ksize);
4658 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4661 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4662 /* Was a single item before, must convert now */
4664 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4665 /* Just overwrite the current item */
4666 if (flags == MDB_CURRENT)
4669 dkey.mv_size = NODEDSZ(leaf);
4670 dkey.mv_data = NODEDATA(leaf);
4671 #if UINT_MAX < SIZE_MAX
4672 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4673 #ifdef MISALIGNED_OK
4674 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4676 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4679 /* if data matches, ignore it */
4680 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4681 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4683 /* create a fake page for the dup items */
4684 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4685 dkey.mv_data = dbuf;
4686 fp = (MDB_page *)&pbuf;
4687 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4688 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4689 fp->mp_lower = PAGEHDRSZ;
4690 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4691 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4692 fp->mp_flags |= P_LEAF2;
4693 fp->mp_pad = data->mv_size;
4694 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4696 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4697 (dkey.mv_size & 1) + (data->mv_size & 1);
4699 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4702 xdata.mv_size = fp->mp_upper;
4707 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4708 /* See if we need to convert from fake page to subDB */
4710 unsigned int offset;
4713 fp = NODEDATA(leaf);
4714 if (flags == MDB_CURRENT) {
4716 fp->mp_flags |= P_DIRTY;
4717 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4718 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4722 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4723 offset = fp->mp_pad;
4724 if (SIZELEFT(fp) >= offset)
4726 offset *= 4; /* space for 4 more */
4728 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4730 offset += offset & 1;
4731 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4732 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4734 /* yes, convert it */
4736 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4737 dummy.md_pad = fp->mp_pad;
4738 dummy.md_flags = MDB_DUPFIXED;
4739 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4740 dummy.md_flags |= MDB_INTEGERKEY;
4743 dummy.md_branch_pages = 0;
4744 dummy.md_leaf_pages = 1;
4745 dummy.md_overflow_pages = 0;
4746 dummy.md_entries = NUMKEYS(fp);
4748 xdata.mv_size = sizeof(MDB_db);
4749 xdata.mv_data = &dummy;
4750 if ((rc = mdb_page_alloc(mc, 1, &mp)))
4752 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4753 flags |= F_DUPDATA|F_SUBDATA;
4754 dummy.md_root = mp->mp_pgno;
4756 /* no, just grow it */
4758 xdata.mv_size = NODEDSZ(leaf) + offset;
4759 xdata.mv_data = &pbuf;
4760 mp = (MDB_page *)&pbuf;
4761 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4764 mp->mp_flags = fp->mp_flags | P_DIRTY;
4765 mp->mp_pad = fp->mp_pad;
4766 mp->mp_lower = fp->mp_lower;
4767 mp->mp_upper = fp->mp_upper + offset;
4769 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4771 nsize = NODEDSZ(leaf) - fp->mp_upper;
4772 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4773 for (i=0; i<NUMKEYS(fp); i++)
4774 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4776 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4780 /* data is on sub-DB, just store it */
4781 flags |= F_DUPDATA|F_SUBDATA;
4785 /* overflow page overwrites need special handling */
4786 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4789 int ovpages, dpages;
4791 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4792 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4793 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4794 mdb_page_get(mc->mc_txn, pg, &omp);
4795 /* Is the ov page writable and large enough? */
4796 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4797 /* yes, overwrite it. Note in this case we don't
4798 * bother to try shrinking the node if the new data
4799 * is smaller than the overflow threshold.
4801 if (F_ISSET(flags, MDB_RESERVE))
4802 data->mv_data = METADATA(omp);
4804 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4807 /* no, free ovpages */
4809 mc->mc_db->md_overflow_pages -= ovpages;
4810 for (i=0; i<ovpages; i++) {
4811 DPRINTF("freed ov page %zu", pg);
4812 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4816 } else if (NODEDSZ(leaf) == data->mv_size) {
4817 /* same size, just replace it. Note that we could
4818 * also reuse this node if the new data is smaller,
4819 * but instead we opt to shrink the node in that case.
4821 if (F_ISSET(flags, MDB_RESERVE))
4822 data->mv_data = NODEDATA(leaf);
4824 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4827 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4828 mc->mc_db->md_entries--;
4830 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4837 nflags = flags & NODE_ADD_FLAGS;
4838 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4839 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4840 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4841 nflags &= ~MDB_APPEND;
4843 nflags |= MDB_SPLIT_REPLACE;
4844 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4846 /* There is room already in this leaf page. */
4847 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4848 if (rc == 0 && !do_sub && insert) {
4849 /* Adjust other cursors pointing to mp */
4850 MDB_cursor *m2, *m3;
4851 MDB_dbi dbi = mc->mc_dbi;
4852 unsigned i = mc->mc_top;
4853 MDB_page *mp = mc->mc_pg[i];
4855 if (mc->mc_flags & C_SUB)
4858 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4859 if (mc->mc_flags & C_SUB)
4860 m3 = &m2->mc_xcursor->mx_cursor;
4863 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4864 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4871 if (rc != MDB_SUCCESS)
4872 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4874 /* Now store the actual data in the child DB. Note that we're
4875 * storing the user data in the keys field, so there are strict
4876 * size limits on dupdata. The actual data fields of the child
4877 * DB are all zero size.
4884 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4885 if (flags & MDB_CURRENT) {
4886 xflags = MDB_CURRENT;
4888 mdb_xcursor_init1(mc, leaf);
4889 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4891 /* converted, write the original data first */
4893 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4897 /* Adjust other cursors pointing to mp */
4899 unsigned i = mc->mc_top;
4900 MDB_page *mp = mc->mc_pg[i];
4902 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4903 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4904 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4905 mdb_xcursor_init1(m2, leaf);
4910 if (flags & MDB_APPENDDUP)
4911 xflags |= MDB_APPEND;
4912 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4913 if (flags & F_SUBDATA) {
4914 void *db = NODEDATA(leaf);
4915 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4918 /* sub-writes might have failed so check rc again.
4919 * Don't increment count if we just replaced an existing item.
