2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2012 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
35 #include <sys/types.h>
37 #include <sys/param.h>
43 #ifdef HAVE_SYS_FILE_H
60 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
61 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
64 #if defined(__APPLE__) || defined (BSD)
71 #include <semaphore.h>
76 #include <valgrind/memcheck.h>
77 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
78 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
79 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
80 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
81 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
83 #define VGMEMP_CREATE(h,r,z)
84 #define VGMEMP_ALLOC(h,a,s)
85 #define VGMEMP_FREE(h,a)
86 #define VGMEMP_DESTROY(h)
87 #define VGMEMP_DEFINED(a,s)
91 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
92 /* Solaris just defines one or the other */
93 # define LITTLE_ENDIAN 1234
94 # define BIG_ENDIAN 4321
95 # ifdef _LITTLE_ENDIAN
96 # define BYTE_ORDER LITTLE_ENDIAN
98 # define BYTE_ORDER BIG_ENDIAN
101 # define BYTE_ORDER __BYTE_ORDER
105 #ifndef LITTLE_ENDIAN
106 #define LITTLE_ENDIAN __LITTLE_ENDIAN
109 #define BIG_ENDIAN __BIG_ENDIAN
112 #if defined(__i386) || defined(__x86_64)
113 #define MISALIGNED_OK 1
119 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
120 # error "Unknown or unsupported endianness (BYTE_ORDER)"
121 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
122 # error "Two's complement, reasonably sized integer types, please"
125 /** @defgroup internal MDB Internals
128 /** @defgroup compat Windows Compatibility Macros
129 * A bunch of macros to minimize the amount of platform-specific ifdefs
130 * needed throughout the rest of the code. When the features this library
131 * needs are similar enough to POSIX to be hidden in a one-or-two line
132 * replacement, this macro approach is used.
136 #define pthread_t DWORD
137 #define pthread_mutex_t HANDLE
138 #define pthread_key_t DWORD
139 #define pthread_self() GetCurrentThreadId()
140 #define pthread_key_create(x,y) (*(x) = TlsAlloc())
141 #define pthread_key_delete(x) TlsFree(x)
142 #define pthread_getspecific(x) TlsGetValue(x)
143 #define pthread_setspecific(x,y) TlsSetValue(x,y)
144 #define pthread_mutex_unlock(x) ReleaseMutex(x)
145 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
146 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
147 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
148 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
149 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
150 #define getpid() GetCurrentProcessId()
151 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
152 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
153 #define ErrCode() GetLastError()
154 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
155 #define close(fd) CloseHandle(fd)
156 #define munmap(ptr,len) UnmapViewOfFile(ptr)
159 #define LOCK_MUTEX_R(env) sem_wait((env)->me_rmutex)
160 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
161 #define LOCK_MUTEX_W(env) sem_wait((env)->me_wmutex)
162 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
163 #define MDB_FDATASYNC(fd) fsync(fd)
166 #define MDB_FDATASYNC(fd) fsync(fd)
168 /** Lock the reader mutex.
170 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
171 /** Unlock the reader mutex.
173 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
175 /** Lock the writer mutex.
176 * Only a single write transaction is allowed at a time. Other writers
177 * will block waiting for this mutex.
179 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
180 /** Unlock the writer mutex.
182 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
183 #endif /* USE_POSIX_SEM */
185 /** Get the error code for the last failed system function.
187 #define ErrCode() errno
189 /** An abstraction for a file handle.
190 * On POSIX systems file handles are small integers. On Windows
191 * they're opaque pointers.
195 /** A value for an invalid file handle.
196 * Mainly used to initialize file variables and signify that they are
199 #define INVALID_HANDLE_VALUE (-1)
201 /** Get the size of a memory page for the system.
202 * This is the basic size that the platform's memory manager uses, and is
203 * fundamental to the use of memory-mapped files.
205 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
208 #if defined(_WIN32) || defined(USE_POSIX_SEM)
211 #define MNAME_LEN (sizeof(pthread_mutex_t))
217 /** A flag for opening a file and requesting synchronous data writes.
218 * This is only used when writing a meta page. It's not strictly needed;
219 * we could just do a normal write and then immediately perform a flush.
220 * But if this flag is available it saves us an extra system call.
222 * @note If O_DSYNC is undefined but exists in /usr/include,
223 * preferably set some compiler flag to get the definition.
224 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
227 # define MDB_DSYNC O_DSYNC
231 /** Function for flushing the data of a file. Define this to fsync
232 * if fdatasync() is not supported.
234 #ifndef MDB_FDATASYNC
235 # define MDB_FDATASYNC fdatasync
239 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
250 /** A page number in the database.
251 * Note that 64 bit page numbers are overkill, since pages themselves
252 * already represent 12-13 bits of addressable memory, and the OS will
253 * always limit applications to a maximum of 63 bits of address space.
255 * @note In the #MDB_node structure, we only store 48 bits of this value,
256 * which thus limits us to only 60 bits of addressable data.
258 typedef MDB_ID pgno_t;
260 /** A transaction ID.
261 * See struct MDB_txn.mt_txnid for details.
263 typedef MDB_ID txnid_t;
265 /** @defgroup debug Debug Macros
269 /** Enable debug output.
270 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
271 * read from and written to the database (used for free space management).
276 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
277 # define DPRINTF (void) /* Vararg macros may be unsupported */
279 static int mdb_debug;
280 static txnid_t mdb_debug_start;
282 /** Print a debug message with printf formatting. */
283 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
284 ((void) ((mdb_debug) && \
285 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
287 # define DPRINTF(fmt, ...) ((void) 0)
289 /** Print a debug string.
290 * The string is printed literally, with no format processing.
292 #define DPUTS(arg) DPRINTF("%s", arg)
295 /** A default memory page size.
296 * The actual size is platform-dependent, but we use this for
297 * boot-strapping. We probably should not be using this any more.
298 * The #GET_PAGESIZE() macro is used to get the actual size.
300 * Note that we don't currently support Huge pages. On Linux,
301 * regular data files cannot use Huge pages, and in general
302 * Huge pages aren't actually pageable. We rely on the OS
303 * demand-pager to read our data and page it out when memory
304 * pressure from other processes is high. So until OSs have
305 * actual paging support for Huge pages, they're not viable.
307 #define MDB_PAGESIZE 4096
309 /** The minimum number of keys required in a database page.
310 * Setting this to a larger value will place a smaller bound on the
311 * maximum size of a data item. Data items larger than this size will
312 * be pushed into overflow pages instead of being stored directly in
313 * the B-tree node. This value used to default to 4. With a page size
314 * of 4096 bytes that meant that any item larger than 1024 bytes would
315 * go into an overflow page. That also meant that on average 2-3KB of
316 * each overflow page was wasted space. The value cannot be lower than
317 * 2 because then there would no longer be a tree structure. With this
318 * value, items larger than 2KB will go into overflow pages, and on
319 * average only 1KB will be wasted.
321 #define MDB_MINKEYS 2
323 /** A stamp that identifies a file as an MDB file.
324 * There's nothing special about this value other than that it is easily
325 * recognizable, and it will reflect any byte order mismatches.
327 #define MDB_MAGIC 0xBEEFC0DE
329 /** The version number for a database's file format. */
330 #define MDB_VERSION 1
332 /** The maximum size of a key in the database.
333 * While data items have essentially unbounded size, we require that
334 * keys all fit onto a regular page. This limit could be raised a bit
335 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
337 #define MAXKEYSIZE 511
342 * This is used for printing a hex dump of a key's contents.
344 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
345 /** Display a key in hex.
347 * Invoke a function to display a key in hex.
349 #define DKEY(x) mdb_dkey(x, kbuf)
351 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
355 /** An invalid page number.
356 * Mainly used to denote an empty tree.
358 #define P_INVALID (~(pgno_t)0)
360 /** Test if a flag \b f is set in a flag word \b w. */
361 #define F_ISSET(w, f) (((w) & (f)) == (f))
363 /** Used for offsets within a single page.
364 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
367 typedef uint16_t indx_t;
369 /** Default size of memory map.
370 * This is certainly too small for any actual applications. Apps should always set
371 * the size explicitly using #mdb_env_set_mapsize().
373 #define DEFAULT_MAPSIZE 1048576
375 /** @defgroup readers Reader Lock Table
376 * Readers don't acquire any locks for their data access. Instead, they
377 * simply record their transaction ID in the reader table. The reader
378 * mutex is needed just to find an empty slot in the reader table. The
379 * slot's address is saved in thread-specific data so that subsequent read
380 * transactions started by the same thread need no further locking to proceed.
382 * Since the database uses multi-version concurrency control, readers don't
383 * actually need any locking. This table is used to keep track of which
384 * readers are using data from which old transactions, so that we'll know
385 * when a particular old transaction is no longer in use. Old transactions
386 * that have discarded any data pages can then have those pages reclaimed
387 * for use by a later write transaction.
389 * The lock table is constructed such that reader slots are aligned with the
390 * processor's cache line size. Any slot is only ever used by one thread.
391 * This alignment guarantees that there will be no contention or cache
392 * thrashing as threads update their own slot info, and also eliminates
393 * any need for locking when accessing a slot.
395 * A writer thread will scan every slot in the table to determine the oldest
396 * outstanding reader transaction. Any freed pages older than this will be
397 * reclaimed by the writer. The writer doesn't use any locks when scanning
398 * this table. This means that there's no guarantee that the writer will
399 * see the most up-to-date reader info, but that's not required for correct
400 * operation - all we need is to know the upper bound on the oldest reader,
401 * we don't care at all about the newest reader. So the only consequence of
402 * reading stale information here is that old pages might hang around a
403 * while longer before being reclaimed. That's actually good anyway, because
404 * the longer we delay reclaiming old pages, the more likely it is that a
405 * string of contiguous pages can be found after coalescing old pages from
406 * many old transactions together.
408 * @todo We don't actually do such coalescing yet, we grab pages from one
409 * old transaction at a time.
412 /** Number of slots in the reader table.
413 * This value was chosen somewhat arbitrarily. 126 readers plus a
414 * couple mutexes fit exactly into 8KB on my development machine.
415 * Applications should set the table size using #mdb_env_set_maxreaders().
417 #define DEFAULT_READERS 126
419 /** The size of a CPU cache line in bytes. We want our lock structures
420 * aligned to this size to avoid false cache line sharing in the
422 * This value works for most CPUs. For Itanium this should be 128.
428 /** The information we store in a single slot of the reader table.
429 * In addition to a transaction ID, we also record the process and
430 * thread ID that owns a slot, so that we can detect stale information,
431 * e.g. threads or processes that went away without cleaning up.
432 * @note We currently don't check for stale records. We simply re-init
433 * the table when we know that we're the only process opening the
436 typedef struct MDB_rxbody {
437 /** The current Transaction ID when this transaction began.
438 * Multiple readers that start at the same time will probably have the
439 * same ID here. Again, it's not important to exclude them from
440 * anything; all we need to know is which version of the DB they
441 * started from so we can avoid overwriting any data used in that
442 * particular version.
445 /** The process ID of the process owning this reader txn. */
447 /** The thread ID of the thread owning this txn. */
451 /** The actual reader record, with cacheline padding. */
452 typedef struct MDB_reader {
455 /** shorthand for mrb_txnid */
456 #define mr_txnid mru.mrx.mrb_txnid
457 #define mr_pid mru.mrx.mrb_pid
458 #define mr_tid mru.mrx.mrb_tid
459 /** cache line alignment */
460 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
464 /** The header for the reader table.
465 * The table resides in a memory-mapped file. (This is a different file
466 * than is used for the main database.)
468 * For POSIX the actual mutexes reside in the shared memory of this
469 * mapped file. On Windows, mutexes are named objects allocated by the
470 * kernel; we store the mutex names in this mapped file so that other
471 * processes can grab them. This same approach is also used on
472 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
473 * process-shared POSIX mutexes. For these cases where a named object
474 * is used, the object name is derived from a 64 bit FNV hash of the
475 * environment pathname. As such, naming collisions are extremely
476 * unlikely. If a collision occurs, the results are unpredictable.
478 typedef struct MDB_txbody {
479 /** Stamp identifying this as an MDB file. It must be set
482 /** Version number of this lock file. Must be set to #MDB_VERSION. */
483 uint32_t mtb_version;
484 #if defined(_WIN32) || defined(USE_POSIX_SEM)
485 char mtb_rmname[MNAME_LEN];
487 /** Mutex protecting access to this table.
488 * This is the reader lock that #LOCK_MUTEX_R acquires.
490 pthread_mutex_t mtb_mutex;
492 /** The ID of the last transaction committed to the database.
493 * This is recorded here only for convenience; the value can always
494 * be determined by reading the main database meta pages.
497 /** The number of slots that have been used in the reader table.
498 * This always records the maximum count, it is not decremented
499 * when readers release their slots.
501 unsigned mtb_numreaders;
504 /** The actual reader table definition. */
505 typedef struct MDB_txninfo {
508 #define mti_magic mt1.mtb.mtb_magic
509 #define mti_version mt1.mtb.mtb_version
510 #define mti_mutex mt1.mtb.mtb_mutex
511 #define mti_rmname mt1.mtb.mtb_rmname
512 #define mti_txnid mt1.mtb.mtb_txnid
513 #define mti_numreaders mt1.mtb.mtb_numreaders
514 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
517 #if defined(_WIN32) || defined(USE_POSIX_SEM)
518 char mt2_wmname[MNAME_LEN];
519 #define mti_wmname mt2.mt2_wmname
521 pthread_mutex_t mt2_wmutex;
522 #define mti_wmutex mt2.mt2_wmutex
524 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
526 MDB_reader mti_readers[1];
530 /** Common header for all page types.
531 * Overflow records occupy a number of contiguous pages with no
532 * headers on any page after the first.
534 typedef struct MDB_page {
535 #define mp_pgno mp_p.p_pgno
536 #define mp_next mp_p.p_next
538 pgno_t p_pgno; /**< page number */
539 void * p_next; /**< for in-memory list of freed structs */
542 /** @defgroup mdb_page Page Flags
544 * Flags for the page headers.
547 #define P_BRANCH 0x01 /**< branch page */
548 #define P_LEAF 0x02 /**< leaf page */
549 #define P_OVERFLOW 0x04 /**< overflow page */
550 #define P_META 0x08 /**< meta page */
551 #define P_DIRTY 0x10 /**< dirty page */
552 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
553 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
555 uint16_t mp_flags; /**< @ref mdb_page */
556 #define mp_lower mp_pb.pb.pb_lower
557 #define mp_upper mp_pb.pb.pb_upper
558 #define mp_pages mp_pb.pb_pages
561 indx_t pb_lower; /**< lower bound of free space */
562 indx_t pb_upper; /**< upper bound of free space */
564 uint32_t pb_pages; /**< number of overflow pages */
566 indx_t mp_ptrs[1]; /**< dynamic size */
569 /** Size of the page header, excluding dynamic data at the end */
570 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
572 /** Address of first usable data byte in a page, after the header */
573 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
575 /** Number of nodes on a page */
576 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
578 /** The amount of space remaining in the page */
579 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
581 /** The percentage of space used in the page, in tenths of a percent. */
582 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
583 ((env)->me_psize - PAGEHDRSZ))
584 /** The minimum page fill factor, in tenths of a percent.
585 * Pages emptier than this are candidates for merging.
587 #define FILL_THRESHOLD 250
589 /** Test if a page is a leaf page */
590 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
591 /** Test if a page is a LEAF2 page */
592 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
593 /** Test if a page is a branch page */
594 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
595 /** Test if a page is an overflow page */
596 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
597 /** Test if a page is a sub page */
598 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
600 /** The number of overflow pages needed to store the given size. */
601 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
603 /** Header for a single key/data pair within a page.
604 * We guarantee 2-byte alignment for nodes.
606 typedef struct MDB_node {
607 /** lo and hi are used for data size on leaf nodes and for
608 * child pgno on branch nodes. On 64 bit platforms, flags
609 * is also used for pgno. (Branch nodes have no flags).
610 * They are in host byte order in case that lets some
611 * accesses be optimized into a 32-bit word access.
613 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
614 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
615 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
616 /** @defgroup mdb_node Node Flags
618 * Flags for node headers.
621 #define F_BIGDATA 0x01 /**< data put on overflow page */
622 #define F_SUBDATA 0x02 /**< data is a sub-database */
623 #define F_DUPDATA 0x04 /**< data has duplicates */
625 /** valid flags for #mdb_node_add() */
626 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
629 unsigned short mn_flags; /**< @ref mdb_node */
630 unsigned short mn_ksize; /**< key size */
631 char mn_data[1]; /**< key and data are appended here */
634 /** Size of the node header, excluding dynamic data at the end */
635 #define NODESIZE offsetof(MDB_node, mn_data)
637 /** Bit position of top word in page number, for shifting mn_flags */
638 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
640 /** Size of a node in a branch page with a given key.