4921 if (!rc && !(flags & MDB_CURRENT))
4922 mc->mc_db->md_entries++;
4923 if (flags & MDB_MULTIPLE) {
4925 if (mcount < data[1].mv_size) {
4926 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4927 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4937 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4942 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4945 if (!mc->mc_flags & C_INITIALIZED)
4948 rc = mdb_cursor_touch(mc);
4952 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4954 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4955 if (flags != MDB_NODUPDATA) {
4956 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4957 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4959 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4960 /* If sub-DB still has entries, we're done */
4961 if (mc->mc_xcursor->mx_db.md_entries) {
4962 if (leaf->mn_flags & F_SUBDATA) {
4963 /* update subDB info */
4964 void *db = NODEDATA(leaf);
4965 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4967 /* shrink fake page */
4968 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4970 mc->mc_db->md_entries--;
4973 /* otherwise fall thru and delete the sub-DB */
4976 if (leaf->mn_flags & F_SUBDATA) {
4977 /* add all the child DB's pages to the free list */
4978 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4979 if (rc == MDB_SUCCESS) {
4980 mc->mc_db->md_entries -=
4981 mc->mc_xcursor->mx_db.md_entries;
4986 return mdb_cursor_del0(mc, leaf);
4989 /** Allocate and initialize new pages for a database.
4990 * @param[in] mc a cursor on the database being added to.
4991 * @param[in] flags flags defining what type of page is being allocated.
4992 * @param[in] num the number of pages to allocate. This is usually 1,
4993 * unless allocating overflow pages for a large record.
4994 * @param[out] mp Address of a page, or NULL on failure.
4995 * @return 0 on success, non-zero on failure.
4998 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5003 if ((rc = mdb_page_alloc(mc, num, &np)))
5005 DPRINTF("allocated new mpage %zu, page size %u",
5006 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5007 np->mp_flags = flags | P_DIRTY;
5008 np->mp_lower = PAGEHDRSZ;
5009 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5012 mc->mc_db->md_branch_pages++;
5013 else if (IS_LEAF(np))
5014 mc->mc_db->md_leaf_pages++;
5015 else if (IS_OVERFLOW(np)) {
5016 mc->mc_db->md_overflow_pages += num;
5024 /** Calculate the size of a leaf node.
5025 * The size depends on the environment's page size; if a data item
5026 * is too large it will be put onto an overflow page and the node
5027 * size will only include the key and not the data. Sizes are always
5028 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5029 * of the #MDB_node headers.
5030 * @param[in] env The environment handle.
5031 * @param[in] key The key for the node.
5032 * @param[in] data The data for the node.
5033 * @return The number of bytes needed to store the node.
5036 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5040 sz = LEAFSIZE(key, data);
5041 if (sz >= env->me_psize / MDB_MINKEYS) {
5042 /* put on overflow page */
5043 sz -= data->mv_size - sizeof(pgno_t);
5047 return sz + sizeof(indx_t);
5050 /** Calculate the size of a branch node.
5051 * The size should depend on the environment's page size but since
5052 * we currently don't support spilling large keys onto overflow
5053 * pages, it's simply the size of the #MDB_node header plus the
5054 * size of the key. Sizes are always rounded up to an even number
5055 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5056 * @param[in] env The environment handle.
5057 * @param[in] key The key for the node.
5058 * @return The number of bytes needed to store the node.
5061 mdb_branch_size(MDB_env *env, MDB_val *key)
5066 if (sz >= env->me_psize / MDB_MINKEYS) {
5067 /* put on overflow page */
5068 /* not implemented */
5069 /* sz -= key->size - sizeof(pgno_t); */
5072 return sz + sizeof(indx_t);
5075 /** Add a node to the page pointed to by the cursor.
5076 * @param[in] mc The cursor for this operation.
5077 * @param[in] indx The index on the page where the new node should be added.
5078 * @param[in] key The key for the new node.
5079 * @param[in] data The data for the new node, if any.
5080 * @param[in] pgno The page number, if adding a branch node.
5081 * @param[in] flags Flags for the node.
5082 * @return 0 on success, non-zero on failure. Possible errors are:
5084 * <li>ENOMEM - failed to allocate overflow pages for the node.
5085 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5086 * should never happen since all callers already calculate the
5087 * page's free space before calling this function.
5091 mdb_node_add(MDB_cursor *mc, indx_t indx,
5092 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5095 size_t node_size = NODESIZE;
5098 MDB_page *mp = mc->mc_pg[mc->mc_top];
5099 MDB_page *ofp = NULL; /* overflow page */
5102 assert(mp->mp_upper >= mp->mp_lower);
5104 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5105 IS_LEAF(mp) ? "leaf" : "branch",
5106 IS_SUBP(mp) ? "sub-" : "",
5107 mp->mp_pgno, indx, data ? data->mv_size : 0,
5108 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5111 /* Move higher keys up one slot. */
5112 int ksize = mc->mc_db->md_pad, dif;
5113 char *ptr = LEAF2KEY(mp, indx, ksize);
5114 dif = NUMKEYS(mp) - indx;
5116 memmove(ptr+ksize, ptr, dif*ksize);
5117 /* insert new key */
5118 memcpy(ptr, key->mv_data, ksize);
5120 /* Just using these for counting */
5121 mp->mp_lower += sizeof(indx_t);
5122 mp->mp_upper -= ksize - sizeof(indx_t);
5127 node_size += key->mv_size;
5131 if (F_ISSET(flags, F_BIGDATA)) {
5132 /* Data already on overflow page. */
5133 node_size += sizeof(pgno_t);
5134 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5135 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5137 /* Put data on overflow page. */
5138 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5139 data->mv_size, node_size+data->mv_size);
5140 node_size += sizeof(pgno_t);
5141 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5143 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5146 node_size += data->mv_size;
5149 node_size += node_size & 1;
5151 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5152 DPRINTF("not enough room in page %zu, got %u ptrs",
5153 mp->mp_pgno, NUMKEYS(mp));
5154 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5155 mp->mp_upper - mp->mp_lower);
5156 DPRINTF("node size = %zu", node_size);
5157 return MDB_PAGE_FULL;
5160 /* Move higher pointers up one slot. */
5161 for (i = NUMKEYS(mp); i > indx; i--)
5162 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5164 /* Adjust free space offsets. */
5165 ofs = mp->mp_upper - node_size;
5166 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5167 mp->mp_ptrs[indx] = ofs;
5169 mp->mp_lower += sizeof(indx_t);
5171 /* Write the node data. */
5172 node = NODEPTR(mp, indx);
5173 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5174 node->mn_flags = flags;
5176 SETDSZ(node,data->mv_size);
5181 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5186 if (F_ISSET(flags, F_BIGDATA))
5187 memcpy(node->mn_data + key->mv_size, data->mv_data,
5189 else if (F_ISSET(flags, MDB_RESERVE))
5190 data->mv_data = node->mn_data + key->mv_size;
5192 memcpy(node->mn_data + key->mv_size, data->mv_data,
5195 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5197 if (F_ISSET(flags, MDB_RESERVE))
5198 data->mv_data = METADATA(ofp);
5200 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5207 /** Delete the specified node from a page.
5208 * @param[in] mp The page to operate on.
5209 * @param[in] indx The index of the node to delete.
5210 * @param[in] ksize The size of a node. Only used if the page is
5211 * part of a #MDB_DUPFIXED database.
5214 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5217 indx_t i, j, numkeys, ptr;
5224 COPY_PGNO(pgno, mp->mp_pgno);
5225 DPRINTF("delete node %u on %s page %zu", indx,
5226 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5229 assert(indx < NUMKEYS(mp));
5232 int x = NUMKEYS(mp) - 1 - indx;
5233 base = LEAF2KEY(mp, indx, ksize);
5235 memmove(base, base + ksize, x * ksize);
5236 mp->mp_lower -= sizeof(indx_t);
5237 mp->mp_upper += ksize - sizeof(indx_t);
5241 node = NODEPTR(mp, indx);
5242 sz = NODESIZE + node->mn_ksize;
5244 if (F_ISSET(node->mn_flags, F_BIGDATA))
5245 sz += sizeof(pgno_t);
5247 sz += NODEDSZ(node);
5251 ptr = mp->mp_ptrs[indx];
5252 numkeys = NUMKEYS(mp);
5253 for (i = j = 0; i < numkeys; i++) {
5255 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5256 if (mp->mp_ptrs[i] < ptr)
5257 mp->mp_ptrs[j] += sz;
5262 base = (char *)mp + mp->mp_upper;
5263 memmove(base + sz, base, ptr - mp->mp_upper);
5265 mp->mp_lower -= sizeof(indx_t);
5269 /** Compact the main page after deleting a node on a subpage.
5270 * @param[in] mp The main page to operate on.
5271 * @param[in] indx The index of the subpage on the main page.
5274 mdb_node_shrink(MDB_page *mp, indx_t indx)
5281 indx_t i, numkeys, ptr;
5283 node = NODEPTR(mp, indx);
5284 sp = (MDB_page *)NODEDATA(node);
5285 osize = NODEDSZ(node);
5287 delta = sp->mp_upper - sp->mp_lower;
5288 SETDSZ(node, osize - delta);
5289 xp = (MDB_page *)((char *)sp + delta);
5291 /* shift subpage upward */
5293 nsize = NUMKEYS(sp) * sp->mp_pad;
5294 memmove(METADATA(xp), METADATA(sp), nsize);
5297 nsize = osize - sp->mp_upper;
5298 numkeys = NUMKEYS(sp);
5299 for (i=numkeys-1; i>=0; i--)
5300 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5302 xp->mp_upper = sp->mp_lower;
5303 xp->mp_lower = sp->mp_lower;
5304 xp->mp_flags = sp->mp_flags;
5305 xp->mp_pad = sp->mp_pad;
5306 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5308 /* shift lower nodes upward */
5309 ptr = mp->mp_ptrs[indx];
5310 numkeys = NUMKEYS(mp);
5311 for (i = 0; i < numkeys; i++) {
5312 if (mp->mp_ptrs[i] <= ptr)
5313 mp->mp_ptrs[i] += delta;
5316 base = (char *)mp + mp->mp_upper;
5317 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5318 mp->mp_upper += delta;
5321 /** Initial setup of a sorted-dups cursor.
5322 * Sorted duplicates are implemented as a sub-database for the given key.
5323 * The duplicate data items are actually keys of the sub-database.
5324 * Operations on the duplicate data items are performed using a sub-cursor
5325 * initialized when the sub-database is first accessed. This function does
5326 * the preliminary setup of the sub-cursor, filling in the fields that
5327 * depend only on the parent DB.
5328 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5331 mdb_xcursor_init0(MDB_cursor *mc)
5333 MDB_xcursor *mx = mc->mc_xcursor;
5335 mx->mx_cursor.mc_xcursor = NULL;
5336 mx->mx_cursor.mc_txn = mc->mc_txn;
5337 mx->mx_cursor.mc_db = &mx->mx_db;
5338 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5339 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5340 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5341 mx->mx_cursor.mc_snum = 0;
5342 mx->mx_cursor.mc_top = 0;
5343 mx->mx_cursor.mc_flags = C_SUB;
5344 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5345 mx->mx_dbx.md_dcmp = NULL;
5346 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5349 /** Final setup of a sorted-dups cursor.
5350 * Sets up the fields that depend on the data from the main cursor.
5351 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5352 * @param[in] node The data containing the #MDB_db record for the
5353 * sorted-dup database.