641 * This is just the node header plus the key, there is no data.
643 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
645 /** Size of a node in a leaf page with a given key and data.
646 * This is node header plus key plus data size.
648 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
650 /** Address of node \b i in page \b p */
651 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
653 /** Address of the key for the node */
654 #define NODEKEY(node) (void *)((node)->mn_data)
656 /** Address of the data for a node */
657 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
659 /** Get the page number pointed to by a branch node */
660 #define NODEPGNO(node) \
661 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
662 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
663 /** Set the page number in a branch node */
664 #define SETPGNO(node,pgno) do { \
665 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
666 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
668 /** Get the size of the data in a leaf node */
669 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
670 /** Set the size of the data for a leaf node */
671 #define SETDSZ(node,size) do { \
672 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
673 /** The size of a key in a node */
674 #define NODEKSZ(node) ((node)->mn_ksize)
676 /** Copy a page number from src to dst */
678 #define COPY_PGNO(dst,src) dst = src
680 #if SIZE_MAX > 4294967295UL
681 #define COPY_PGNO(dst,src) do { \
682 unsigned short *s, *d; \
683 s = (unsigned short *)&(src); \
684 d = (unsigned short *)&(dst); \
691 #define COPY_PGNO(dst,src) do { \
692 unsigned short *s, *d; \
693 s = (unsigned short *)&(src); \
694 d = (unsigned short *)&(dst); \
700 /** The address of a key in a LEAF2 page.
701 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
702 * There are no node headers, keys are stored contiguously.
704 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
706 /** Set the \b node's key into \b key, if requested. */
707 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
708 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
710 /** Information about a single database in the environment. */
711 typedef struct MDB_db {
712 uint32_t md_pad; /**< also ksize for LEAF2 pages */
713 uint16_t md_flags; /**< @ref mdb_open */
714 uint16_t md_depth; /**< depth of this tree */
715 pgno_t md_branch_pages; /**< number of internal pages */
716 pgno_t md_leaf_pages; /**< number of leaf pages */
717 pgno_t md_overflow_pages; /**< number of overflow pages */
718 size_t md_entries; /**< number of data items */
719 pgno_t md_root; /**< the root page of this tree */
722 /** Handle for the DB used to track free pages. */
724 /** Handle for the default DB. */
727 /** Meta page content. */
728 typedef struct MDB_meta {
729 /** Stamp identifying this as an MDB file. It must be set
732 /** Version number of this lock file. Must be set to #MDB_VERSION. */
734 void *mm_address; /**< address for fixed mapping */
735 size_t mm_mapsize; /**< size of mmap region */
736 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
737 /** The size of pages used in this DB */
738 #define mm_psize mm_dbs[0].md_pad
739 /** Any persistent environment flags. @ref mdb_env */
740 #define mm_flags mm_dbs[0].md_flags
741 pgno_t mm_last_pg; /**< last used page in file */
742 txnid_t mm_txnid; /**< txnid that committed this page */
745 /** Buffer for a stack-allocated dirty page.
746 * The members define size and alignment, and silence type
747 * aliasing warnings. They are not used directly; that could
748 * mean incorrectly using several union members in parallel.
750 typedef union MDB_pagebuf {
751 char mb_raw[MDB_PAGESIZE];
754 char mm_pad[PAGEHDRSZ];
759 /** Auxiliary DB info.
760 * The information here is mostly static/read-only. There is
761 * only a single copy of this record in the environment.
763 typedef struct MDB_dbx {
764 MDB_val md_name; /**< name of the database */
765 MDB_cmp_func *md_cmp; /**< function for comparing keys */
766 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
767 MDB_rel_func *md_rel; /**< user relocate function */
768 void *md_relctx; /**< user-provided context for md_rel */
771 /** A database transaction.
772 * Every operation requires a transaction handle.
775 MDB_txn *mt_parent; /**< parent of a nested txn */
776 MDB_txn *mt_child; /**< nested txn under this txn */
777 pgno_t mt_next_pgno; /**< next unallocated page */
778 /** The ID of this transaction. IDs are integers incrementing from 1.
779 * Only committed write transactions increment the ID. If a transaction
780 * aborts, the ID may be re-used by the next writer.
783 MDB_env *mt_env; /**< the DB environment */
784 /** The list of pages that became unused during this transaction.
788 MDB_ID2L dirty_list; /**< modified pages */
789 MDB_reader *reader; /**< this thread's slot in the reader table */
791 /** Array of records for each DB known in the environment. */
793 /** Array of MDB_db records for each known DB */
795 /** @defgroup mt_dbflag Transaction DB Flags
799 #define DB_DIRTY 0x01 /**< DB was written in this txn */
800 #define DB_STALE 0x02 /**< DB record is older than txnID */
802 /** Array of cursors for each DB */
803 MDB_cursor **mt_cursors;
804 /** Array of flags for each DB */
805 unsigned char *mt_dbflags;
806 /** Number of DB records in use. This number only ever increments;
807 * we don't decrement it when individual DB handles are closed.
811 /** @defgroup mdb_txn Transaction Flags
815 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
816 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
818 unsigned int mt_flags; /**< @ref mdb_txn */
819 /** Tracks which of the two meta pages was used at the start
820 * of this transaction.
822 unsigned int mt_toggle;
825 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
826 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
827 * raise this on a 64 bit machine.
829 #define CURSOR_STACK 32
833 /** Cursors are used for all DB operations */
835 /** Next cursor on this DB in this txn */
837 /** Original cursor if this is a shadow */
839 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
840 struct MDB_xcursor *mc_xcursor;
841 /** The transaction that owns this cursor */
843 /** The database handle this cursor operates on */
845 /** The database record for this cursor */
847 /** The database auxiliary record for this cursor */
849 /** The @ref mt_dbflag for this database */
850 unsigned char *mc_dbflag;
851 unsigned short mc_snum; /**< number of pushed pages */
852 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
853 /** @defgroup mdb_cursor Cursor Flags
855 * Cursor state flags.
858 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
859 #define C_EOF 0x02 /**< No more data */
860 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
861 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
862 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
863 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
865 unsigned int mc_flags; /**< @ref mdb_cursor */
866 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
867 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
870 /** Context for sorted-dup records.
871 * We could have gone to a fully recursive design, with arbitrarily
872 * deep nesting of sub-databases. But for now we only handle these
873 * levels - main DB, optional sub-DB, sorted-duplicate DB.
875 typedef struct MDB_xcursor {
876 /** A sub-cursor for traversing the Dup DB */
877 MDB_cursor mx_cursor;
878 /** The database record for this Dup DB */
880 /** The auxiliary DB record for this Dup DB */
882 /** The @ref mt_dbflag for this Dup DB */
883 unsigned char mx_dbflag;
886 /** A set of pages freed by an earlier transaction. */
887 typedef struct MDB_oldpages {
888 /** Usually we only read one record from the FREEDB at a time, but
889 * in case we read more, this will chain them together.
891 struct MDB_oldpages *mo_next;
892 /** The ID of the transaction in which these pages were freed. */
894 /** An #MDB_IDL of the pages */
895 pgno_t mo_pages[1]; /* dynamic */
898 /** The database environment. */
900 HANDLE me_fd; /**< The main data file */
901 HANDLE me_lfd; /**< The lock file */
902 HANDLE me_mfd; /**< just for writing the meta pages */
903 /** Failed to update the meta page. Probably an I/O error. */
904 #define MDB_FATAL_ERROR 0x80000000U
905 uint32_t me_flags; /**< @ref mdb_env */
906 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
907 unsigned int me_maxreaders; /**< size of the reader table */
908 MDB_dbi me_numdbs; /**< number of DBs opened */
909 MDB_dbi me_maxdbs; /**< size of the DB table */
910 char *me_path; /**< path to the DB files */
911 char *me_map; /**< the memory map of the data file */
912 MDB_txninfo *me_txns; /**< the memory map of the lock file */
913 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
914 MDB_txn *me_txn; /**< current write transaction */
915 size_t me_mapsize; /**< size of the data memory map */
916 off_t me_size; /**< current file size */
917 pgno_t me_maxpg; /**< me_mapsize / me_psize */
918 txnid_t me_pgfirst; /**< ID of first old page record we used */
919 txnid_t me_pglast; /**< ID of last old page record we used */
920 MDB_dbx *me_dbxs; /**< array of static DB info */
921 uint16_t *me_dbflags; /**< array of DB flags */
922 MDB_oldpages *me_pghead; /**< list of old page records */
923 MDB_oldpages *me_pgfree; /**< list of page records to free */
924 pthread_key_t me_txkey; /**< thread-key for readers */
925 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
926 /** IDL of pages that became unused in a write txn */
928 /** ID2L of pages that were written during a write txn */
929 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
931 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
935 sem_t *me_rmutex; /* Apple doesn't support shared mutexes */
939 /** max number of pages to commit in one writev() call */
940 #define MDB_COMMIT_PAGES 64
941 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
942 #undef MDB_COMMIT_PAGES
943 #define MDB_COMMIT_PAGES IOV_MAX
946 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
947 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
948 static int mdb_page_touch(MDB_cursor *mc);
950 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
951 static int mdb_page_search_root(MDB_cursor *mc,
952 MDB_val *key, int modify);
953 #define MDB_PS_MODIFY 1
954 #define MDB_PS_ROOTONLY 2
955 static int mdb_page_search(MDB_cursor *mc,
956 MDB_val *key, int flags);
957 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
959 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
960 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
961 pgno_t newpgno, unsigned int nflags);
963 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
964 static int mdb_env_pick_meta(const MDB_env *env);
965 static int mdb_env_write_meta(MDB_txn *txn);
967 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
968 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
969 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
970 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
971 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
972 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
973 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
974 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
975 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
977 static int mdb_rebalance(MDB_cursor *mc);
978 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
980 static void mdb_cursor_pop(MDB_cursor *mc);
981 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
983 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
984 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
985 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
986 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
987 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
989 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
990 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
992 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
993 static void mdb_xcursor_init0(MDB_cursor *mc);
994 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
996 static int mdb_drop0(MDB_cursor *mc, int subs);
997 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1000 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1004 static SECURITY_DESCRIPTOR mdb_null_sd;
1005 static SECURITY_ATTRIBUTES mdb_all_sa;
1006 static int mdb_sec_inited;
1009 /** Return the library version info. */
1011 mdb_version(int *major, int *minor, int *patch)
1013 if (major) *major = MDB_VERSION_MAJOR;
1014 if (minor) *minor = MDB_VERSION_MINOR;
1015 if (patch) *patch = MDB_VERSION_PATCH;
1016 return MDB_VERSION_STRING;
1019 /** Table of descriptions for MDB @ref errors */
1020 static char *const mdb_errstr[] = {
1021 "MDB_KEYEXIST: Key/data pair already exists",
1022 "MDB_NOTFOUND: No matching key/data pair found",
1023 "MDB_PAGE_NOTFOUND: Requested page not found",
1024 "MDB_CORRUPTED: Located page was wrong type",
1025 "MDB_PANIC: Update of meta page failed",
1026 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1027 "MDB_INVALID: File is not an MDB file",
1028 "MDB_MAP_FULL: Environment mapsize limit reached",
1029 "MDB_DBS_FULL: Environment maxdbs limit reached",
1030 "MDB_READERS_FULL: Environment maxreaders limit reached",
1031 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1032 "MDB_TXN_FULL: Nested transaction has too many dirty pages - transaction too big",
1033 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1034 "MDB_PAGE_FULL: Internal error - page has no more space"
1038 mdb_strerror(int err)
1042 return ("Successful return: 0");
1044 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1045 i = err - MDB_KEYEXIST;
1046 return mdb_errstr[i];
1049 return strerror(err);
1053 /** Display a key in hexadecimal and return the address of the result.
1054 * @param[in] key the key to display
1055 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1056 * @return The key in hexadecimal form.
1059 mdb_dkey(MDB_val *key, char *buf)
1062 unsigned char *c = key->mv_data;
1064 if (key->mv_size > MAXKEYSIZE)
1065 return "MAXKEYSIZE";
1066 /* may want to make this a dynamic check: if the key is mostly
1067 * printable characters, print it as-is instead of converting to hex.
1071 for (i=0; i<key->mv_size; i++)
1072 ptr += sprintf(ptr, "%02x", *c++);
1074 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1079 /** Display all the keys in the page. */
1081 mdb_page_list(MDB_page *mp)
1084 unsigned int i, nkeys, nsize;
1088 nkeys = NUMKEYS(mp);
1089 fprintf(stderr, "numkeys %d\n", nkeys);
1090 for (i=0; i<nkeys; i++) {
1091 node = NODEPTR(mp, i);
1092 key.mv_size = node->mn_ksize;
1093 key.mv_data = node->mn_data;
1094 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1095 if (F_ISSET(node->mn_flags, F_BIGDATA))
1096 nsize += sizeof(pgno_t);
1098 nsize += NODEDSZ(node);
1099 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1104 mdb_cursor_chk(MDB_cursor *mc)
1110 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1111 for (i=0; i<mc->mc_top; i++) {
1113 node = NODEPTR(mp, mc->mc_ki[i]);
1114 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1117 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1123 /** Count all the pages in each DB and in the freelist
1124 * and make sure it matches the actual number of pages
1127 static void mdb_audit(MDB_txn *txn)
1131 MDB_ID freecount, count;
1136 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1137 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1138 freecount += *(MDB_ID *)data.mv_data;
1141 for (i = 0; i<txn->mt_numdbs; i++) {
1142 MDB_xcursor mx, *mxp;
1143 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1144 mdb_cursor_init(&mc, txn, i, mxp);
1145 if (txn->mt_dbs[i].md_root == P_INVALID)
1147 count += txn->mt_dbs[i].md_branch_pages +
1148 txn->mt_dbs[i].md_leaf_pages +
1149 txn->mt_dbs[i].md_overflow_pages;
1150 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1151 mdb_page_search(&mc, NULL, 0);
1155 mp = mc.mc_pg[mc.mc_top];
1156 for (j=0; j<NUMKEYS(mp); j++) {
1157 MDB_node *leaf = NODEPTR(mp, j);
1158 if (leaf->mn_flags & F_SUBDATA) {
1160 memcpy(&db, NODEDATA(leaf), sizeof(db));
1161 count += db.md_branch_pages + db.md_leaf_pages +
1162 db.md_overflow_pages;
1166 while (mdb_cursor_sibling(&mc, 1) == 0);
1169 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1170 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1171 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1177 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1179 return txn->mt_dbxs[dbi].md_cmp(a, b);
1183 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1185 if (txn->mt_dbxs[dbi].md_dcmp)
1186 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1188 return EINVAL; /* too bad you can't distinguish this from a valid result */
1191 /** Allocate a single page.
1192 * Re-use old malloc'd pages first, otherwise just malloc.
1195 mdb_page_malloc(MDB_cursor *mc) {
1197 size_t sz = mc->mc_txn->mt_env->me_psize;
1198 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1199 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1200 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1201 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1202 } else if ((ret = malloc(sz)) != NULL) {
1203 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1208 /** Allocate pages for writing.
1209 * If there are free pages available from older transactions, they
1210 * will be re-used first. Otherwise a new page will be allocated.
1211 * @param[in] mc cursor A cursor handle identifying the transaction and
1212 * database for which we are allocating.
1213 * @param[in] num the number of pages to allocate.
1214 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1215 * will always be satisfied by a single contiguous chunk of memory.
1216 * @return 0 on success, non-zero on failure.
1219 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1221 MDB_txn *txn = mc->mc_txn;
1223 pgno_t pgno = P_INVALID;
1228 /* The free list won't have any content at all until txn 2 has
1229 * committed. The pages freed by txn 2 will be unreferenced
1230 * after txn 3 commits, and so will be safe to re-use in txn 4.
1232 if (txn->mt_txnid > 3) {
1234 if (!txn->mt_env->me_pghead &&
1235 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1236 /* See if there's anything in the free DB */
1240 txnid_t *kptr, oldest, last;
1242 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1243 if (!txn->mt_env->me_pgfirst) {
1244 mdb_page_search(&m2, NULL, 0);
1245 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1246 kptr = (txnid_t *)NODEKEY(leaf);
1253 last = txn->mt_env->me_pglast + 1;
1255 key.mv_data = &last;
1256 key.mv_size = sizeof(last);
1257 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1260 last = *(txnid_t *)key.mv_data;
1265 oldest = txn->mt_txnid - 1;
1266 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
1267 txnid_t mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
1268 if (mr && mr < oldest)
1273 if (oldest > last) {
1274 /* It's usable, grab it.