5356 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5358 MDB_xcursor *mx = mc->mc_xcursor;
5360 if (node->mn_flags & F_SUBDATA) {
5361 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5362 mx->mx_cursor.mc_pg[0] = 0;
5363 mx->mx_cursor.mc_snum = 0;
5364 mx->mx_cursor.mc_flags = C_SUB;
5366 MDB_page *fp = NODEDATA(node);
5367 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5368 mx->mx_db.md_flags = 0;
5369 mx->mx_db.md_depth = 1;
5370 mx->mx_db.md_branch_pages = 0;
5371 mx->mx_db.md_leaf_pages = 1;
5372 mx->mx_db.md_overflow_pages = 0;
5373 mx->mx_db.md_entries = NUMKEYS(fp);
5374 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5375 mx->mx_cursor.mc_snum = 1;
5376 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5377 mx->mx_cursor.mc_top = 0;
5378 mx->mx_cursor.mc_pg[0] = fp;
5379 mx->mx_cursor.mc_ki[0] = 0;
5380 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5381 mx->mx_db.md_flags = MDB_DUPFIXED;
5382 mx->mx_db.md_pad = fp->mp_pad;
5383 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5384 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5387 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5389 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5391 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5392 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5393 #if UINT_MAX < SIZE_MAX
5394 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5395 #ifdef MISALIGNED_OK
5396 mx->mx_dbx.md_cmp = mdb_cmp_long;
5398 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5403 /** Initialize a cursor for a given transaction and database. */
5405 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5410 mc->mc_db = &txn->mt_dbs[dbi];
5411 mc->mc_dbx = &txn->mt_dbxs[dbi];
5412 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5417 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5419 mc->mc_xcursor = mx;
5420 mdb_xcursor_init0(mc);
5422 mc->mc_xcursor = NULL;
5424 if (*mc->mc_dbflag & DB_STALE) {
5425 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5430 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5433 MDB_xcursor *mx = NULL;
5434 size_t size = sizeof(MDB_cursor);
5436 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5439 /* Allow read access to the freelist */
5440 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5443 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5444 size += sizeof(MDB_xcursor);
5446 if ((mc = malloc(size)) != NULL) {
5447 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5448 mx = (MDB_xcursor *)(mc + 1);
5450 mdb_cursor_init(mc, txn, dbi, mx);
5451 if (txn->mt_cursors) {
5452 mc->mc_next = txn->mt_cursors[dbi];
5453 txn->mt_cursors[dbi] = mc;
5455 mc->mc_flags |= C_ALLOCD;
5466 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5468 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5471 if (txn->mt_cursors)
5474 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5478 /* Return the count of duplicate data items for the current key */
5480 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5484 if (mc == NULL || countp == NULL)
5487 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5490 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5491 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5494 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5497 *countp = mc->mc_xcursor->mx_db.md_entries;
5503 mdb_cursor_close(MDB_cursor *mc)
5506 /* remove from txn, if tracked */
5507 if (mc->mc_txn->mt_cursors) {
5508 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5509 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5511 *prev = mc->mc_next;
5513 if (mc->mc_flags & C_ALLOCD)
5519 mdb_cursor_txn(MDB_cursor *mc)
5521 if (!mc) return NULL;
5526 mdb_cursor_dbi(MDB_cursor *mc)
5532 /** Replace the key for a node with a new key.
5533 * @param[in] mp The page containing the node to operate on.
5534 * @param[in] indx The index of the node to operate on.
5535 * @param[in] key The new key to use.
5536 * @return 0 on success, non-zero on failure.
5539 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5545 indx_t ptr, i, numkeys;
5548 node = NODEPTR(mp, indx);
5549 ptr = mp->mp_ptrs[indx];
5553 char kbuf2[(MAXKEYSIZE*2+1)];
5554 k2.mv_data = NODEKEY(node);
5555 k2.mv_size = node->mn_ksize;
5556 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5558 mdb_dkey(&k2, kbuf2),
5564 delta0 = delta = key->mv_size - node->mn_ksize;
5566 /* Must be 2-byte aligned. If new key is
5567 * shorter by 1, the shift will be skipped.
5569 delta += (delta & 1);
5571 if (delta > 0 && SIZELEFT(mp) < delta) {
5572 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5573 return MDB_PAGE_FULL;
5576 numkeys = NUMKEYS(mp);
5577 for (i = 0; i < numkeys; i++) {
5578 if (mp->mp_ptrs[i] <= ptr)
5579 mp->mp_ptrs[i] -= delta;
5582 base = (char *)mp + mp->mp_upper;
5583 len = ptr - mp->mp_upper + NODESIZE;
5584 memmove(base - delta, base, len);
5585 mp->mp_upper -= delta;
5587 node = NODEPTR(mp, indx);
5590 /* But even if no shift was needed, update ksize */
5592 node->mn_ksize = key->mv_size;
5595 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5600 /** Move a node from csrc to cdst.
5603 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5609 unsigned short flags;
5613 /* Mark src and dst as dirty. */
5614 if ((rc = mdb_page_touch(csrc)) ||
5615 (rc = mdb_page_touch(cdst)))
5618 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5619 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5620 key.mv_size = csrc->mc_db->md_pad;
5621 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5623 data.mv_data = NULL;
5627 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5628 assert(!((long)srcnode&1));
5629 srcpg = NODEPGNO(srcnode);
5630 flags = srcnode->mn_flags;
5631 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5632 unsigned int snum = csrc->mc_snum;
5634 /* must find the lowest key below src */
5635 mdb_page_search_root(csrc, NULL, 0);
5636 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5637 key.mv_size = csrc->mc_db->md_pad;
5638 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5640 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5641 key.mv_size = NODEKSZ(s2);
5642 key.mv_data = NODEKEY(s2);
5644 csrc->mc_snum = snum--;
5645 csrc->mc_top = snum;
5647 key.mv_size = NODEKSZ(srcnode);
5648 key.mv_data = NODEKEY(srcnode);
5650 data.mv_size = NODEDSZ(srcnode);
5651 data.mv_data = NODEDATA(srcnode);
5653 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5654 unsigned int snum = cdst->mc_snum;
5657 /* must find the lowest key below dst */
5658 mdb_page_search_root(cdst, NULL, 0);
5659 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5660 bkey.mv_size = cdst->mc_db->md_pad;
5661 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5663 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5664 bkey.mv_size = NODEKSZ(s2);
5665 bkey.mv_data = NODEKEY(s2);
5667 cdst->mc_snum = snum--;
5668 cdst->mc_top = snum;
5669 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5672 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5673 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5674 csrc->mc_ki[csrc->mc_top],
5676 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5677 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5679 /* Add the node to the destination page.
5681 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5682 if (rc != MDB_SUCCESS)
5685 /* Delete the node from the source page.
5687 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5690 /* Adjust other cursors pointing to mp */
5691 MDB_cursor *m2, *m3;
5692 MDB_dbi dbi = csrc->mc_dbi;
5693 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5695 if (csrc->mc_flags & C_SUB)
5698 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5699 if (m2 == csrc) continue;
5700 if (csrc->mc_flags & C_SUB)
5701 m3 = &m2->mc_xcursor->mx_cursor;
5704 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5705 csrc->mc_ki[csrc->mc_top]) {
5706 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5707 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5712 /* Update the parent separators.
5714 if (csrc->mc_ki[csrc->mc_top] == 0) {
5715 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5716 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5717 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5719 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5720 key.mv_size = NODEKSZ(srcnode);
5721 key.mv_data = NODEKEY(srcnode);
5723 DPRINTF("update separator for source page %zu to [%s]",
5724 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5725 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5726 &key)) != MDB_SUCCESS)
5729 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5731 nullkey.mv_size = 0;
5732 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5733 assert(rc == MDB_SUCCESS);
5737 if (cdst->mc_ki[cdst->mc_top] == 0) {
5738 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5739 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5740 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5742 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5743 key.mv_size = NODEKSZ(srcnode);
5744 key.mv_data = NODEKEY(srcnode);
5746 DPRINTF("update separator for destination page %zu to [%s]",
5747 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5748 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5749 &key)) != MDB_SUCCESS)
5752 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5754 nullkey.mv_size = 0;
5755 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5756 assert(rc == MDB_SUCCESS);
5763 /** Merge one page into another.
5764 * The nodes from the page pointed to by \b csrc will
5765 * be copied to the page pointed to by \b cdst and then
5766 * the \b csrc page will be freed.
5767 * @param[in] csrc Cursor pointing to the source page.
5768 * @param[in] cdst Cursor pointing to the destination page.
5771 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5779 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5780 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5782 assert(csrc->mc_snum > 1); /* can't merge root page */
5783 assert(cdst->mc_snum > 1);
5785 /* Mark dst as dirty. */
5786 if ((rc = mdb_page_touch(cdst)))
5789 /* Move all nodes from src to dst.
5791 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5792 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5793 key.mv_size = csrc->mc_db->md_pad;
5794 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5795 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5796 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5797 if (rc != MDB_SUCCESS)
5799 key.mv_data = (char *)key.mv_data + key.mv_size;
5802 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5803 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5804 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5805 unsigned int snum = csrc->mc_snum;
5807 /* must find the lowest key below src */
5808 mdb_page_search_root(csrc, NULL, 0);
5809 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5810 key.mv_size = csrc->mc_db->md_pad;
5811 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5813 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5814 key.mv_size = NODEKSZ(s2);
5815 key.mv_data = NODEKEY(s2);
5817 csrc->mc_snum = snum--;
5818 csrc->mc_top = snum;
5820 key.mv_size = srcnode->mn_ksize;
5821 key.mv_data = NODEKEY(srcnode);
5824 data.mv_size = NODEDSZ(srcnode);
5825 data.mv_data = NODEDATA(srcnode);
5826 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5827 if (rc != MDB_SUCCESS)
5832 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5833 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);
5835 /* Unlink the src page from parent and add to free list.
5837 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5838 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5840 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5844 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5845 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5846 csrc->mc_db->md_leaf_pages--;
5848 csrc->mc_db->md_branch_pages--;
5850 /* Adjust other cursors pointing to mp */
5851 MDB_cursor *m2, *m3;
5852 MDB_dbi dbi = csrc->mc_dbi;
5853 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5855 if (csrc->mc_flags & C_SUB)
5858 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5859 if (csrc->mc_flags & C_SUB)
5860 m3 = &m2->mc_xcursor->mx_cursor;
5863 if (m3 == csrc) continue;
5864 if (m3->mc_snum < csrc->mc_snum) continue;
5865 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5866 m3->mc_pg[csrc->mc_top] = mp;
5867 m3->mc_ki[csrc->mc_top] += nkeys;
5871 mdb_cursor_pop(csrc);
5873 return mdb_rebalance(csrc);
5876 /** Copy the contents of a cursor.
5877 * @param[in] csrc The cursor to copy from.
5878 * @param[out] cdst The cursor to copy to.
5881 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5885 cdst->mc_txn = csrc->mc_txn;
5886 cdst->mc_dbi = csrc->mc_dbi;
5887 cdst->mc_db = csrc->mc_db;
5888 cdst->mc_dbx = csrc->mc_dbx;
5889 cdst->mc_snum = csrc->mc_snum;
5890 cdst->mc_top = csrc->mc_top;
5891 cdst->mc_flags = csrc->mc_flags;
5893 for (i=0; i<csrc->mc_snum; i++) {
5894 cdst->mc_pg[i] = csrc->mc_pg[i];
5895 cdst->mc_ki[i] = csrc->mc_ki[i];
5899 /** Rebalance the tree after a delete operation.
5900 * @param[in] mc Cursor pointing to the page where rebalancing
5902 * @return 0 on success, non-zero on failure.
5905 mdb_rebalance(MDB_cursor *mc)
5915 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5916 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5917 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5918 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5922 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5925 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5926 DPRINTF("no need to rebalance page %zu, above fill threshold",
5932 if (mc->mc_snum < 2) {
5933 MDB_page *mp = mc->mc_pg[0];
5934 if (NUMKEYS(mp) == 0) {
5935 DPUTS("tree is completely empty");
5936 mc->mc_db->md_root = P_INVALID;
5937 mc->mc_db->md_depth = 0;
5938 mc->mc_db->md_leaf_pages = 0;
5939 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5943 /* Adjust other cursors pointing to mp */
5944 MDB_cursor *m2, *m3;
5945 MDB_dbi dbi = mc->mc_dbi;
5947 if (mc->mc_flags & C_SUB)
5950 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5951 if (m2 == mc) continue;
5952 if (mc->mc_flags & C_SUB)
5953 m3 = &m2->mc_xcursor->mx_cursor;
5956 if (m3->mc_snum < mc->mc_snum) continue;
5957 if (m3->mc_pg[0] == mp) {
5963 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5964 DPUTS("collapsing root page!");
5965 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5966 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5967 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5970 mc->mc_db->md_depth--;
5971 mc->mc_db->md_branch_pages--;
5973 /* Adjust other cursors pointing to mp */
5974 MDB_cursor *m2, *m3;
5975 MDB_dbi dbi = mc->mc_dbi;
5977 if (mc->mc_flags & C_SUB)
5980 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5981 if (m2 == mc) continue;
5982 if (mc->mc_flags & C_SUB)
5983 m3 = &m2->mc_xcursor->mx_cursor;
5986 if (m3->mc_snum < mc->mc_snum) continue;
5987 if (m3->mc_pg[0] == mp) {
5988 m3->mc_pg[0] = mc->mc_pg[0];
5993 DPUTS("root page doesn't need rebalancing");
5997 /* The parent (branch page) must have at least 2 pointers,
5998 * otherwise the tree is invalid.