1279 if (!txn->mt_env->me_pgfirst) {
1280 mdb_node_read(txn, leaf, &data);
1282 txn->mt_env->me_pglast = last;
1283 if (!txn->mt_env->me_pgfirst)
1284 txn->mt_env->me_pgfirst = last;
1285 idl = (MDB_ID *) data.mv_data;
1286 /* We might have a zero-length IDL due to freelist growth
1287 * during a prior commit
1289 if (!idl[0]) goto again;
1290 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1293 mop->mo_next = txn->mt_env->me_pghead;
1294 mop->mo_txnid = last;
1295 txn->mt_env->me_pghead = mop;
1296 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1301 DPRINTF("IDL read txn %zu root %zu num %zu",
1302 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1303 for (i=0; i<idl[0]; i++) {
1304 DPRINTF("IDL %zu", idl[i+1]);
1311 if (txn->mt_env->me_pghead) {
1312 MDB_oldpages *mop = txn->mt_env->me_pghead;
1314 /* FIXME: For now, always use fresh pages. We
1315 * really ought to search the free list for a
1320 /* peel pages off tail, so we only have to truncate the list */
1321 pgno = MDB_IDL_LAST(mop->mo_pages);
1322 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1324 if (mop->mo_pages[2] > mop->mo_pages[1])
1325 mop->mo_pages[0] = 0;
1329 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1330 txn->mt_env->me_pghead = mop->mo_next;
1331 if (mc->mc_dbi == FREE_DBI) {
1332 mop->mo_next = txn->mt_env->me_pgfree;
1333 txn->mt_env->me_pgfree = mop;
1342 if (pgno == P_INVALID) {
1343 /* DB size is maxed out */
1344 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1345 DPUTS("DB size maxed out");
1346 return MDB_MAP_FULL;
1349 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1350 if (pgno == P_INVALID) {
1351 pgno = txn->mt_next_pgno;
1352 txn->mt_next_pgno += num;
1354 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1357 if (txn->mt_env->me_dpages && num == 1) {
1358 np = txn->mt_env->me_dpages;
1359 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1360 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1361 txn->mt_env->me_dpages = np->mp_next;
1363 size_t sz = txn->mt_env->me_psize * num;
1364 if ((np = malloc(sz)) == NULL)
1366 VGMEMP_ALLOC(txn->mt_env, np, sz);
1368 if (pgno == P_INVALID) {
1369 np->mp_pgno = txn->mt_next_pgno;
1370 txn->mt_next_pgno += num;
1375 mid.mid = np->mp_pgno;
1377 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1378 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1380 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1387 /** Copy a page: avoid copying unused portions of the page.
1388 * @param[in] dst page to copy into
1389 * @param[in] src page to copy from
1392 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1394 dst->mp_flags = src->mp_flags | P_DIRTY;
1395 dst->mp_pages = src->mp_pages;
1397 if (IS_LEAF2(src)) {
1398 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1400 unsigned int i, nkeys = NUMKEYS(src);
1401 for (i=0; i<nkeys; i++)
1402 dst->mp_ptrs[i] = src->mp_ptrs[i];
1403 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1404 psize - src->mp_upper);
1408 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1409 * @param[in] mc cursor pointing to the page to be touched
1410 * @return 0 on success, non-zero on failure.
1413 mdb_page_touch(MDB_cursor *mc)
1415 MDB_page *mp = mc->mc_pg[mc->mc_top];
1419 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1421 if ((rc = mdb_page_alloc(mc, 1, &np)))
1423 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1424 assert(mp->mp_pgno != np->mp_pgno);
1425 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1427 /* If page isn't full, just copy the used portion */
1428 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1431 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1433 np->mp_flags |= P_DIRTY;
1438 /* Adjust other cursors pointing to mp */
1439 if (mc->mc_flags & C_SUB) {
1440 MDB_cursor *m2, *m3;
1441 MDB_dbi dbi = mc->mc_dbi-1;
1443 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1444 if (m2 == mc) continue;
1445 m3 = &m2->mc_xcursor->mx_cursor;
1446 if (m3->mc_snum < mc->mc_snum) continue;
1447 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1448 m3->mc_pg[mc->mc_top] = mp;
1454 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1455 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1456 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1457 m2->mc_pg[mc->mc_top] = mp;
1461 mc->mc_pg[mc->mc_top] = mp;
1462 /** If this page has a parent, update the parent to point to
1466 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1468 mc->mc_db->md_root = mp->mp_pgno;
1469 } else if (mc->mc_txn->mt_parent) {
1472 /* If txn has a parent, make sure the page is in our
1475 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1476 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1477 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1478 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1479 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1480 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1481 mc->mc_pg[mc->mc_top] = mp;
1487 np = mdb_page_malloc(mc);
1490 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1491 mid.mid = np->mp_pgno;
1493 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1501 mdb_env_sync(MDB_env *env, int force)
1504 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1505 if (env->me_flags & MDB_WRITEMAP) {
1506 int flags = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
1507 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1510 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1514 if (MDB_FDATASYNC(env->me_fd))
1521 /** Make shadow copies of all of parent txn's cursors */
1523 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1525 MDB_cursor *mc, *m2;
1526 unsigned int i, j, size;
1528 for (i=0;i<src->mt_numdbs; i++) {
1529 if (src->mt_cursors[i]) {
1530 size = sizeof(MDB_cursor);
1531 if (src->mt_cursors[i]->mc_xcursor)
1532 size += sizeof(MDB_xcursor);
1533 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1540 mc->mc_db = &dst->mt_dbs[i];
1541 mc->mc_dbx = m2->mc_dbx;
1542 mc->mc_dbflag = &dst->mt_dbflags[i];
1543 mc->mc_snum = m2->mc_snum;
1544 mc->mc_top = m2->mc_top;
1545 mc->mc_flags = m2->mc_flags | C_SHADOW;
1546 for (j=0; j<mc->mc_snum; j++) {
1547 mc->mc_pg[j] = m2->mc_pg[j];
1548 mc->mc_ki[j] = m2->mc_ki[j];
1550 if (m2->mc_xcursor) {
1551 MDB_xcursor *mx, *mx2;
1552 mx = (MDB_xcursor *)(mc+1);
1553 mc->mc_xcursor = mx;
1554 mx2 = m2->mc_xcursor;
1555 mx->mx_db = mx2->mx_db;
1556 mx->mx_dbx = mx2->mx_dbx;
1557 mx->mx_dbflag = mx2->mx_dbflag;
1558 mx->mx_cursor.mc_txn = dst;
1559 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1560 mx->mx_cursor.mc_db = &mx->mx_db;
1561 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1562 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1563 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1564 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1565 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1566 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1567 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1568 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1571 mc->mc_xcursor = NULL;
1573 mc->mc_next = dst->mt_cursors[i];
1574 dst->mt_cursors[i] = mc;
1581 /** Merge shadow cursors back into parent's */
1583 mdb_cursor_merge(MDB_txn *txn)
1586 for (i=0; i<txn->mt_numdbs; i++) {
1587 if (txn->mt_cursors[i]) {
1589 while ((mc = txn->mt_cursors[i])) {
1590 txn->mt_cursors[i] = mc->mc_next;
1591 if (mc->mc_flags & C_SHADOW) {
1592 MDB_cursor *m2 = mc->mc_orig;
1594 m2->mc_snum = mc->mc_snum;
1595 m2->mc_top = mc->mc_top;
1596 for (j=0; j<mc->mc_snum; j++) {
1597 m2->mc_pg[j] = mc->mc_pg[j];
1598 m2->mc_ki[j] = mc->mc_ki[j];
1601 if (mc->mc_flags & C_ALLOCD)
1609 mdb_txn_reset0(MDB_txn *txn);
1611 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1612 * @param[in] txn the transaction handle to initialize
1613 * @return 0 on success, non-zero on failure. This can only
1614 * fail for read-only transactions, and then only if the
1615 * reader table is full.
1618 mdb_txn_renew0(MDB_txn *txn)
1620 MDB_env *env = txn->mt_env;
1624 txn->mt_numdbs = env->me_numdbs;
1625 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1627 if (txn->mt_flags & MDB_TXN_RDONLY) {
1628 MDB_reader *r = pthread_getspecific(env->me_txkey);
1630 pid_t pid = getpid();
1631 pthread_t tid = pthread_self();
1634 for (i=0; i<env->me_txns->mti_numreaders; i++)
1635 if (env->me_txns->mti_readers[i].mr_pid == 0)
1637 if (i == env->me_maxreaders) {
1638 UNLOCK_MUTEX_R(env);
1639 return MDB_READERS_FULL;
1641 env->me_txns->mti_readers[i].mr_pid = pid;
1642 env->me_txns->mti_readers[i].mr_tid = tid;
1643 if (i >= env->me_txns->mti_numreaders)
1644 env->me_txns->mti_numreaders = i+1;
1645 UNLOCK_MUTEX_R(env);
1646 r = &env->me_txns->mti_readers[i];
1647 pthread_setspecific(env->me_txkey, r);
1649 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1650 txn->mt_toggle = txn->mt_txnid & 1;
1651 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1652 txn->mt_u.reader = r;
1656 txn->mt_txnid = env->me_txns->mti_txnid;
1657 txn->mt_toggle = txn->mt_txnid & 1;
1658 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1661 if (txn->mt_txnid == mdb_debug_start)
1664 txn->mt_u.dirty_list = env->me_dirty_list;
1665 txn->mt_u.dirty_list[0].mid = 0;
1666 txn->mt_free_pgs = env->me_free_pgs;
1667 txn->mt_free_pgs[0] = 0;
1671 /* Copy the DB info and flags */
1672 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1673 for (i=2; i<txn->mt_numdbs; i++)
1674 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1675 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1676 if (txn->mt_numdbs > 2)
1677 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1683 mdb_txn_renew(MDB_txn *txn)
1690 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1691 DPUTS("environment had fatal error, must shutdown!");
1695 rc = mdb_txn_renew0(txn);
1696 if (rc == MDB_SUCCESS) {
1697 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1698 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1699 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1705 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1710 if (env->me_flags & MDB_FATAL_ERROR) {
1711 DPUTS("environment had fatal error, must shutdown!");
1714 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1717 /* parent already has an active child txn */
1718 if (parent->mt_child) {
1721 /* nested TXNs not supported here */
1722 if (env->me_flags & MDB_WRITEMAP)
1725 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1726 if (!(flags & MDB_RDONLY))
1727 size += env->me_maxdbs * sizeof(MDB_cursor *);
1729 if ((txn = calloc(1, size)) == NULL) {
1730 DPRINTF("calloc: %s", strerror(ErrCode()));
1733 txn->mt_dbs = (MDB_db *)(txn+1);
1734 if (flags & MDB_RDONLY) {
1735 txn->mt_flags |= MDB_TXN_RDONLY;
1736 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1738 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1739 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1744 txn->mt_free_pgs = mdb_midl_alloc();
1745 if (!txn->mt_free_pgs) {
1749 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1750 if (!txn->mt_u.dirty_list) {
1751 free(txn->mt_free_pgs);
1755 txn->mt_txnid = parent->mt_txnid;
1756 txn->mt_toggle = parent->mt_toggle;
1757 txn->mt_u.dirty_list[0].mid = 0;
1758 txn->mt_free_pgs[0] = 0;
1759 txn->mt_next_pgno = parent->mt_next_pgno;
1760 parent->mt_child = txn;
1761 txn->mt_parent = parent;
1762 txn->mt_numdbs = parent->mt_numdbs;
1763 txn->mt_dbxs = parent->mt_dbxs;
1764 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1765 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1766 mdb_cursor_shadow(parent, txn);
1769 rc = mdb_txn_renew0(txn);
1775 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1776 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1777 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1783 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1784 * @param[in] txn the transaction handle to reset
1787 mdb_txn_reset0(MDB_txn *txn)
1789 MDB_env *env = txn->mt_env;
1791 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1792 txn->mt_u.reader->mr_txnid = 0;
1798 /* close(free) all cursors */
1799 for (i=0; i<txn->mt_numdbs; i++) {
1800 if (txn->mt_cursors[i]) {
1802 while ((mc = txn->mt_cursors[i])) {
1803 txn->mt_cursors[i] = mc->mc_next;
1804 if (mc->mc_flags & C_ALLOCD)
1810 if (!(env->me_flags & MDB_WRITEMAP)) {
1811 /* return all dirty pages to dpage list */
1812 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1813 dp = txn->mt_u.dirty_list[i].mptr;
1814 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1815 dp->mp_next = txn->mt_env->me_dpages;
1816 VGMEMP_FREE(txn->mt_env, dp);
1817 txn->mt_env->me_dpages = dp;
1819 /* large pages just get freed directly */
1820 VGMEMP_FREE(txn->mt_env, dp);
1826 if (txn->mt_parent) {
1827 txn->mt_parent->mt_child = NULL;
1828 mdb_midl_free(txn->mt_free_pgs);
1829 free(txn->mt_u.dirty_list);
1832 if (mdb_midl_shrink(&txn->mt_free_pgs))
1833 env->me_free_pgs = txn->mt_free_pgs;
1836 while ((mop = txn->mt_env->me_pghead)) {
1837 txn->mt_env->me_pghead = mop->mo_next;
1840 txn->mt_env->me_pgfirst = 0;
1841 txn->mt_env->me_pglast = 0;
1844 /* The writer mutex was locked in mdb_txn_begin. */
1845 UNLOCK_MUTEX_W(env);
1850 mdb_txn_reset(MDB_txn *txn)
1855 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1856 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1857 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1859 mdb_txn_reset0(txn);
1863 mdb_txn_abort(MDB_txn *txn)
1868 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1869 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1870 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1873 mdb_txn_abort(txn->mt_child);
1875 mdb_txn_reset0(txn);
1880 mdb_txn_commit(MDB_txn *txn)
1888 pgno_t next, freecnt;
1891 assert(txn != NULL);
1892 assert(txn->mt_env != NULL);
1894 if (txn->mt_child) {
1895 mdb_txn_commit(txn->mt_child);
1896 txn->mt_child = NULL;
1901 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1902 if (txn->mt_numdbs > env->me_numdbs) {
1903 /* update the DB flags */
1905 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
1906 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
1913 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1914 DPUTS("error flag is set, can't commit");
1916 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1921 /* Merge (and close) our cursors with parent's */
1922 mdb_cursor_merge(txn);
1924 if (txn->mt_parent) {
1930 /* Update parent's DB table */
1931 ip = &txn->mt_parent->mt_dbs[2];
1932 jp = &txn->mt_dbs[2];
1933 for (i = 2; i < txn->mt_numdbs; i++) {
1934 if (ip->md_root != jp->md_root)
1938 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1940 /* Append our free list to parent's */
1941 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1943 mdb_midl_free(txn->mt_free_pgs);
1945 /* Merge our dirty list with parent's */
1946 dst = txn->mt_parent->mt_u.dirty_list;
1947 src = txn->mt_u.dirty_list;
1948 x = mdb_mid2l_search(dst, src[1].mid);
1949 for (y=1; y<=src[0].mid; y++) {
1950 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1954 dst[x].mptr = src[y].mptr;
1957 for (; y<=src[0].mid; y++) {
1958 if (++x >= MDB_IDL_UM_MAX) {
1960 return MDB_TXN_FULL;
1965 free(txn->mt_u.dirty_list);
1966 txn->mt_parent->mt_child = NULL;
1971 if (txn != env->me_txn) {
1972 DPUTS("attempt to commit unknown transaction");
1977 if (!txn->mt_u.dirty_list[0].mid)
1980 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1981 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1983 /* Update DB root pointers. Their pages have already been
1984 * touched so this is all in-place and cannot fail.
1986 if (txn->mt_numdbs > 2) {
1989 data.mv_size = sizeof(MDB_db);
1991 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
1992 for (i = 2; i < txn->mt_numdbs; i++) {
1993 if (txn->mt_dbflags[i] & DB_DIRTY) {
1994 data.mv_data = &txn->mt_dbs[i];
1995 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2000 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2002 /* should only be one record now */
2003 if (env->me_pghead) {
2004 /* make sure first page of freeDB is touched and on freelist */
2005 mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2008 /* Delete IDLs we used from the free list */
2009 if (env->me_pgfirst) {
2014 key.mv_size = sizeof(cur);
2015 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
2018 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
2019 rc = mdb_cursor_del(&mc, 0);
2025 env->me_pgfirst = 0;
2029 /* save to free list */
2031 freecnt = txn->mt_free_pgs[0];
2032 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2035 /* make sure last page of freeDB is touched and on freelist */
2036 key.mv_size = MAXKEYSIZE+1;
2038 mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2040 mdb_midl_sort(txn->mt_free_pgs);
2044 MDB_IDL idl = txn->mt_free_pgs;
2045 DPRINTF("IDL write txn %zu root %zu num %zu",
2046 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2047 for (i=0; i<idl[0]; i++) {
2048 DPRINTF("IDL %zu", idl[i+1]);
2052 /* write to last page of freeDB */
2053 key.mv_size = sizeof(pgno_t);
2054 key.mv_data = &txn->mt_txnid;
2055 data.mv_data = txn->mt_free_pgs;
2056 /* The free list can still grow during this call,
2057 * despite the pre-emptive touches above. So check
2058 * and make sure the entire thing got written.