6000 ptop = mc->mc_top-1;
6001 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6003 /* Leaf page fill factor is below the threshold.
6004 * Try to move keys from left or right neighbor, or
6005 * merge with a neighbor page.
6010 mdb_cursor_copy(mc, &mn);
6011 mn.mc_xcursor = NULL;
6013 if (mc->mc_ki[ptop] == 0) {
6014 /* We're the leftmost leaf in our parent.
6016 DPUTS("reading right neighbor");
6018 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6019 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6021 mn.mc_ki[mn.mc_top] = 0;
6022 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6024 /* There is at least one neighbor to the left.
6026 DPUTS("reading left neighbor");
6028 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6029 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6031 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6032 mc->mc_ki[mc->mc_top] = 0;
6035 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6036 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);
6038 /* If the neighbor page is above threshold and has at least two
6039 * keys, move one key from it.
6041 * Otherwise we should try to merge them.
6043 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6044 return mdb_node_move(&mn, mc);
6045 else { /* FIXME: if (has_enough_room()) */
6046 mc->mc_flags &= ~C_INITIALIZED;
6047 if (mc->mc_ki[ptop] == 0)
6048 return mdb_page_merge(&mn, mc);
6050 return mdb_page_merge(mc, &mn);
6054 /** Complete a delete operation started by #mdb_cursor_del(). */
6056 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6060 /* add overflow pages to free list */
6061 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6065 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6066 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6067 mc->mc_db->md_overflow_pages -= ovpages;
6068 for (i=0; i<ovpages; i++) {
6069 DPRINTF("freed ov page %zu", pg);
6070 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6074 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6075 mc->mc_db->md_entries--;
6076 rc = mdb_rebalance(mc);
6077 if (rc != MDB_SUCCESS)
6078 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6084 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6085 MDB_val *key, MDB_val *data)
6090 MDB_val rdata, *xdata;
6094 assert(key != NULL);
6096 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6098 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6101 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6105 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6109 mdb_cursor_init(&mc, txn, dbi, &mx);
6120 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6122 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6126 /** Split a page and insert a new node.
6127 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6128 * The cursor will be updated to point to the actual page and index where
6129 * the node got inserted after the split.
6130 * @param[in] newkey The key for the newly inserted node.
6131 * @param[in] newdata The data for the newly inserted node.
6132 * @param[in] newpgno The page number, if the new node is a branch node.
6133 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6134 * @return 0 on success, non-zero on failure.
6137 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6138 unsigned int nflags)
6141 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6144 unsigned int i, j, split_indx, nkeys, pmax;
6146 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6148 MDB_page *mp, *rp, *pp;
6153 mp = mc->mc_pg[mc->mc_top];
6154 newindx = mc->mc_ki[mc->mc_top];
6156 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6157 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6158 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6160 /* Create a right sibling. */
6161 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6163 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6165 if (mc->mc_snum < 2) {
6166 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6168 /* shift current top to make room for new parent */
6169 mc->mc_pg[1] = mc->mc_pg[0];
6170 mc->mc_ki[1] = mc->mc_ki[0];
6173 mc->mc_db->md_root = pp->mp_pgno;
6174 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6175 mc->mc_db->md_depth++;
6178 /* Add left (implicit) pointer. */
6179 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6180 /* undo the pre-push */
6181 mc->mc_pg[0] = mc->mc_pg[1];
6182 mc->mc_ki[0] = mc->mc_ki[1];
6183 mc->mc_db->md_root = mp->mp_pgno;
6184 mc->mc_db->md_depth--;
6191 ptop = mc->mc_top-1;
6192 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6195 mc->mc_flags |= C_SPLITTING;
6196 mdb_cursor_copy(mc, &mn);
6197 mn.mc_pg[mn.mc_top] = rp;
6198 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6200 if (nflags & MDB_APPEND) {
6201 mn.mc_ki[mn.mc_top] = 0;
6203 split_indx = newindx;
6208 nkeys = NUMKEYS(mp);
6209 split_indx = nkeys / 2;
6210 if (newindx < split_indx)
6216 unsigned int lsize, rsize, ksize;
6217 /* Move half of the keys to the right sibling */
6219 x = mc->mc_ki[mc->mc_top] - split_indx;
6220 ksize = mc->mc_db->md_pad;
6221 split = LEAF2KEY(mp, split_indx, ksize);
6222 rsize = (nkeys - split_indx) * ksize;
6223 lsize = (nkeys - split_indx) * sizeof(indx_t);
6224 mp->mp_lower -= lsize;
6225 rp->mp_lower += lsize;
6226 mp->mp_upper += rsize - lsize;
6227 rp->mp_upper -= rsize - lsize;
6228 sepkey.mv_size = ksize;
6229 if (newindx == split_indx) {
6230 sepkey.mv_data = newkey->mv_data;
6232 sepkey.mv_data = split;
6235 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6236 memcpy(rp->mp_ptrs, split, rsize);
6237 sepkey.mv_data = rp->mp_ptrs;
6238 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6239 memcpy(ins, newkey->mv_data, ksize);
6240 mp->mp_lower += sizeof(indx_t);
6241 mp->mp_upper -= ksize - sizeof(indx_t);
6244 memcpy(rp->mp_ptrs, split, x * ksize);
6245 ins = LEAF2KEY(rp, x, ksize);
6246 memcpy(ins, newkey->mv_data, ksize);
6247 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6248 rp->mp_lower += sizeof(indx_t);
6249 rp->mp_upper -= ksize - sizeof(indx_t);
6250 mc->mc_ki[mc->mc_top] = x;
6251 mc->mc_pg[mc->mc_top] = rp;
6256 /* For leaf pages, check the split point based on what
6257 * fits where, since otherwise mdb_node_add can fail.
6259 * This check is only needed when the data items are
6260 * relatively large, such that being off by one will
6261 * make the difference between success or failure.
6263 * It's also relevant if a page happens to be laid out
6264 * such that one half of its nodes are all "small" and
6265 * the other half of its nodes are "large." If the new
6266 * item is also "large" and falls on the half with
6267 * "large" nodes, it also may not fit.
6270 unsigned int psize, nsize;
6271 /* Maximum free space in an empty page */
6272 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6273 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6274 if ((nkeys < 20) || (nsize > pmax/16)) {
6275 if (newindx <= split_indx) {
6278 for (i=0; i<split_indx; i++) {
6279 node = NODEPTR(mp, i);
6280 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6281 if (F_ISSET(node->mn_flags, F_BIGDATA))
6282 psize += sizeof(pgno_t);
6284 psize += NODEDSZ(node);
6288 split_indx = newindx;
6299 for (i=nkeys-1; i>=split_indx; i--) {
6300 node = NODEPTR(mp, i);
6301 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6302 if (F_ISSET(node->mn_flags, F_BIGDATA))
6303 psize += sizeof(pgno_t);
6305 psize += NODEDSZ(node);
6309 split_indx = newindx;
6320 /* First find the separating key between the split pages.
6321 * The case where newindx == split_indx is ambiguous; the
6322 * new item could go to the new page or stay on the original
6323 * page. If newpos == 1 it goes to the new page.
6325 if (newindx == split_indx && newpos) {
6326 sepkey.mv_size = newkey->mv_size;
6327 sepkey.mv_data = newkey->mv_data;
6329 node = NODEPTR(mp, split_indx);
6330 sepkey.mv_size = node->mn_ksize;
6331 sepkey.mv_data = NODEKEY(node);
6335 DPRINTF("separator is [%s]", DKEY(&sepkey));
6337 /* Copy separator key to the parent.
6339 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6343 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6346 if (mn.mc_snum == mc->mc_snum) {
6347 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6348 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6349 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6350 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6355 /* Right page might now have changed parent.
6356 * Check if left page also changed parent.
6358 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6359 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6360 for (i=0; i<ptop; i++) {
6361 mc->mc_pg[i] = mn.mc_pg[i];
6362 mc->mc_ki[i] = mn.mc_ki[i];
6364 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6365 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6369 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6372 mc->mc_flags ^= C_SPLITTING;
6373 if (rc != MDB_SUCCESS) {
6376 if (nflags & MDB_APPEND) {
6377 mc->mc_pg[mc->mc_top] = rp;
6378 mc->mc_ki[mc->mc_top] = 0;
6379 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6382 for (i=0; i<mc->mc_top; i++)
6383 mc->mc_ki[i] = mn.mc_ki[i];
6390 /* Move half of the keys to the right sibling. */
6392 /* grab a page to hold a temporary copy */
6393 copy = mdb_page_malloc(mc);
6397 copy->mp_pgno = mp->mp_pgno;
6398 copy->mp_flags = mp->mp_flags;
6399 copy->mp_lower = PAGEHDRSZ;
6400 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6401 mc->mc_pg[mc->mc_top] = copy;
6402 for (i = j = 0; i <= nkeys; j++) {
6403 if (i == split_indx) {
6404 /* Insert in right sibling. */
6405 /* Reset insert index for right sibling. */
6406 if (i != newindx || (newpos ^ ins_new)) {
6408 mc->mc_pg[mc->mc_top] = rp;
6412 if (i == newindx && !ins_new) {
6413 /* Insert the original entry that caused the split. */
6414 rkey.mv_data = newkey->mv_data;
6415 rkey.mv_size = newkey->mv_size;
6424 /* Update index for the new key. */
6425 mc->mc_ki[mc->mc_top] = j;
6426 } else if (i == nkeys) {
6429 node = NODEPTR(mp, i);
6430 rkey.mv_data = NODEKEY(node);
6431 rkey.mv_size = node->mn_ksize;
6433 xdata.mv_data = NODEDATA(node);
6434 xdata.mv_size = NODEDSZ(node);
6437 pgno = NODEPGNO(node);
6438 flags = node->mn_flags;
6443 if (!IS_LEAF(mp) && j == 0) {
6444 /* First branch index doesn't need key data. */
6448 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6452 nkeys = NUMKEYS(copy);
6453 for (i=0; i<nkeys; i++)
6454 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6455 mp->mp_lower = copy->mp_lower;
6456 mp->mp_upper = copy->mp_upper;
6457 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6458 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6460 /* reset back to original page */
6461 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6462 mc->mc_pg[mc->mc_top] = mp;
6463 if (nflags & MDB_RESERVE) {
6464 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6465 if (!(node->mn_flags & F_BIGDATA))
6466 newdata->mv_data = NODEDATA(node);
6472 /* return tmp page to freelist */
6473 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
6474 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
6475 mc->mc_txn->mt_env->me_dpages = copy;
6478 /* Adjust other cursors pointing to mp */
6479 MDB_cursor *m2, *m3;
6480 MDB_dbi dbi = mc->mc_dbi;
6481 int fixup = NUMKEYS(mp);
6483 if (mc->mc_flags & C_SUB)
6486 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6487 if (m2 == mc) continue;
6488 if (mc->mc_flags & C_SUB)
6489 m3 = &m2->mc_xcursor->mx_cursor;
6492 if (!(m3->mc_flags & C_INITIALIZED))
6494 if (m3->mc_flags & C_SPLITTING)
6499 for (k=m3->mc_top; k>=0; k--) {
6500 m3->mc_ki[k+1] = m3->mc_ki[k];
6501 m3->mc_pg[k+1] = m3->mc_pg[k];
6503 if (m3->mc_ki[0] >= split_indx) {
6508 m3->mc_pg[0] = mc->mc_pg[0];
6512 if (m3->mc_pg[mc->mc_top] == mp) {
6513 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6514 m3->mc_ki[mc->mc_top]++;
6515 if (m3->mc_ki[mc->mc_top] >= fixup) {
6516 m3->mc_pg[mc->mc_top] = rp;
6517 m3->mc_ki[mc->mc_top] -= fixup;
6518 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6520 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6521 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6530 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6531 MDB_val *key, MDB_val *data, unsigned int flags)
6536 assert(key != NULL);
6537 assert(data != NULL);
6539 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6542 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6546 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6550 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6553 mdb_cursor_init(&mc, txn, dbi, &mx);
6554 return mdb_cursor_put(&mc, key, data, flags);
6557 /** Only a subset of the @ref mdb_env flags can be changed
6558 * at runtime. Changing other flags requires closing the environment
6559 * and re-opening it with the new flags.
6561 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
6563 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6565 if ((flag & CHANGEABLE) != flag)
6568 env->me_flags |= flag;
6570 env->me_flags &= ~flag;
6575 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6580 *arg = env->me_flags;
6585 mdb_env_get_path(MDB_env *env, const char **arg)
6590 *arg = env->me_path;
6594 /** Common code for #mdb_stat() and #mdb_env_stat().