2061 freecnt = txn->mt_free_pgs[0];
2062 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2063 rc = mdb_cursor_put(&mc, &key, &data, 0);
2068 } while (freecnt != txn->mt_free_pgs[0]);
2070 /* should only be one record now */
2072 if (env->me_pghead) {
2078 mop = env->me_pghead;
2080 key.mv_size = sizeof(id);
2082 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2083 data.mv_data = mop->mo_pages;
2084 orig = mop->mo_pages[0];
2085 /* These steps may grow the freelist again
2086 * due to freed overflow pages...
2088 mdb_cursor_put(&mc, &key, &data, 0);
2089 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2090 /* could have been used again here */
2091 if (mop->mo_pages[0] != orig) {
2092 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2093 data.mv_data = mop->mo_pages;
2095 mdb_cursor_put(&mc, &key, &data, 0);
2097 env->me_pghead = NULL;
2100 /* was completely used up */
2101 mdb_cursor_del(&mc, 0);
2105 env->me_pgfirst = 0;
2109 while (env->me_pgfree) {
2110 MDB_oldpages *mop = env->me_pgfree;
2111 env->me_pgfree = mop->mo_next;
2115 /* Check for growth of freelist again */
2116 if (freecnt != txn->mt_free_pgs[0])
2119 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2120 if (mdb_midl_shrink(&txn->mt_free_pgs))
2121 env->me_free_pgs = txn->mt_free_pgs;
2128 if (env->me_flags & MDB_WRITEMAP) {
2129 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2130 dp = txn->mt_u.dirty_list[i].mptr;
2131 /* clear dirty flag */
2132 dp->mp_flags &= ~P_DIRTY;
2133 txn->mt_u.dirty_list[i].mid = 0;
2135 txn->mt_u.dirty_list[0].mid = 0;
2139 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2145 /* Windows actually supports scatter/gather I/O, but only on
2146 * unbuffered file handles. Since we're relying on the OS page
2147 * cache for all our data, that's self-defeating. So we just
2148 * write pages one at a time. We use the ov structure to set
2149 * the write offset, to at least save the overhead of a Seek
2153 memset(&ov, 0, sizeof(ov));
2154 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2156 dp = txn->mt_u.dirty_list[i].mptr;
2157 DPRINTF("committing page %zu", dp->mp_pgno);
2158 size = dp->mp_pgno * env->me_psize;
2159 ov.Offset = size & 0xffffffff;
2160 ov.OffsetHigh = size >> 16;
2161 ov.OffsetHigh >>= 16;
2162 /* clear dirty flag */
2163 dp->mp_flags &= ~P_DIRTY;
2164 wsize = env->me_psize;
2165 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2166 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2169 DPRINTF("WriteFile: %d", n);
2176 struct iovec iov[MDB_COMMIT_PAGES];
2180 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2181 dp = txn->mt_u.dirty_list[i].mptr;
2182 if (dp->mp_pgno != next) {
2184 rc = writev(env->me_fd, iov, n);
2188 DPUTS("short write, filesystem full?");
2190 DPRINTF("writev: %s", strerror(n));
2197 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2200 DPRINTF("committing page %zu", dp->mp_pgno);
2201 iov[n].iov_len = env->me_psize;
2202 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2203 iov[n].iov_base = (char *)dp;
2204 size += iov[n].iov_len;
2205 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2206 /* clear dirty flag */
2207 dp->mp_flags &= ~P_DIRTY;
2208 if (++n >= MDB_COMMIT_PAGES) {
2218 rc = writev(env->me_fd, iov, n);
2222 DPUTS("short write, filesystem full?");
2224 DPRINTF("writev: %s", strerror(n));
2231 /* Drop the dirty pages.
2233 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2234 dp = txn->mt_u.dirty_list[i].mptr;
2235 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2236 dp->mp_next = txn->mt_env->me_dpages;
2237 VGMEMP_FREE(txn->mt_env, dp);
2238 txn->mt_env->me_dpages = dp;
2240 VGMEMP_FREE(txn->mt_env, dp);
2243 txn->mt_u.dirty_list[i].mid = 0;
2245 txn->mt_u.dirty_list[0].mid = 0;
2248 if ((n = mdb_env_sync(env, 0)) != 0 ||
2249 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2256 if (txn->mt_numdbs > env->me_numdbs) {
2257 /* update the DB flags */
2259 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2260 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2264 UNLOCK_MUTEX_W(env);
2270 /** Read the environment parameters of a DB environment before
2271 * mapping it into memory.
2272 * @param[in] env the environment handle
2273 * @param[out] meta address of where to store the meta information
2274 * @return 0 on success, non-zero on failure.
2277 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2284 /* We don't know the page size yet, so use a minimum value.
2288 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2290 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2295 else if (rc != MDB_PAGESIZE) {
2299 DPRINTF("read: %s", strerror(err));
2303 p = (MDB_page *)&pbuf;
2305 if (!F_ISSET(p->mp_flags, P_META)) {
2306 DPRINTF("page %zu not a meta page", p->mp_pgno);
2311 if (m->mm_magic != MDB_MAGIC) {
2312 DPUTS("meta has invalid magic");
2316 if (m->mm_version != MDB_VERSION) {
2317 DPRINTF("database is version %u, expected version %u",
2318 m->mm_version, MDB_VERSION);
2319 return MDB_VERSION_MISMATCH;
2322 memcpy(meta, m, sizeof(*m));
2326 /** Write the environment parameters of a freshly created DB environment.
2327 * @param[in] env the environment handle
2328 * @param[out] meta address of where to store the meta information
2329 * @return 0 on success, non-zero on failure.
2332 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2339 DPUTS("writing new meta page");
2341 GET_PAGESIZE(psize);
2343 meta->mm_magic = MDB_MAGIC;
2344 meta->mm_version = MDB_VERSION;
2345 meta->mm_psize = psize;
2346 meta->mm_last_pg = 1;
2347 meta->mm_flags = env->me_flags & 0xffff;
2348 meta->mm_flags |= MDB_INTEGERKEY;
2349 meta->mm_dbs[0].md_root = P_INVALID;
2350 meta->mm_dbs[1].md_root = P_INVALID;
2352 p = calloc(2, psize);
2354 p->mp_flags = P_META;
2357 memcpy(m, meta, sizeof(*meta));
2359 q = (MDB_page *)((char *)p + psize);
2362 q->mp_flags = P_META;
2365 memcpy(m, meta, sizeof(*meta));
2370 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2371 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2374 rc = write(env->me_fd, p, psize * 2);
2375 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2381 /** Update the environment info to commit a transaction.
2382 * @param[in] txn the transaction that's being committed
2383 * @return 0 on success, non-zero on failure.
2386 mdb_env_write_meta(MDB_txn *txn)
2389 MDB_meta meta, metab;
2391 int rc, len, toggle;
2397 assert(txn != NULL);
2398 assert(txn->mt_env != NULL);
2400 toggle = !txn->mt_toggle;
2401 DPRINTF("writing meta page %d for root page %zu",
2402 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2406 if (env->me_flags & MDB_WRITEMAP) {
2407 MDB_meta *mp = env->me_metas[toggle];
2408 mp->mm_dbs[0] = txn->mt_dbs[0];
2409 mp->mm_dbs[1] = txn->mt_dbs[1];
2410 mp->mm_last_pg = txn->mt_next_pgno - 1;
2411 mp->mm_txnid = txn->mt_txnid;
2412 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2413 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2416 ptr += env->me_psize;
2417 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2424 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2425 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2427 ptr = (char *)&meta;
2428 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2429 len = sizeof(MDB_meta) - off;
2432 meta.mm_dbs[0] = txn->mt_dbs[0];
2433 meta.mm_dbs[1] = txn->mt_dbs[1];
2434 meta.mm_last_pg = txn->mt_next_pgno - 1;
2435 meta.mm_txnid = txn->mt_txnid;
2438 off += env->me_psize;
2441 /* Write to the SYNC fd */
2444 memset(&ov, 0, sizeof(ov));
2446 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2449 rc = pwrite(env->me_mfd, ptr, len, off);
2454 DPUTS("write failed, disk error?");
2455 /* On a failure, the pagecache still contains the new data.
2456 * Write some old data back, to prevent it from being used.
2457 * Use the non-SYNC fd; we know it will fail anyway.
2459 meta.mm_last_pg = metab.mm_last_pg;
2460 meta.mm_txnid = metab.mm_txnid;
2462 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2464 r2 = pwrite(env->me_fd, ptr, len, off);
2467 env->me_flags |= MDB_FATAL_ERROR;
2471 /* Memory ordering issues are irrelevant; since the entire writer
2472 * is wrapped by wmutex, all of these changes will become visible
2473 * after the wmutex is unlocked. Since the DB is multi-version,
2474 * readers will get consistent data regardless of how fresh or
2475 * how stale their view of these values is.
2477 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2482 /** Check both meta pages to see which one is newer.
2483 * @param[in] env the environment handle
2484 * @return meta toggle (0 or 1).
2487 mdb_env_pick_meta(const MDB_env *env)
2489 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2493 mdb_env_create(MDB_env **env)
2497 e = calloc(1, sizeof(MDB_env));
2501 e->me_free_pgs = mdb_midl_alloc();
2502 if (!e->me_free_pgs) {
2506 e->me_maxreaders = DEFAULT_READERS;
2508 e->me_fd = INVALID_HANDLE_VALUE;
2509 e->me_lfd = INVALID_HANDLE_VALUE;
2510 e->me_mfd = INVALID_HANDLE_VALUE;
2511 VGMEMP_CREATE(e,0,0);
2517 mdb_env_set_mapsize(MDB_env *env, size_t size)
2521 env->me_mapsize = size;
2523 env->me_maxpg = env->me_mapsize / env->me_psize;
2528 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2532 env->me_maxdbs = dbs;
2537 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2539 if (env->me_map || readers < 1)
2541 env->me_maxreaders = readers;
2546 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2548 if (!env || !readers)
2550 *readers = env->me_maxreaders;
2554 /** Further setup required for opening an MDB environment
2557 mdb_env_open2(MDB_env *env, unsigned int flags)
2559 int i, newenv = 0, prot;
2563 env->me_flags = flags;
2565 memset(&meta, 0, sizeof(meta));
2567 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2570 DPUTS("new mdbenv");
2574 if (!env->me_mapsize) {
2575 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2581 LONG sizelo, sizehi;
2582 sizelo = env->me_mapsize & 0xffffffff;
2583 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2585 /* Windows won't create mappings for zero length files.
2586 * Just allocate the maxsize right now.
2589 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2590 if (!SetEndOfFile(env->me_fd))
2592 SetFilePointer(env->me_fd, 0, NULL, 0);
2594 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2595 PAGE_READWRITE : PAGE_READONLY,
2596 sizehi, sizelo, NULL);
2599 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2600 FILE_MAP_WRITE : FILE_MAP_READ,
2601 0, 0, env->me_mapsize, meta.mm_address);
2608 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2611 if (flags & MDB_WRITEMAP) {
2613 ftruncate(env->me_fd, env->me_mapsize);
2615 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2617 if (env->me_map == MAP_FAILED) {
2624 meta.mm_mapsize = env->me_mapsize;
2625 if (flags & MDB_FIXEDMAP)
2626 meta.mm_address = env->me_map;
2627 i = mdb_env_init_meta(env, &meta);
2628 if (i != MDB_SUCCESS) {
2629 munmap(env->me_map, env->me_mapsize);
2633 env->me_psize = meta.mm_psize;
2635 env->me_maxpg = env->me_mapsize / env->me_psize;
2637 p = (MDB_page *)env->me_map;
2638 env->me_metas[0] = METADATA(p);
2639 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2643 int toggle = mdb_env_pick_meta(env);
2644 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2646 DPRINTF("opened database version %u, pagesize %u",
2647 env->me_metas[0]->mm_version, env->me_psize);
2648 DPRINTF("using meta page %d", toggle);
2649 DPRINTF("depth: %u", db->md_depth);
2650 DPRINTF("entries: %zu", db->md_entries);
2651 DPRINTF("branch pages: %zu", db->md_branch_pages);
2652 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2653 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2654 DPRINTF("root: %zu", db->md_root);
2662 /** Release a reader thread's slot in the reader lock table.
2663 * This function is called automatically when a thread exits.
2664 * @param[in] ptr This points to the slot in the reader lock table.
2667 mdb_env_reader_dest(void *ptr)
2669 MDB_reader *reader = ptr;
2671 reader->mr_txnid = 0;
2677 /** Junk for arranging thread-specific callbacks on Windows. This is
2678 * necessarily platform and compiler-specific. Windows supports up
2679 * to 1088 keys. Let's assume nobody opens more than 64 environments
2680 * in a single process, for now. They can override this if needed.
2682 #ifndef MAX_TLS_KEYS
2683 #define MAX_TLS_KEYS 64
2685 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2686 static int mdb_tls_nkeys;
2688 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2692 case DLL_PROCESS_ATTACH: break;
2693 case DLL_THREAD_ATTACH: break;
2694 case DLL_THREAD_DETACH:
2695 for (i=0; i<mdb_tls_nkeys; i++) {
2696 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2697 mdb_env_reader_dest(r);
2700 case DLL_PROCESS_DETACH: break;
2705 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2707 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2711 /* Force some symbol references.
2712 * _tls_used forces the linker to create the TLS directory if not already done
2713 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2715 #pragma comment(linker, "/INCLUDE:_tls_used")
2716 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2717 #pragma const_seg(".CRT$XLB")
2718 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2719 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2722 #pragma comment(linker, "/INCLUDE:__tls_used")
2723 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2724 #pragma data_seg(".CRT$XLB")
2725 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2727 #endif /* WIN 32/64 */
2728 #endif /* !__GNUC__ */
2731 /** Downgrade the exclusive lock on the region back to shared */
2733 mdb_env_share_locks(MDB_env *env)
2735 int toggle = mdb_env_pick_meta(env);
2737 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2742 /* First acquire a shared lock. The Unlock will
2743 * then release the existing exclusive lock.
2745 memset(&ov, 0, sizeof(ov));
2746 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2747 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2751 struct flock lock_info;
2752 /* The shared lock replaces the existing lock */
2753 memset((void *)&lock_info, 0, sizeof(lock_info));
2754 lock_info.l_type = F_RDLCK;
2755 lock_info.l_whence = SEEK_SET;
2756 lock_info.l_start = 0;
2757 lock_info.l_len = 1;
2758 fcntl(env->me_lfd, F_SETLK, &lock_info);
2764 mdb_env_excl_lock(MDB_env *env, int *excl)
2767 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2771 memset(&ov, 0, sizeof(ov));
2772 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2777 struct flock lock_info;
2778 memset((void *)&lock_info, 0, sizeof(lock_info));
2779 lock_info.l_type = F_WRLCK;
2780 lock_info.l_whence = SEEK_SET;
2781 lock_info.l_start = 0;
2782 lock_info.l_len = 1;
2783 if (!fcntl(env->me_lfd, F_SETLK, &lock_info)) {
2786 lock_info.l_type = F_RDLCK;
2787 if (fcntl(env->me_lfd, F_SETLKW, &lock_info)) {
2795 #if defined(_WIN32) || defined(USE_POSIX_SEM)
2797 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2799 * @(#) $Revision: 5.1 $
2800 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2801 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2803 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2807 * Please do not copyright this code. This code is in the public domain.
2809 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2810 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2811 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2812 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2813 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2814 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2815 * PERFORMANCE OF THIS SOFTWARE.
2818 * chongo <Landon Curt Noll> /\oo/\
2819 * http://www.isthe.com/chongo/
2821 * Share and Enjoy! :-)
2824 typedef unsigned long long mdb_hash_t;
2825 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2827 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2828 * @param[in] str string to hash
2829 * @param[in] hval initial value for hash
2830 * @return 64 bit hash
2832 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2833 * hval arg on the first call.
2836 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
2838 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
2839 unsigned char *end = s + val->mv_size;
2841 * FNV-1a hash each octet of the string
2844 /* xor the bottom with the current octet */
2845 hval ^= (mdb_hash_t)*s++;
2847 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2848 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2849 (hval << 7) + (hval << 8) + (hval << 40);
2851 /* return our new hash value */
2855 /** Hash the string and output the hash in hex.
2856 * @param[in] str string to hash
2857 * @param[out] hexbuf an array of 17 chars to hold the hash
2860 mdb_hash_hex(MDB_val *val, char *hexbuf)
2863 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
2864 for (i=0; i<8; i++) {
2865 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2871 /** Open and/or initialize the lock region for the environment.
2872 * @param[in] env The MDB environment.
2873 * @param[in] lpath The pathname of the file used for the lock region.
2874 * @param[in] mode The Unix permissions for the file, if we create it.
2875 * @param[out] excl Set to true if we got an exclusive lock on the region.
2876 * @return 0 on success, non-zero on failure.
2879 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2887 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2888 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2889 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2893 /* Try to get exclusive lock. If we succeed, then
2894 * nobody is using the lock region and we should initialize it.
2896 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2897 size = GetFileSize(env->me_lfd, NULL);
2903 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1)
2905 /* Lose record locks when exec*() */
2906 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
2907 fcntl(env->me_lfd, F_SETFD, fdflags);
2909 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
2910 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1)
2914 /* Try to get exclusive lock. If we succeed, then
2915 * nobody is using the lock region and we should initialize it.
2917 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2919 size = lseek(env->me_lfd, 0, SEEK_END);
2921 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2922 if (size < rsize && *excl) {
2924 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2925 if (!SetEndOfFile(env->me_lfd)) {
2930 if (ftruncate(env->me_lfd, rsize) != 0) {
2937 size = rsize - sizeof(MDB_txninfo);
2938 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2943 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2949 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2951 if (!env->me_txns) {
2956 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2958 if (m == MAP_FAILED) {
2959 env->me_txns = NULL;
2968 BY_HANDLE_FILE_INFORMATION stbuf;
2977 if (!mdb_sec_inited) {
2978 InitializeSecurityDescriptor(&mdb_null_sd,
2979 SECURITY_DESCRIPTOR_REVISION);
2980 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2981 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2982 mdb_all_sa.bInheritHandle = FALSE;
2983 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2986 GetFileInformationByHandle(env->me_lfd, &stbuf);
2987 idbuf.volume = stbuf.dwVolumeSerialNumber;
2988 idbuf.nhigh = stbuf.nFileIndexHigh;
2989 idbuf.nlow = stbuf.nFileIndexLow;
2990 val.mv_data = &idbuf;
2991 val.mv_size = sizeof(idbuf);
2992 mdb_hash_hex(&val, hexbuf);
2993 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
2994 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2995 if (!env->me_rmutex) {
2999 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3000 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3001 if (!env->me_wmutex) {
3006 #ifdef USE_POSIX_SEM
3015 fstat(env->me_lfd, &stbuf);
3016 idbuf.dev = stbuf.st_dev;
3017 idbuf.ino = stbuf.st_ino;
3018 val.mv_data = &idbuf;
3019 val.mv_size = sizeof(idbuf);
3020 mdb_hash_hex(&val, hexbuf);
3021 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3022 if (sem_unlink(env->me_txns->mti_rmname)) {
3024 if (rc != ENOENT && rc != EINVAL)
3027 env->me_rmutex = sem_open(env->me_txns->mti_rmname, O_CREAT, mode, 1);
3028 if (env->me_rmutex == SEM_FAILED) {
3032 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3033 if (sem_unlink(env->me_txns->mti_wmname)) {
3035 if (rc != ENOENT && rc != EINVAL)
3038 env->me_wmutex = sem_open(env->me_txns->mti_wmname, O_CREAT, mode, 1);
3039 if (env->me_wmutex == SEM_FAILED) {
3043 #else /* USE_POSIX_SEM */
3044 pthread_mutexattr_t mattr;
3046 pthread_mutexattr_init(&mattr);
3047 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
3051 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
3052 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
3053 #endif /* USE_POSIX_SEM */
3055 env->me_txns->mti_version = MDB_VERSION;
3056 env->me_txns->mti_magic = MDB_MAGIC;
3057 env->me_txns->mti_txnid = 0;
3058 env->me_txns->mti_numreaders = 0;
3061 if (env->me_txns->mti_magic != MDB_MAGIC) {
3062 DPUTS("lock region has invalid magic");
3066 if (env->me_txns->mti_version != MDB_VERSION) {
3067 DPRINTF("lock region is version %u, expected version %u",
3068 env->me_txns->mti_version, MDB_VERSION);
3069 rc = MDB_VERSION_MISMATCH;
3073 if (rc != EACCES && rc != EAGAIN) {
3077 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3078 if (!env->me_rmutex) {
3082 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3083 if (!env->me_wmutex) {
3088 #ifdef USE_POSIX_SEM
3089 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3090 if (env->me_rmutex == SEM_FAILED) {
3094 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3095 if (env->me_wmutex == SEM_FAILED) {
3105 env->me_lfd = INVALID_HANDLE_VALUE;
3110 /** The name of the lock file in the DB environment */
3111 #define LOCKNAME "/lock.mdb"
3112 /** The name of the data file in the DB environment */
3113 #define DATANAME "/data.mdb"
3114 /** The suffix of the lock file when no subdir is used */
3115 #define LOCKSUFF "-lock"
3118 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3120 int oflags, rc, len, excl;
3121 char *lpath, *dpath;
3124 if (flags & MDB_NOSUBDIR) {
3125 rc = len + sizeof(LOCKSUFF) + len + 1;
3127 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3132 if (flags & MDB_NOSUBDIR) {
3133 dpath = lpath + len + sizeof(LOCKSUFF);
3134 sprintf(lpath, "%s" LOCKSUFF, path);
3135 strcpy(dpath, path);
3137 dpath = lpath + len + sizeof(LOCKNAME);
3138 sprintf(lpath, "%s" LOCKNAME, path);
3139 sprintf(dpath, "%s" DATANAME, path);
3142 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3147 if (F_ISSET(flags, MDB_RDONLY)) {
3148 oflags = GENERIC_READ;
3149 len = OPEN_EXISTING;
3151 oflags = GENERIC_READ|GENERIC_WRITE;
3154 mode = FILE_ATTRIBUTE_NORMAL;
3155 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3156 NULL, len, mode, NULL);
3158 if (F_ISSET(flags, MDB_RDONLY))
3161 oflags = O_RDWR | O_CREAT;
3163 env->me_fd = open(dpath, oflags, mode);
3165 if (env->me_fd == INVALID_HANDLE_VALUE) {
3170 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
3171 if (flags & (MDB_RDONLY|MDB_NOSYNC|MDB_NOMETASYNC)) {
3172 env->me_mfd = env->me_fd;
3174 /* synchronous fd for meta writes */
3176 env->me_mfd = CreateFile(dpath, oflags,
3177 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3178 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3180 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3182 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3187 env->me_path = strdup(path);
3188 DPRINTF("opened dbenv %p", (void *) env);
3189 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3191 /* Windows TLS callbacks need help finding their TLS info. */
3192 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3193 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3200 mdb_env_share_locks(env);
3202 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3203 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3204 if (!env->me_dbxs || !env->me_dbflags)
3210 if (env->me_fd != INVALID_HANDLE_VALUE) {
3212 env->me_fd = INVALID_HANDLE_VALUE;
3214 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3216 env->me_lfd = INVALID_HANDLE_VALUE;
3224 mdb_env_close(MDB_env *env)
3231 VGMEMP_DESTROY(env);
3232 while (env->me_dpages) {
3233 dp = env->me_dpages;
3234 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3235 env->me_dpages = dp->mp_next;
3239 free(env->me_dbflags);
3243 pthread_key_delete(env->me_txkey);
3245 /* Delete our key from the global list */
3247 for (i=0; i<mdb_tls_nkeys; i++)
3248 if (mdb_tls_keys[i] == env->me_txkey) {
3249 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3257 munmap(env->me_map, env->me_mapsize);
3259 if (env->me_mfd != env->me_fd)
3263 pid_t pid = getpid();
3265 for (i=0; i<env->me_txns->mti_numreaders; i++)
3266 if (env->me_txns->mti_readers[i].mr_pid == pid)
3267 env->me_txns->mti_readers[i].mr_pid = 0;
3269 CloseHandle(env->me_rmutex);
3270 CloseHandle(env->me_wmutex);
3271 /* Windows automatically destroys the mutexes when
3272 * the last handle closes.
3275 #ifdef USE_POSIX_SEM
3276 sem_close(env->me_rmutex);
3277 sem_close(env->me_wmutex);
3279 if (!mdb_env_excl_lock(env, &excl) && excl) {
3280 /* we are the only remaining user of the environment.
3281 clean up semaphores. */
3282 sem_unlink(env->me_txns->mti_rmname);
3283 sem_unlink(env->me_txns->mti_wmname);
3288 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3291 mdb_midl_free(env->me_free_pgs);
3295 /** Compare two items pointing at aligned size_t's */
3297 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3299 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3300 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3303 /** Compare two items pointing at aligned int's */
3305 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3307 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3308 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3311 /** Compare two items pointing at ints of unknown alignment.
3312 * Nodes and keys are guaranteed to be 2-byte aligned.
3315 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3317 #if BYTE_ORDER == LITTLE_ENDIAN
3318 unsigned short *u, *c;
3321 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3322 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3325 } while(!x && u > (unsigned short *)a->mv_data);
3328 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3332 /** Compare two items lexically */
3334 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3341 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3347 diff = memcmp(a->mv_data, b->mv_data, len);
3348 return diff ? diff : len_diff<0 ? -1 : len_diff;
3351 /** Compare two items in reverse byte order */
3353 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3355 const unsigned char *p1, *p2, *p1_lim;
3359 p1_lim = (const unsigned char *)a->mv_data;
3360 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3361 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3363 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3369 while (p1 > p1_lim) {
3370 diff = *--p1 - *--p2;
3374 return len_diff<0 ? -1 : len_diff;
3377 /** Search for key within a page, using binary search.
3378 * Returns the smallest entry larger or equal to the key.
3379 * If exactp is non-null, stores whether the found entry was an exact match
3380 * in *exactp (1 or 0).
3381 * Updates the cursor index with the index of the found entry.
3382 * If no entry larger or equal to the key is found, returns NULL.
3385 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3387 unsigned int i = 0, nkeys;
3390 MDB_page *mp = mc->mc_pg[mc->mc_top];
3391 MDB_node *node = NULL;
3396 nkeys = NUMKEYS(mp);
3401 COPY_PGNO(pgno, mp->mp_pgno);
3402 DPRINTF("searching %u keys in %s %spage %zu",
3403 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3410 low = IS_LEAF(mp) ? 0 : 1;
3412 cmp = mc->mc_dbx->md_cmp;
3414 /* Branch pages have no data, so if using integer keys,
3415 * alignment is guaranteed. Use faster mdb_cmp_int.
3417 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3418 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3425 nodekey.mv_size = mc->mc_db->md_pad;
3426 node = NODEPTR(mp, 0); /* fake */
3427 while (low <= high) {
3428 i = (low + high) >> 1;
3429 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3430 rc = cmp(key, &nodekey);
3431 DPRINTF("found leaf index %u [%s], rc = %i",
3432 i, DKEY(&nodekey), rc);
3441 while (low <= high) {
3442 i = (low + high) >> 1;
3444 node = NODEPTR(mp, i);
3445 nodekey.mv_size = NODEKSZ(node);
3446 nodekey.mv_data = NODEKEY(node);
3448 rc = cmp(key, &nodekey);
3451 DPRINTF("found leaf index %u [%s], rc = %i",
3452 i, DKEY(&nodekey), rc);
3454 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3455 i, DKEY(&nodekey), NODEPGNO(node), rc);
3466 if (rc > 0) { /* Found entry is less than the key. */
3467 i++; /* Skip to get the smallest entry larger than key. */
3469 node = NODEPTR(mp, i);
3472 *exactp = (rc == 0);
3473 /* store the key index */
3474 mc->mc_ki[mc->mc_top] = i;
3476 /* There is no entry larger or equal to the key. */
3479 /* nodeptr is fake for LEAF2 */
3485 mdb_cursor_adjust(MDB_cursor *mc, func)
3489 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3490 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3497 /** Pop a page off the top of the cursor's stack. */
3499 mdb_cursor_pop(MDB_cursor *mc)
3503 MDB_page *top = mc->mc_pg[mc->mc_top];
3509 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3510 mc->mc_dbi, (void *) mc);
3514 /** Push a page onto the top of the cursor's stack. */
3516 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3518 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3519 mc->mc_dbi, (void *) mc);
3521 if (mc->mc_snum >= CURSOR_STACK) {
3522 assert(mc->mc_snum < CURSOR_STACK);
3523 return MDB_CURSOR_FULL;
3526 mc->mc_top = mc->mc_snum++;
3527 mc->mc_pg[mc->mc_top] = mp;
3528 mc->mc_ki[mc->mc_top] = 0;
3533 /** Find the address of the page corresponding to a given page number.
3534 * @param[in] txn the transaction for this access.
3535 * @param[in] pgno the page number for the page to retrieve.
3536 * @param[out] ret address of a pointer where the page's address will be stored.
3537 * @return 0 on success, non-zero on failure.
3540 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3544 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
3545 if (pgno < txn->mt_next_pgno)
3546 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3549 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3551 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3552 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3553 p = txn->mt_u.dirty_list[x].mptr;
3557 if (pgno < txn->mt_next_pgno)
3558 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3563 DPRINTF("page %zu not found", pgno);
3566 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3569 /** Search for the page a given key should be in.
3570 * Pushes parent pages on the cursor stack. This function continues a
3571 * search on a cursor that has already been initialized. (Usually by
3572 * #mdb_page_search() but also by #mdb_node_move().)
3573 * @param[in,out] mc the cursor for this operation.
3574 * @param[in] key the key to search for. If NULL, search for the lowest
3575 * page. (This is used by #mdb_cursor_first().)
3576 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3577 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3578 * @return 0 on success, non-zero on failure.
3581 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3583 MDB_page *mp = mc->mc_pg[mc->mc_top];
3588 while (IS_BRANCH(mp)) {
3592 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3593 assert(NUMKEYS(mp) > 1);
3594 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3596 if (key == NULL) /* Initialize cursor to first page. */
3598 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3599 /* cursor to last page */
3603 node = mdb_node_search(mc, key, &exact);
3605 i = NUMKEYS(mp) - 1;
3607 i = mc->mc_ki[mc->mc_top];
3616 DPRINTF("following index %u for key [%s]",
3618 assert(i < NUMKEYS(mp));
3619 node = NODEPTR(mp, i);
3621 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3624 mc->mc_ki[mc->mc_top] = i;
3625 if ((rc = mdb_cursor_push(mc, mp)))
3629 if ((rc = mdb_page_touch(mc)) != 0)
3631 mp = mc->mc_pg[mc->mc_top];
3636 DPRINTF("internal error, index points to a %02X page!?",
3638 return MDB_CORRUPTED;
3641 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3642 key ? DKEY(key) : NULL);
3647 /** Search for the page a given key should be in.
3648 * Pushes parent pages on the cursor stack. This function just sets up
3649 * the search; it finds the root page for \b mc's database and sets this
3650 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3651 * called to complete the search.
3652 * @param[in,out] mc the cursor for this operation.
3653 * @param[in] key the key to search for. If NULL, search for the lowest
3654 * page. (This is used by #mdb_cursor_first().)
3655 * @param[in] modify If true, visited pages are updated with new page numbers.
3656 * @return 0 on success, non-zero on failure.
3659 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
3664 /* Make sure the txn is still viable, then find the root from
3665 * the txn's db table.
3667 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3668 DPUTS("transaction has failed, must abort");
3671 /* Make sure we're using an up-to-date root */
3672 if (mc->mc_dbi > MAIN_DBI) {
3673 if ((*mc->mc_dbflag & DB_STALE) ||
3674 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
3676 unsigned char dbflag = 0;
3677 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3678 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
3681 if (*mc->mc_dbflag & DB_STALE) {
3684 MDB_node *leaf = mdb_node_search(&mc2,
3685 &mc->mc_dbx->md_name, &exact);
3687 return MDB_NOTFOUND;
3688 mdb_node_read(mc->mc_txn, leaf, &data);
3689 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3691 if (flags & MDB_PS_MODIFY)
3693 *mc->mc_dbflag = dbflag;
3696 root = mc->mc_db->md_root;
3698 if (root == P_INVALID) { /* Tree is empty. */
3699 DPUTS("tree is empty");
3700 return MDB_NOTFOUND;
3705 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
3706 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3712 DPRINTF("db %u root page %zu has flags 0x%X",
3713 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3715 if (flags & MDB_PS_MODIFY) {
3716 if ((rc = mdb_page_touch(mc)))
3720 if (flags & MDB_PS_ROOTONLY)
3723 return mdb_page_search_root(mc, key, flags);
3726 /** Return the data associated with a given node.
3727 * @param[in] txn The transaction for this operation.
3728 * @param[in] leaf The node being read.
3729 * @param[out] data Updated to point to the node's data.
3730 * @return 0 on success, non-zero on failure.
3733 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3735 MDB_page *omp; /* overflow page */
3739 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3740 data->mv_size = NODEDSZ(leaf);
3741 data->mv_data = NODEDATA(leaf);
3745 /* Read overflow data.