6595 * @param[in] env the environment to operate in.
6596 * @param[in] db the #MDB_db record containing the stats to return.
6597 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6598 * @return 0, this function always succeeds.
6601 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6603 arg->ms_psize = env->me_psize;
6604 arg->ms_depth = db->md_depth;
6605 arg->ms_branch_pages = db->md_branch_pages;
6606 arg->ms_leaf_pages = db->md_leaf_pages;
6607 arg->ms_overflow_pages = db->md_overflow_pages;
6608 arg->ms_entries = db->md_entries;
6613 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6617 if (env == NULL || arg == NULL)
6620 toggle = mdb_env_pick_meta(env);
6622 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6625 /** Set the default comparison functions for a database.
6626 * Called immediately after a database is opened to set the defaults.
6627 * The user can then override them with #mdb_set_compare() or
6628 * #mdb_set_dupsort().
6629 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6630 * @param[in] dbi A database handle returned by #mdb_open()
6633 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6635 uint16_t f = txn->mt_dbs[dbi].md_flags;
6637 txn->mt_dbxs[dbi].md_cmp =
6638 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6639 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6641 txn->mt_dbxs[dbi].md_dcmp =
6642 !(f & MDB_DUPSORT) ? 0 :
6643 ((f & MDB_INTEGERDUP)
6644 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6645 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6648 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6653 int rc, dbflag, exact;
6654 unsigned int unused = 0;
6657 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6658 mdb_default_cmp(txn, FREE_DBI);
6664 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
6665 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
6666 mdb_default_cmp(txn, MAIN_DBI);
6670 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6671 mdb_default_cmp(txn, MAIN_DBI);
6674 /* Is the DB already open? */
6676 for (i=2; i<txn->mt_numdbs; i++) {
6677 if (!txn->mt_dbxs[i].md_name.mv_size) {
6678 /* Remember this free slot */
6679 if (!unused) unused = i;
6682 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6683 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6689 /* If no free slot and max hit, fail */
6690 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
6691 return MDB_DBS_FULL;
6693 /* Find the DB info */
6697 key.mv_data = (void *)name;
6698 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6699 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6700 if (rc == MDB_SUCCESS) {
6701 /* make sure this is actually a DB */
6702 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6703 if (!(node->mn_flags & F_SUBDATA))
6705 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6706 /* Create if requested */
6708 data.mv_size = sizeof(MDB_db);
6709 data.mv_data = &dummy;
6710 memset(&dummy, 0, sizeof(dummy));
6711 dummy.md_root = P_INVALID;
6712 dummy.md_flags = flags & 0xffff;
6713 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6717 /* OK, got info, add to table */
6718 if (rc == MDB_SUCCESS) {
6719 unsigned int slot = unused ? unused : txn->mt_numdbs;
6720 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
6721 txn->mt_dbxs[slot].md_name.mv_size = len;
6722 txn->mt_dbxs[slot].md_rel = NULL;
6723 txn->mt_dbflags[slot] = dbflag;
6724 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
6726 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
6727 mdb_default_cmp(txn, slot);
6730 txn->mt_env->me_numdbs++;
6737 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6739 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6742 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6745 void mdb_close(MDB_env *env, MDB_dbi dbi)
6748 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6750 ptr = env->me_dbxs[dbi].md_name.mv_data;
6751 env->me_dbxs[dbi].md_name.mv_data = NULL;
6752 env->me_dbxs[dbi].md_name.mv_size = 0;
6756 /** Add all the DB's pages to the free list.
6757 * @param[in] mc Cursor on the DB to free.
6758 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6759 * @return 0 on success, non-zero on failure.
6762 mdb_drop0(MDB_cursor *mc, int subs)
6766 rc = mdb_page_search(mc, NULL, 0);
6767 if (rc == MDB_SUCCESS) {
6772 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6773 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6776 mdb_cursor_copy(mc, &mx);
6777 while (mc->mc_snum > 0) {
6778 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6779 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6780 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6781 if (ni->mn_flags & F_SUBDATA) {
6782 mdb_xcursor_init1(mc, ni);
6783 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6789 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6791 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6794 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6799 rc = mdb_cursor_sibling(mc, 1);
6801 /* no more siblings, go back to beginning
6802 * of previous level. (stack was already popped
6803 * by mdb_cursor_sibling)
6805 for (i=1; i<mc->mc_top; i++)
6806 mc->mc_pg[i] = mx.mc_pg[i];
6810 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6811 mc->mc_db->md_root);
6816 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6821 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6824 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6827 rc = mdb_cursor_open(txn, dbi, &mc);
6831 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6835 /* Can't delete the main DB */
6836 if (del && dbi > MAIN_DBI) {
6837 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6839 mdb_close(txn->mt_env, dbi);
6841 /* reset the DB record, mark it dirty */
6842 txn->mt_dbflags[dbi] |= DB_DIRTY;
6843 txn->mt_dbs[dbi].md_depth = 0;
6844 txn->mt_dbs[dbi].md_branch_pages = 0;
6845 txn->mt_dbs[dbi].md_leaf_pages = 0;
6846 txn->mt_dbs[dbi].md_overflow_pages = 0;
6847 txn->mt_dbs[dbi].md_entries = 0;
6848 txn->mt_dbs[dbi].md_root = P_INVALID;
6850 if (!txn->mt_u.dirty_list[0].mid) {
6853 /* make sure we have at least one dirty page in this txn
6854 * otherwise these changes will be ignored.
6856 key.mv_size = sizeof(txnid_t);
6857 key.mv_data = &txn->mt_txnid;
6858 data.mv_size = sizeof(MDB_ID);
6859 data.mv_data = txn->mt_free_pgs;
6860 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
6861 rc = mdb_cursor_put(&m2, &key, &data, 0);
6865 mdb_cursor_close(mc);
6869 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6871 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6874 txn->mt_dbxs[dbi].md_cmp = cmp;
6878 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6880 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6883 txn->mt_dbxs[dbi].md_dcmp = cmp;
6887 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6889 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6892 txn->mt_dbxs[dbi].md_rel = rel;
6896 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6898 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6901 txn->mt_dbxs[dbi].md_relctx = ctx;