3747 data->mv_size = NODEDSZ(leaf);
3748 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3749 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3750 DPRINTF("read overflow page %zu failed", pgno);
3753 data->mv_data = METADATA(omp);
3759 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3760 MDB_val *key, MDB_val *data)
3769 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3771 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3774 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3778 mdb_cursor_init(&mc, txn, dbi, &mx);
3779 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3782 /** Find a sibling for a page.
3783 * Replaces the page at the top of the cursor's stack with the
3784 * specified sibling, if one exists.
3785 * @param[in] mc The cursor for this operation.
3786 * @param[in] move_right Non-zero if the right sibling is requested,
3787 * otherwise the left sibling.
3788 * @return 0 on success, non-zero on failure.
3791 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3797 if (mc->mc_snum < 2) {
3798 return MDB_NOTFOUND; /* root has no siblings */
3802 DPRINTF("parent page is page %zu, index %u",
3803 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3805 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3806 : (mc->mc_ki[mc->mc_top] == 0)) {
3807 DPRINTF("no more keys left, moving to %s sibling",
3808 move_right ? "right" : "left");
3809 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3813 mc->mc_ki[mc->mc_top]++;
3815 mc->mc_ki[mc->mc_top]--;
3816 DPRINTF("just moving to %s index key %u",
3817 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3819 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3821 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3822 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3825 mdb_cursor_push(mc, mp);
3830 /** Move the cursor to the next data item. */
3832 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3838 if (mc->mc_flags & C_EOF) {
3839 return MDB_NOTFOUND;
3842 assert(mc->mc_flags & C_INITIALIZED);
3844 mp = mc->mc_pg[mc->mc_top];
3846 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3847 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3848 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3849 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3850 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3851 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3855 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3856 if (op == MDB_NEXT_DUP)
3857 return MDB_NOTFOUND;
3861 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3863 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3864 DPUTS("=====> move to next sibling page");
3865 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3866 mc->mc_flags |= C_EOF;
3867 mc->mc_flags &= ~C_INITIALIZED;
3868 return MDB_NOTFOUND;
3870 mp = mc->mc_pg[mc->mc_top];
3871 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3873 mc->mc_ki[mc->mc_top]++;
3875 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3876 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3879 key->mv_size = mc->mc_db->md_pad;
3880 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3884 assert(IS_LEAF(mp));
3885 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3887 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3888 mdb_xcursor_init1(mc, leaf);
3891 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3894 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3895 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3896 if (rc != MDB_SUCCESS)
3901 MDB_GET_KEY(leaf, key);
3905 /** Move the cursor to the previous data item. */
3907 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3913 assert(mc->mc_flags & C_INITIALIZED);
3915 mp = mc->mc_pg[mc->mc_top];
3917 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3918 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3919 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3920 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3921 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3922 if (op != MDB_PREV || rc == MDB_SUCCESS)
3925 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3926 if (op == MDB_PREV_DUP)
3927 return MDB_NOTFOUND;
3932 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3934 if (mc->mc_ki[mc->mc_top] == 0) {
3935 DPUTS("=====> move to prev sibling page");
3936 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3937 mc->mc_flags &= ~C_INITIALIZED;
3938 return MDB_NOTFOUND;
3940 mp = mc->mc_pg[mc->mc_top];
3941 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3942 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3944 mc->mc_ki[mc->mc_top]--;
3946 mc->mc_flags &= ~C_EOF;
3948 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3949 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3952 key->mv_size = mc->mc_db->md_pad;
3953 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3957 assert(IS_LEAF(mp));
3958 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3960 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3961 mdb_xcursor_init1(mc, leaf);
3964 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3967 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3968 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3969 if (rc != MDB_SUCCESS)
3974 MDB_GET_KEY(leaf, key);
3978 /** Set the cursor on a specific data item. */
3980 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3981 MDB_cursor_op op, int *exactp)
3985 MDB_node *leaf = NULL;
3990 assert(key->mv_size > 0);
3992 /* See if we're already on the right page */
3993 if (mc->mc_flags & C_INITIALIZED) {
3996 mp = mc->mc_pg[mc->mc_top];
3998 mc->mc_ki[mc->mc_top] = 0;
3999 return MDB_NOTFOUND;
4001 if (mp->mp_flags & P_LEAF2) {
4002 nodekey.mv_size = mc->mc_db->md_pad;
4003 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4005 leaf = NODEPTR(mp, 0);
4006 MDB_GET_KEY(leaf, &nodekey);
4008 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4010 /* Probably happens rarely, but first node on the page
4011 * was the one we wanted.
4013 mc->mc_ki[mc->mc_top] = 0;
4020 unsigned int nkeys = NUMKEYS(mp);
4022 if (mp->mp_flags & P_LEAF2) {
4023 nodekey.mv_data = LEAF2KEY(mp,
4024 nkeys-1, nodekey.mv_size);
4026 leaf = NODEPTR(mp, nkeys-1);
4027 MDB_GET_KEY(leaf, &nodekey);
4029 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4031 /* last node was the one we wanted */
4032 mc->mc_ki[mc->mc_top] = nkeys-1;
4038 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4039 /* This is definitely the right page, skip search_page */
4040 if (mp->mp_flags & P_LEAF2) {
4041 nodekey.mv_data = LEAF2KEY(mp,
4042 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4044 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4045 MDB_GET_KEY(leaf, &nodekey);
4047 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4049 /* current node was the one we wanted */
4059 /* If any parents have right-sibs, search.
4060 * Otherwise, there's nothing further.
4062 for (i=0; i<mc->mc_top; i++)
4064 NUMKEYS(mc->mc_pg[i])-1)
4066 if (i == mc->mc_top) {
4067 /* There are no other pages */
4068 mc->mc_ki[mc->mc_top] = nkeys;
4069 return MDB_NOTFOUND;
4073 /* There are no other pages */
4074 mc->mc_ki[mc->mc_top] = 0;
4075 return MDB_NOTFOUND;
4079 rc = mdb_page_search(mc, key, 0);
4080 if (rc != MDB_SUCCESS)
4083 mp = mc->mc_pg[mc->mc_top];
4084 assert(IS_LEAF(mp));
4087 leaf = mdb_node_search(mc, key, exactp);
4088 if (exactp != NULL && !*exactp) {
4089 /* MDB_SET specified and not an exact match. */
4090 return MDB_NOTFOUND;
4094 DPUTS("===> inexact leaf not found, goto sibling");
4095 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4096 return rc; /* no entries matched */
4097 mp = mc->mc_pg[mc->mc_top];
4098 assert(IS_LEAF(mp));
4099 leaf = NODEPTR(mp, 0);
4103 mc->mc_flags |= C_INITIALIZED;
4104 mc->mc_flags &= ~C_EOF;
4107 key->mv_size = mc->mc_db->md_pad;
4108 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4112 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4113 mdb_xcursor_init1(mc, leaf);
4116 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4117 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4118 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4121 if (op == MDB_GET_BOTH) {
4127 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4128 if (rc != MDB_SUCCESS)
4131 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4133 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4135 rc = mc->mc_dbx->md_dcmp(data, &d2);
4137 if (op == MDB_GET_BOTH || rc > 0)
4138 return MDB_NOTFOUND;
4143 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4144 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4149 /* The key already matches in all other cases */
4150 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4151 MDB_GET_KEY(leaf, key);
4152 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4157 /** Move the cursor to the first item in the database. */
4159 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4164 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4165 rc = mdb_page_search(mc, NULL, 0);
4166 if (rc != MDB_SUCCESS)
4169 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4171 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4172 mc->mc_flags |= C_INITIALIZED;
4173 mc->mc_flags &= ~C_EOF;
4175 mc->mc_ki[mc->mc_top] = 0;
4177 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4178 key->mv_size = mc->mc_db->md_pad;
4179 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4184 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4185 mdb_xcursor_init1(mc, leaf);
4186 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4191 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4192 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4196 MDB_GET_KEY(leaf, key);
4200 /** Move the cursor to the last item in the database. */
4202 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4207 if (!(mc->mc_flags & C_EOF)) {
4209 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4212 lkey.mv_size = MAXKEYSIZE+1;
4213 lkey.mv_data = NULL;
4214 rc = mdb_page_search(mc, &lkey, 0);
4215 if (rc != MDB_SUCCESS)
4218 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4220 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4221 mc->mc_flags |= C_INITIALIZED|C_EOF;
4223 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4225 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4226 key->mv_size = mc->mc_db->md_pad;
4227 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4232 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4233 mdb_xcursor_init1(mc, leaf);
4234 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4239 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4240 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4245 MDB_GET_KEY(leaf, key);
4250 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4260 case MDB_GET_BOTH_RANGE:
4261 if (data == NULL || mc->mc_xcursor == NULL) {
4269 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4271 } else if (op == MDB_SET_RANGE)
4272 rc = mdb_cursor_set(mc, key, data, op, NULL);
4274 rc = mdb_cursor_set(mc, key, data, op, &exact);
4276 case MDB_GET_MULTIPLE:
4278 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4279 !(mc->mc_flags & C_INITIALIZED)) {
4284 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4285 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4288 case MDB_NEXT_MULTIPLE:
4290 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4294 if (!(mc->mc_flags & C_INITIALIZED))
4295 rc = mdb_cursor_first(mc, key, data);
4297 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4298 if (rc == MDB_SUCCESS) {
4299 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4302 mx = &mc->mc_xcursor->mx_cursor;
4303 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4305 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4306 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4314 case MDB_NEXT_NODUP:
4315 if (!(mc->mc_flags & C_INITIALIZED))
4316 rc = mdb_cursor_first(mc, key, data);
4318 rc = mdb_cursor_next(mc, key, data, op);
4322 case MDB_PREV_NODUP:
4323 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4324 rc = mdb_cursor_last(mc, key, data);
4325 mc->mc_flags &= ~C_EOF;
4327 rc = mdb_cursor_prev(mc, key, data, op);
4330 rc = mdb_cursor_first(mc, key, data);
4334 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4335 !(mc->mc_flags & C_INITIALIZED) ||
4336 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4340 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4343 rc = mdb_cursor_last(mc, key, data);
4347 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4348 !(mc->mc_flags & C_INITIALIZED) ||
4349 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4353 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4356 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4364 /** Touch all the pages in the cursor stack.
4365 * Makes sure all the pages are writable, before attempting a write operation.
4366 * @param[in] mc The cursor to operate on.
4369 mdb_cursor_touch(MDB_cursor *mc)
4373 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4375 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4376 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4379 *mc->mc_dbflag = DB_DIRTY;
4381 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4382 rc = mdb_page_touch(mc);
4386 mc->mc_top = mc->mc_snum-1;
4391 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4394 MDB_node *leaf = NULL;
4395 MDB_val xdata, *rdata, dkey;
4398 int do_sub = 0, insert = 0;
4399 unsigned int mcount = 0;
4403 char dbuf[MAXKEYSIZE+1];
4404 unsigned int nflags;
4407 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4410 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4411 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4415 if (flags == MDB_CURRENT) {
4416 if (!(mc->mc_flags & C_INITIALIZED))
4419 } else if (mc->mc_db->md_root == P_INVALID) {
4421 /* new database, write a root leaf page */
4422 DPUTS("allocating new root leaf page");
4423 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4427 mdb_cursor_push(mc, np);
4428 mc->mc_db->md_root = np->mp_pgno;
4429 mc->mc_db->md_depth++;
4430 *mc->mc_dbflag = DB_DIRTY;
4431 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4433 np->mp_flags |= P_LEAF2;
4434 mc->mc_flags |= C_INITIALIZED;
4440 if (flags & MDB_APPEND) {
4442 rc = mdb_cursor_last(mc, &k2, &d2);
4444 rc = mc->mc_dbx->md_cmp(key, &k2);
4447 mc->mc_ki[mc->mc_top]++;
4453 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4455 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4456 DPRINTF("duplicate key [%s]", DKEY(key));
4458 return MDB_KEYEXIST;
4460 if (rc && rc != MDB_NOTFOUND)
4464 /* Cursor is positioned, now make sure all pages are writable */
4465 rc2 = mdb_cursor_touch(mc);
4470 /* The key already exists */
4471 if (rc == MDB_SUCCESS) {
4472 /* there's only a key anyway, so this is a no-op */
4473 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4474 unsigned int ksize = mc->mc_db->md_pad;
4475 if (key->mv_size != ksize)
4477 if (flags == MDB_CURRENT) {
4478 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4479 memcpy(ptr, key->mv_data, ksize);
4484 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4487 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4488 /* Was a single item before, must convert now */
4490 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4491 /* Just overwrite the current item */
4492 if (flags == MDB_CURRENT)
4495 dkey.mv_size = NODEDSZ(leaf);
4496 dkey.mv_data = NODEDATA(leaf);
4497 #if UINT_MAX < SIZE_MAX
4498 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4499 #ifdef MISALIGNED_OK
4500 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4502 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4505 /* if data matches, ignore it */
4506 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4507 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4509 /* create a fake page for the dup items */
4510 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4511 dkey.mv_data = dbuf;
4512 fp = (MDB_page *)&pbuf;
4513 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4514 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4515 fp->mp_lower = PAGEHDRSZ;
4516 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4517 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4518 fp->mp_flags |= P_LEAF2;
4519 fp->mp_pad = data->mv_size;
4520 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4522 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4523 (dkey.mv_size & 1) + (data->mv_size & 1);
4525 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4528 xdata.mv_size = fp->mp_upper;
4533 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4534 /* See if we need to convert from fake page to subDB */
4536 unsigned int offset;
4539 fp = NODEDATA(leaf);
4540 if (flags == MDB_CURRENT) {
4542 fp->mp_flags |= P_DIRTY;
4543 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4544 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4548 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4549 offset = fp->mp_pad;
4550 if (SIZELEFT(fp) >= offset)
4552 offset *= 4; /* space for 4 more */
4554 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4556 offset += offset & 1;
4557 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4558 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4560 /* yes, convert it */
4562 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4563 dummy.md_pad = fp->mp_pad;
4564 dummy.md_flags = MDB_DUPFIXED;
4565 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4566 dummy.md_flags |= MDB_INTEGERKEY;
4569 dummy.md_branch_pages = 0;
4570 dummy.md_leaf_pages = 1;
4571 dummy.md_overflow_pages = 0;
4572 dummy.md_entries = NUMKEYS(fp);
4574 xdata.mv_size = sizeof(MDB_db);
4575 xdata.mv_data = &dummy;
4576 if ((rc = mdb_page_alloc(mc, 1, &mp)))
4578 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4579 flags |= F_DUPDATA|F_SUBDATA;
4580 dummy.md_root = mp->mp_pgno;
4582 /* no, just grow it */
4584 xdata.mv_size = NODEDSZ(leaf) + offset;
4585 xdata.mv_data = &pbuf;
4586 mp = (MDB_page *)&pbuf;
4587 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4590 mp->mp_flags = fp->mp_flags | P_DIRTY;
4591 mp->mp_pad = fp->mp_pad;
4592 mp->mp_lower = fp->mp_lower;
4593 mp->mp_upper = fp->mp_upper + offset;
4595 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4597 nsize = NODEDSZ(leaf) - fp->mp_upper;
4598 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4599 for (i=0; i<NUMKEYS(fp); i++)
4600 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4602 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4606 /* data is on sub-DB, just store it */
4607 flags |= F_DUPDATA|F_SUBDATA;
4611 /* overflow page overwrites need special handling */
4612 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4615 int ovpages, dpages;
4617 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4618 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4619 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4620 mdb_page_get(mc->mc_txn, pg, &omp);
4621 /* Is the ov page writable and large enough? */
4622 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4623 /* yes, overwrite it. Note in this case we don't
4624 * bother to try shrinking the node if the new data
4625 * is smaller than the overflow threshold.
4627 if (F_ISSET(flags, MDB_RESERVE))
4628 data->mv_data = METADATA(omp);
4630 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4633 /* no, free ovpages */
4635 mc->mc_db->md_overflow_pages -= ovpages;
4636 for (i=0; i<ovpages; i++) {
4637 DPRINTF("freed ov page %zu", pg);
4638 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4642 } else if (NODEDSZ(leaf) == data->mv_size) {
4643 /* same size, just replace it. Note that we could
4644 * also reuse this node if the new data is smaller,
4645 * but instead we opt to shrink the node in that case.
4647 if (F_ISSET(flags, MDB_RESERVE))
4648 data->mv_data = NODEDATA(leaf);
4650 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4653 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4654 mc->mc_db->md_entries--;
4656 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4663 nflags = flags & NODE_ADD_FLAGS;
4664 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4665 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4666 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4667 nflags &= ~MDB_APPEND;
4669 nflags |= MDB_SPLIT_REPLACE;
4670 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4672 /* There is room already in this leaf page. */
4673 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4674 if (rc == 0 && !do_sub && insert) {
4675 /* Adjust other cursors pointing to mp */
4676 MDB_cursor *m2, *m3;
4677 MDB_dbi dbi = mc->mc_dbi;
4678 unsigned i = mc->mc_top;
4679 MDB_page *mp = mc->mc_pg[i];
4681 if (mc->mc_flags & C_SUB)
4684 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4685 if (mc->mc_flags & C_SUB)
4686 m3 = &m2->mc_xcursor->mx_cursor;
4689 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4690 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4697 if (rc != MDB_SUCCESS)
4698 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4700 /* Now store the actual data in the child DB. Note that we're
4701 * storing the user data in the keys field, so there are strict
4702 * size limits on dupdata. The actual data fields of the child
4703 * DB are all zero size.
4710 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4711 if (flags & MDB_CURRENT) {
4712 xflags = MDB_CURRENT;
4714 mdb_xcursor_init1(mc, leaf);
4715 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4717 /* converted, write the original data first */
4719 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4723 /* Adjust other cursors pointing to mp */
4725 unsigned i = mc->mc_top;
4726 MDB_page *mp = mc->mc_pg[i];
4728 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4729 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4730 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4731 mdb_xcursor_init1(m2, leaf);
4736 if (flags & MDB_APPENDDUP)
4737 xflags |= MDB_APPEND;
4738 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4739 if (flags & F_SUBDATA) {
4740 void *db = NODEDATA(leaf);
4741 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4744 /* sub-writes might have failed so check rc again.
4745 * Don't increment count if we just replaced an existing item.
4747 if (!rc && !(flags & MDB_CURRENT))
4748 mc->mc_db->md_entries++;
4749 if (flags & MDB_MULTIPLE) {
4751 if (mcount < data[1].mv_size) {
4752 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4753 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4763 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4768 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4771 if (!mc->mc_flags & C_INITIALIZED)
4774 rc = mdb_cursor_touch(mc);
4778 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4780 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4781 if (flags != MDB_NODUPDATA) {
4782 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4783 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4785 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4786 /* If sub-DB still has entries, we're done */
4787 if (mc->mc_xcursor->mx_db.md_entries) {
4788 if (leaf->mn_flags & F_SUBDATA) {
4789 /* update subDB info */
4790 void *db = NODEDATA(leaf);
4791 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4793 /* shrink fake page */
4794 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4796 mc->mc_db->md_entries--;
4799 /* otherwise fall thru and delete the sub-DB */
4802 if (leaf->mn_flags & F_SUBDATA) {
4803 /* add all the child DB's pages to the free list */
4804 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4805 if (rc == MDB_SUCCESS) {
4806 mc->mc_db->md_entries -=
4807 mc->mc_xcursor->mx_db.md_entries;
4812 return mdb_cursor_del0(mc, leaf);
4815 /** Allocate and initialize new pages for a database.
4816 * @param[in] mc a cursor on the database being added to.
4817 * @param[in] flags flags defining what type of page is being allocated.
4818 * @param[in] num the number of pages to allocate. This is usually 1,
4819 * unless allocating overflow pages for a large record.
4820 * @param[out] mp Address of a page, or NULL on failure.
4821 * @return 0 on success, non-zero on failure.
4824 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
4829 if ((rc = mdb_page_alloc(mc, num, &np)))
4831 DPRINTF("allocated new mpage %zu, page size %u",
4832 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4833 np->mp_flags = flags | P_DIRTY;
4834 np->mp_lower = PAGEHDRSZ;
4835 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4838 mc->mc_db->md_branch_pages++;
4839 else if (IS_LEAF(np))
4840 mc->mc_db->md_leaf_pages++;
4841 else if (IS_OVERFLOW(np)) {
4842 mc->mc_db->md_overflow_pages += num;
4850 /** Calculate the size of a leaf node.
4851 * The size depends on the environment's page size; if a data item
4852 * is too large it will be put onto an overflow page and the node
4853 * size will only include the key and not the data. Sizes are always
4854 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4855 * of the #MDB_node headers.
4856 * @param[in] env The environment handle.
4857 * @param[in] key The key for the node.
4858 * @param[in] data The data for the node.
4859 * @return The number of bytes needed to store the node.
4862 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4866 sz = LEAFSIZE(key, data);
4867 if (sz >= env->me_psize / MDB_MINKEYS) {
4868 /* put on overflow page */
4869 sz -= data->mv_size - sizeof(pgno_t);
4873 return sz + sizeof(indx_t);
4876 /** Calculate the size of a branch node.
4877 * The size should depend on the environment's page size but since
4878 * we currently don't support spilling large keys onto overflow
4879 * pages, it's simply the size of the #MDB_node header plus the
4880 * size of the key. Sizes are always rounded up to an even number
4881 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4882 * @param[in] env The environment handle.
4883 * @param[in] key The key for the node.
4884 * @return The number of bytes needed to store the node.
4887 mdb_branch_size(MDB_env *env, MDB_val *key)
4892 if (sz >= env->me_psize / MDB_MINKEYS) {
4893 /* put on overflow page */
4894 /* not implemented */
4895 /* sz -= key->size - sizeof(pgno_t); */
4898 return sz + sizeof(indx_t);
4901 /** Add a node to the page pointed to by the cursor.
4902 * @param[in] mc The cursor for this operation.
4903 * @param[in] indx The index on the page where the new node should be added.
4904 * @param[in] key The key for the new node.
4905 * @param[in] data The data for the new node, if any.
4906 * @param[in] pgno The page number, if adding a branch node.
4907 * @param[in] flags Flags for the node.
4908 * @return 0 on success, non-zero on failure. Possible errors are:
4910 * <li>ENOMEM - failed to allocate overflow pages for the node.
4911 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
4912 * should never happen since all callers already calculate the
4913 * page's free space before calling this function.
4917 mdb_node_add(MDB_cursor *mc, indx_t indx,
4918 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4921 size_t node_size = NODESIZE;
4924 MDB_page *mp = mc->mc_pg[mc->mc_top];
4925 MDB_page *ofp = NULL; /* overflow page */
4928 assert(mp->mp_upper >= mp->mp_lower);
4930 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4931 IS_LEAF(mp) ? "leaf" : "branch",
4932 IS_SUBP(mp) ? "sub-" : "",
4933 mp->mp_pgno, indx, data ? data->mv_size : 0,
4934 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4937 /* Move higher keys up one slot. */
4938 int ksize = mc->mc_db->md_pad, dif;
4939 char *ptr = LEAF2KEY(mp, indx, ksize);
4940 dif = NUMKEYS(mp) - indx;
4942 memmove(ptr+ksize, ptr, dif*ksize);
4943 /* insert new key */
4944 memcpy(ptr, key->mv_data, ksize);
4946 /* Just using these for counting */
4947 mp->mp_lower += sizeof(indx_t);
4948 mp->mp_upper -= ksize - sizeof(indx_t);
4953 node_size += key->mv_size;
4957 if (F_ISSET(flags, F_BIGDATA)) {
4958 /* Data already on overflow page. */
4959 node_size += sizeof(pgno_t);
4960 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4961 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4963 /* Put data on overflow page. */
4964 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4965 data->mv_size, node_size+data->mv_size);
4966 node_size += sizeof(pgno_t);
4967 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
4969 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4972 node_size += data->mv_size;
4975 node_size += node_size & 1;
4977 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4978 DPRINTF("not enough room in page %zu, got %u ptrs",
4979 mp->mp_pgno, NUMKEYS(mp));
4980 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4981 mp->mp_upper - mp->mp_lower);
4982 DPRINTF("node size = %zu", node_size);
4983 return MDB_PAGE_FULL;
4986 /* Move higher pointers up one slot. */
4987 for (i = NUMKEYS(mp); i > indx; i--)
4988 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
4990 /* Adjust free space offsets. */
4991 ofs = mp->mp_upper - node_size;
4992 assert(ofs >= mp->mp_lower + sizeof(indx_t));
4993 mp->mp_ptrs[indx] = ofs;
4995 mp->mp_lower += sizeof(indx_t);
4997 /* Write the node data. */
4998 node = NODEPTR(mp, indx);
4999 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5000 node->mn_flags = flags;
5002 SETDSZ(node,data->mv_size);
5007 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5012 if (F_ISSET(flags, F_BIGDATA))
5013 memcpy(node->mn_data + key->mv_size, data->mv_data,
5015 else if (F_ISSET(flags, MDB_RESERVE))
5016 data->mv_data = node->mn_data + key->mv_size;
5018 memcpy(node->mn_data + key->mv_size, data->mv_data,
5021 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5023 if (F_ISSET(flags, MDB_RESERVE))
5024 data->mv_data = METADATA(ofp);
5026 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5033 /** Delete the specified node from a page.
5034 * @param[in] mp The page to operate on.
5035 * @param[in] indx The index of the node to delete.
5036 * @param[in] ksize The size of a node. Only used if the page is
5037 * part of a #MDB_DUPFIXED database.
5040 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5043 indx_t i, j, numkeys, ptr;
5050 COPY_PGNO(pgno, mp->mp_pgno);
5051 DPRINTF("delete node %u on %s page %zu", indx,
5052 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5055 assert(indx < NUMKEYS(mp));
5058 int x = NUMKEYS(mp) - 1 - indx;
5059 base = LEAF2KEY(mp, indx, ksize);
5061 memmove(base, base + ksize, x * ksize);
5062 mp->mp_lower -= sizeof(indx_t);
5063 mp->mp_upper += ksize - sizeof(indx_t);
5067 node = NODEPTR(mp, indx);
5068 sz = NODESIZE + node->mn_ksize;
5070 if (F_ISSET(node->mn_flags, F_BIGDATA))
5071 sz += sizeof(pgno_t);
5073 sz += NODEDSZ(node);
5077 ptr = mp->mp_ptrs[indx];
5078 numkeys = NUMKEYS(mp);
5079 for (i = j = 0; i < numkeys; i++) {
5081 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5082 if (mp->mp_ptrs[i] < ptr)
5083 mp->mp_ptrs[j] += sz;
5088 base = (char *)mp + mp->mp_upper;
5089 memmove(base + sz, base, ptr - mp->mp_upper);
5091 mp->mp_lower -= sizeof(indx_t);
5095 /** Compact the main page after deleting a node on a subpage.
5096 * @param[in] mp The main page to operate on.
5097 * @param[in] indx The index of the subpage on the main page.
5100 mdb_node_shrink(MDB_page *mp, indx_t indx)
5107 indx_t i, numkeys, ptr;
5109 node = NODEPTR(mp, indx);
5110 sp = (MDB_page *)NODEDATA(node);
5111 osize = NODEDSZ(node);
5113 delta = sp->mp_upper - sp->mp_lower;
5114 SETDSZ(node, osize - delta);
5115 xp = (MDB_page *)((char *)sp + delta);
5117 /* shift subpage upward */
5119 nsize = NUMKEYS(sp) * sp->mp_pad;
5120 memmove(METADATA(xp), METADATA(sp), nsize);
5123 nsize = osize - sp->mp_upper;
5124 numkeys = NUMKEYS(sp);
5125 for (i=numkeys-1; i>=0; i--)
5126 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5128 xp->mp_upper = sp->mp_lower;
5129 xp->mp_lower = sp->mp_lower;
5130 xp->mp_flags = sp->mp_flags;
5131 xp->mp_pad = sp->mp_pad;
5132 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5134 /* shift lower nodes upward */
5135 ptr = mp->mp_ptrs[indx];
5136 numkeys = NUMKEYS(mp);
5137 for (i = 0; i < numkeys; i++) {
5138 if (mp->mp_ptrs[i] <= ptr)
5139 mp->mp_ptrs[i] += delta;
5142 base = (char *)mp + mp->mp_upper;
5143 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5144 mp->mp_upper += delta;
5147 /** Initial setup of a sorted-dups cursor.
5148 * Sorted duplicates are implemented as a sub-database for the given key.
5149 * The duplicate data items are actually keys of the sub-database.
5150 * Operations on the duplicate data items are performed using a sub-cursor
5151 * initialized when the sub-database is first accessed. This function does
5152 * the preliminary setup of the sub-cursor, filling in the fields that
5153 * depend only on the parent DB.
5154 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5157 mdb_xcursor_init0(MDB_cursor *mc)
5159 MDB_xcursor *mx = mc->mc_xcursor;
5161 mx->mx_cursor.mc_xcursor = NULL;
5162 mx->mx_cursor.mc_txn = mc->mc_txn;
5163 mx->mx_cursor.mc_db = &mx->mx_db;
5164 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5165 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5166 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5167 mx->mx_cursor.mc_snum = 0;
5168 mx->mx_cursor.mc_top = 0;
5169 mx->mx_cursor.mc_flags = C_SUB;
5170 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5171 mx->mx_dbx.md_dcmp = NULL;
5172 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5175 /** Final setup of a sorted-dups cursor.
5176 * Sets up the fields that depend on the data from the main cursor.
5177 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5178 * @param[in] node The data containing the #MDB_db record for the
5179 * sorted-dup database.
5182 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5184 MDB_xcursor *mx = mc->mc_xcursor;
5186 if (node->mn_flags & F_SUBDATA) {
5187 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5188 mx->mx_cursor.mc_pg[0] = 0;
5189 mx->mx_cursor.mc_snum = 0;
5190 mx->mx_cursor.mc_flags = C_SUB;
5192 MDB_page *fp = NODEDATA(node);
5193 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5194 mx->mx_db.md_flags = 0;
5195 mx->mx_db.md_depth = 1;
5196 mx->mx_db.md_branch_pages = 0;
5197 mx->mx_db.md_leaf_pages = 1;
5198 mx->mx_db.md_overflow_pages = 0;
5199 mx->mx_db.md_entries = NUMKEYS(fp);
5200 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5201 mx->mx_cursor.mc_snum = 1;
5202 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5203 mx->mx_cursor.mc_top = 0;
5204 mx->mx_cursor.mc_pg[0] = fp;
5205 mx->mx_cursor.mc_ki[0] = 0;
5206 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5207 mx->mx_db.md_flags = MDB_DUPFIXED;
5208 mx->mx_db.md_pad = fp->mp_pad;
5209 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5210 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5213 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5215 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5217 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5218 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5219 #if UINT_MAX < SIZE_MAX
5220 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5221 #ifdef MISALIGNED_OK
5222 mx->mx_dbx.md_cmp = mdb_cmp_long;
5224 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5229 /** Initialize a cursor for a given transaction and database. */
5231 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5236 mc->mc_db = &txn->mt_dbs[dbi];
5237 mc->mc_dbx = &txn->mt_dbxs[dbi];
5238 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5243 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5245 mc->mc_xcursor = mx;
5246 mdb_xcursor_init0(mc);
5248 mc->mc_xcursor = NULL;
5250 if (*mc->mc_dbflag & DB_STALE) {
5251 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5256 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5259 MDB_xcursor *mx = NULL;
5260 size_t size = sizeof(MDB_cursor);
5262 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5265 /* Allow read access to the freelist */
5266 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5269 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5270 size += sizeof(MDB_xcursor);
5272 if ((mc = malloc(size)) != NULL) {
5273 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5274 mx = (MDB_xcursor *)(mc + 1);
5276 mdb_cursor_init(mc, txn, dbi, mx);
5277 if (txn->mt_cursors) {
5278 mc->mc_next = txn->mt_cursors[dbi];
5279 txn->mt_cursors[dbi] = mc;
5281 mc->mc_flags |= C_ALLOCD;
5292 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5294 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5297 if (txn->mt_cursors)
5300 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5304 /* Return the count of duplicate data items for the current key */
5306 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5310 if (mc == NULL || countp == NULL)
5313 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5316 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5317 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5320 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5323 *countp = mc->mc_xcursor->mx_db.md_entries;
5329 mdb_cursor_close(MDB_cursor *mc)
5332 /* remove from txn, if tracked */
5333 if (mc->mc_txn->mt_cursors) {
5334 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5335 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5337 *prev = mc->mc_next;
5339 if (mc->mc_flags & C_ALLOCD)
5345 mdb_cursor_txn(MDB_cursor *mc)
5347 if (!mc) return NULL;
5352 mdb_cursor_dbi(MDB_cursor *mc)
5358 /** Replace the key for a node with a new key.
5359 * @param[in] mp The page containing the node to operate on.
5360 * @param[in] indx The index of the node to operate on.
5361 * @param[in] key The new key to use.
5362 * @return 0 on success, non-zero on failure.
5365 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5371 indx_t ptr, i, numkeys;
5374 node = NODEPTR(mp, indx);
5375 ptr = mp->mp_ptrs[indx];
5379 char kbuf2[(MAXKEYSIZE*2+1)];
5380 k2.mv_data = NODEKEY(node);
5381 k2.mv_size = node->mn_ksize;
5382 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5384 mdb_dkey(&k2, kbuf2),
5390 delta0 = delta = key->mv_size - node->mn_ksize;
5392 /* Must be 2-byte aligned. If new key is
5393 * shorter by 1, the shift will be skipped.
5395 delta += (delta & 1);
5397 if (delta > 0 && SIZELEFT(mp) < delta) {
5398 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5399 return MDB_PAGE_FULL;
5402 numkeys = NUMKEYS(mp);
5403 for (i = 0; i < numkeys; i++) {
5404 if (mp->mp_ptrs[i] <= ptr)
5405 mp->mp_ptrs[i] -= delta;
5408 base = (char *)mp + mp->mp_upper;
5409 len = ptr - mp->mp_upper + NODESIZE;
5410 memmove(base - delta, base, len);
5411 mp->mp_upper -= delta;
5413 node = NODEPTR(mp, indx);
5416 /* But even if no shift was needed, update ksize */
5418 node->mn_ksize = key->mv_size;
5421 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5426 /** Move a node from csrc to cdst.
5429 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5435 unsigned short flags;
5439 /* Mark src and dst as dirty. */
5440 if ((rc = mdb_page_touch(csrc)) ||
5441 (rc = mdb_page_touch(cdst)))
5444 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5445 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5446 key.mv_size = csrc->mc_db->md_pad;
5447 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5449 data.mv_data = NULL;
5453 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5454 assert(!((long)srcnode&1));
5455 srcpg = NODEPGNO(srcnode);
5456 flags = srcnode->mn_flags;
5457 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5458 unsigned int snum = csrc->mc_snum;
5460 /* must find the lowest key below src */
5461 mdb_page_search_root(csrc, NULL, 0);
5462 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5463 key.mv_size = csrc->mc_db->md_pad;
5464 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5466 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5467 key.mv_size = NODEKSZ(s2);
5468 key.mv_data = NODEKEY(s2);
5470 csrc->mc_snum = snum--;
5471 csrc->mc_top = snum;
5473 key.mv_size = NODEKSZ(srcnode);
5474 key.mv_data = NODEKEY(srcnode);
5476 data.mv_size = NODEDSZ(srcnode);
5477 data.mv_data = NODEDATA(srcnode);
5479 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5480 unsigned int snum = cdst->mc_snum;
5483 /* must find the lowest key below dst */
5484 mdb_page_search_root(cdst, NULL, 0);
5485 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5486 bkey.mv_size = cdst->mc_db->md_pad;
5487 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5489 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5490 bkey.mv_size = NODEKSZ(s2);
5491 bkey.mv_data = NODEKEY(s2);
5493 cdst->mc_snum = snum--;
5494 cdst->mc_top = snum;
5495 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5498 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5499 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5500 csrc->mc_ki[csrc->mc_top],
5502 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5503 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5505 /* Add the node to the destination page.
5507 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5508 if (rc != MDB_SUCCESS)
5511 /* Delete the node from the source page.
5513 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5516 /* Adjust other cursors pointing to mp */
5517 MDB_cursor *m2, *m3;
5518 MDB_dbi dbi = csrc->mc_dbi;
5519 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5521 if (csrc->mc_flags & C_SUB)
5524 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5525 if (m2 == csrc) continue;
5526 if (csrc->mc_flags & C_SUB)
5527 m3 = &m2->mc_xcursor->mx_cursor;
5530 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5531 csrc->mc_ki[csrc->mc_top]) {
5532 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5533 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5538 /* Update the parent separators.
5540 if (csrc->mc_ki[csrc->mc_top] == 0) {
5541 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5542 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5543 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5545 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5546 key.mv_size = NODEKSZ(srcnode);
5547 key.mv_data = NODEKEY(srcnode);
5549 DPRINTF("update separator for source page %zu to [%s]",
5550 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5551 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5552 &key)) != MDB_SUCCESS)
5555 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5557 nullkey.mv_size = 0;
5558 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5559 assert(rc == MDB_SUCCESS);
5563 if (cdst->mc_ki[cdst->mc_top] == 0) {
5564 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5565 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5566 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5568 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5569 key.mv_size = NODEKSZ(srcnode);
5570 key.mv_data = NODEKEY(srcnode);
5572 DPRINTF("update separator for destination page %zu to [%s]",
5573 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5574 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5575 &key)) != MDB_SUCCESS)
5578 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5580 nullkey.mv_size = 0;
5581 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5582 assert(rc == MDB_SUCCESS);
5589 /** Merge one page into another.
5590 * The nodes from the page pointed to by \b csrc will
5591 * be copied to the page pointed to by \b cdst and then
5592 * the \b csrc page will be freed.
5593 * @param[in] csrc Cursor pointing to the source page.
5594 * @param[in] cdst Cursor pointing to the destination page.
5597 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5605 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5606 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5608 assert(csrc->mc_snum > 1); /* can't merge root page */
5609 assert(cdst->mc_snum > 1);
5611 /* Mark dst as dirty. */
5612 if ((rc = mdb_page_touch(cdst)))
5615 /* Move all nodes from src to dst.
5617 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5618 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5619 key.mv_size = csrc->mc_db->md_pad;
5620 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5621 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5622 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5623 if (rc != MDB_SUCCESS)
5625 key.mv_data = (char *)key.mv_data + key.mv_size;
5628 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5629 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5630 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5631 unsigned int snum = csrc->mc_snum;
5633 /* must find the lowest key below src */
5634 mdb_page_search_root(csrc, NULL, 0);
5635 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5636 key.mv_size = csrc->mc_db->md_pad;
5637 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5639 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5640 key.mv_size = NODEKSZ(s2);
5641 key.mv_data = NODEKEY(s2);
5643 csrc->mc_snum = snum--;
5644 csrc->mc_top = snum;
5646 key.mv_size = srcnode->mn_ksize;
5647 key.mv_data = NODEKEY(srcnode);
5650 data.mv_size = NODEDSZ(srcnode);
5651 data.mv_data = NODEDATA(srcnode);
5652 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5653 if (rc != MDB_SUCCESS)
5658 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5659 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);
5661 /* Unlink the src page from parent and add to free list.
5663 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5664 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5666 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5670 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5671 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5672 csrc->mc_db->md_leaf_pages--;
5674 csrc->mc_db->md_branch_pages--;
5676 /* Adjust other cursors pointing to mp */
5677 MDB_cursor *m2, *m3;
5678 MDB_dbi dbi = csrc->mc_dbi;
5679 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5681 if (csrc->mc_flags & C_SUB)
5684 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5685 if (csrc->mc_flags & C_SUB)
5686 m3 = &m2->mc_xcursor->mx_cursor;
5689 if (m3 == csrc) continue;
5690 if (m3->mc_snum < csrc->mc_snum) continue;
5691 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5692 m3->mc_pg[csrc->mc_top] = mp;
5693 m3->mc_ki[csrc->mc_top] += nkeys;
5697 mdb_cursor_pop(csrc);
5699 return mdb_rebalance(csrc);
5702 /** Copy the contents of a cursor.
5703 * @param[in] csrc The cursor to copy from.
5704 * @param[out] cdst The cursor to copy to.
5707 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5711 cdst->mc_txn = csrc->mc_txn;
5712 cdst->mc_dbi = csrc->mc_dbi;
5713 cdst->mc_db = csrc->mc_db;
5714 cdst->mc_dbx = csrc->mc_dbx;
5715 cdst->mc_snum = csrc->mc_snum;
5716 cdst->mc_top = csrc->mc_top;
5717 cdst->mc_flags = csrc->mc_flags;
5719 for (i=0; i<csrc->mc_snum; i++) {
5720 cdst->mc_pg[i] = csrc->mc_pg[i];
5721 cdst->mc_ki[i] = csrc->mc_ki[i];
5725 /** Rebalance the tree after a delete operation.
5726 * @param[in] mc Cursor pointing to the page where rebalancing
5728 * @return 0 on success, non-zero on failure.
5731 mdb_rebalance(MDB_cursor *mc)
5741 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5742 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5743 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5744 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5748 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5751 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5752 DPRINTF("no need to rebalance page %zu, above fill threshold",
5758 if (mc->mc_snum < 2) {
5759 MDB_page *mp = mc->mc_pg[0];
5760 if (NUMKEYS(mp) == 0) {
5761 DPUTS("tree is completely empty");
5762 mc->mc_db->md_root = P_INVALID;
5763 mc->mc_db->md_depth = 0;
5764 mc->mc_db->md_leaf_pages = 0;
5765 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5769 /* Adjust other cursors pointing to mp */
5770 MDB_cursor *m2, *m3;
5771 MDB_dbi dbi = mc->mc_dbi;
5773 if (mc->mc_flags & C_SUB)
5776 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5777 if (m2 == mc) continue;
5778 if (mc->mc_flags & C_SUB)
5779 m3 = &m2->mc_xcursor->mx_cursor;
5782 if (m3->mc_snum < mc->mc_snum) continue;
5783 if (m3->mc_pg[0] == mp) {
5789 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5790 DPUTS("collapsing root page!");
5791 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5792 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5793 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5796 mc->mc_db->md_depth--;
5797 mc->mc_db->md_branch_pages--;
5799 /* Adjust other cursors pointing to mp */
5800 MDB_cursor *m2, *m3;
5801 MDB_dbi dbi = mc->mc_dbi;
5803 if (mc->mc_flags & C_SUB)
5806 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5807 if (m2 == mc) continue;
5808 if (mc->mc_flags & C_SUB)
5809 m3 = &m2->mc_xcursor->mx_cursor;
5812 if (m3->mc_snum < mc->mc_snum) continue;
5813 if (m3->mc_pg[0] == mp) {
5814 m3->mc_pg[0] = mc->mc_pg[0];
5819 DPUTS("root page doesn't need rebalancing");
5823 /* The parent (branch page) must have at least 2 pointers,
5824 * otherwise the tree is invalid.
5826 ptop = mc->mc_top-1;
5827 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5829 /* Leaf page fill factor is below the threshold.
5830 * Try to move keys from left or right neighbor, or
5831 * merge with a neighbor page.
5836 mdb_cursor_copy(mc, &mn);
5837 mn.mc_xcursor = NULL;
5839 if (mc->mc_ki[ptop] == 0) {
5840 /* We're the leftmost leaf in our parent.
5842 DPUTS("reading right neighbor");
5844 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5845 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5847 mn.mc_ki[mn.mc_top] = 0;
5848 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5850 /* There is at least one neighbor to the left.
5852 DPUTS("reading left neighbor");
5854 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5855 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5857 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5858 mc->mc_ki[mc->mc_top] = 0;
5861 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5862 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);
5864 /* If the neighbor page is above threshold and has at least two
5865 * keys, move one key from it.
5867 * Otherwise we should try to merge them.
5869 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5870 return mdb_node_move(&mn, mc);
5871 else { /* FIXME: if (has_enough_room()) */
5872 mc->mc_flags &= ~C_INITIALIZED;
5873 if (mc->mc_ki[ptop] == 0)
5874 return mdb_page_merge(&mn, mc);
5876 return mdb_page_merge(mc, &mn);
5880 /** Complete a delete operation started by #mdb_cursor_del(). */
5882 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5886 /* add overflow pages to free list */
5887 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5891 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5892 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5893 mc->mc_db->md_overflow_pages -= ovpages;
5894 for (i=0; i<ovpages; i++) {
5895 DPRINTF("freed ov page %zu", pg);
5896 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5900 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5901 mc->mc_db->md_entries--;
5902 rc = mdb_rebalance(mc);
5903 if (rc != MDB_SUCCESS)
5904 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5910 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5911 MDB_val *key, MDB_val *data)
5916 MDB_val rdata, *xdata;
5920 assert(key != NULL);
5922 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5924 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5927 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5931 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5935 mdb_cursor_init(&mc, txn, dbi, &mx);
5946 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5948 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5952 /** Split a page and insert a new node.
5953 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5954 * The cursor will be updated to point to the actual page and index where
5955 * the node got inserted after the split.
5956 * @param[in] newkey The key for the newly inserted node.
5957 * @param[in] newdata The data for the newly inserted node.
5958 * @param[in] newpgno The page number, if the new node is a branch node.
5959 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
5960 * @return 0 on success, non-zero on failure.
5963 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5964 unsigned int nflags)
5967 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
5970 unsigned int i, j, split_indx, nkeys, pmax;
5972 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5974 MDB_page *mp, *rp, *pp;
5979 mp = mc->mc_pg[mc->mc_top];
5980 newindx = mc->mc_ki[mc->mc_top];
5982 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
5983 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
5984 DKEY(newkey), mc->mc_ki[mc->mc_top]);
5986 /* Create a right sibling. */
5987 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
5989 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
5991 if (mc->mc_snum < 2) {
5992 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
5994 /* shift current top to make room for new parent */
5995 mc->mc_pg[1] = mc->mc_pg[0];
5996 mc->mc_ki[1] = mc->mc_ki[0];
5999 mc->mc_db->md_root = pp->mp_pgno;
6000 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6001 mc->mc_db->md_depth++;
6004 /* Add left (implicit) pointer. */
6005 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6006 /* undo the pre-push */
6007 mc->mc_pg[0] = mc->mc_pg[1];
6008 mc->mc_ki[0] = mc->mc_ki[1];
6009 mc->mc_db->md_root = mp->mp_pgno;
6010 mc->mc_db->md_depth--;
6017 ptop = mc->mc_top-1;
6018 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6021 mc->mc_flags |= C_SPLITTING;
6022 mdb_cursor_copy(mc, &mn);
6023 mn.mc_pg[mn.mc_top] = rp;
6024 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6026 if (nflags & MDB_APPEND) {
6027 mn.mc_ki[mn.mc_top] = 0;
6029 split_indx = newindx;
6034 nkeys = NUMKEYS(mp);
6035 split_indx = nkeys / 2;
6036 if (newindx < split_indx)
6042 unsigned int lsize, rsize, ksize;
6043 /* Move half of the keys to the right sibling */
6045 x = mc->mc_ki[mc->mc_top] - split_indx;
6046 ksize = mc->mc_db->md_pad;
6047 split = LEAF2KEY(mp, split_indx, ksize);
6048 rsize = (nkeys - split_indx) * ksize;
6049 lsize = (nkeys - split_indx) * sizeof(indx_t);
6050 mp->mp_lower -= lsize;
6051 rp->mp_lower += lsize;
6052 mp->mp_upper += rsize - lsize;
6053 rp->mp_upper -= rsize - lsize;
6054 sepkey.mv_size = ksize;
6055 if (newindx == split_indx) {
6056 sepkey.mv_data = newkey->mv_data;
6058 sepkey.mv_data = split;
6061 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6062 memcpy(rp->mp_ptrs, split, rsize);
6063 sepkey.mv_data = rp->mp_ptrs;
6064 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6065 memcpy(ins, newkey->mv_data, ksize);
6066 mp->mp_lower += sizeof(indx_t);
6067 mp->mp_upper -= ksize - sizeof(indx_t);
6070 memcpy(rp->mp_ptrs, split, x * ksize);
6071 ins = LEAF2KEY(rp, x, ksize);
6072 memcpy(ins, newkey->mv_data, ksize);
6073 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6074 rp->mp_lower += sizeof(indx_t);
6075 rp->mp_upper -= ksize - sizeof(indx_t);
6076 mc->mc_ki[mc->mc_top] = x;
6077 mc->mc_pg[mc->mc_top] = rp;
6082 /* For leaf pages, check the split point based on what
6083 * fits where, since otherwise mdb_node_add can fail.
6085 * This check is only needed when the data items are
6086 * relatively large, such that being off by one will
6087 * make the difference between success or failure.
6088 * When the size of the data items is much smaller than
6089 * one-half of a page, this check is irrelevant.
6092 unsigned int psize, nsize;
6093 /* Maximum free space in an empty page */
6094 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6095 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6096 if ((nkeys < 20) || (nsize > pmax/4)) {
6097 if (newindx <= split_indx) {
6100 for (i=0; i<split_indx; i++) {
6101 node = NODEPTR(mp, i);
6102 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6103 if (F_ISSET(node->mn_flags, F_BIGDATA))
6104 psize += sizeof(pgno_t);
6106 psize += NODEDSZ(node);
6110 split_indx = newindx;
6121 for (i=nkeys-1; i>=split_indx; i--) {
6122 node = NODEPTR(mp, i);
6123 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6124 if (F_ISSET(node->mn_flags, F_BIGDATA))
6125 psize += sizeof(pgno_t);
6127 psize += NODEDSZ(node);
6131 split_indx = newindx;
6141 /* First find the separating key between the split pages.
6142 * The case where newindx == split_indx is ambiguous; the
6143 * new item could go to the new page or stay on the original
6144 * page. If newpos == 1 it goes to the new page.
6146 if (newindx == split_indx && newpos) {
6147 sepkey.mv_size = newkey->mv_size;
6148 sepkey.mv_data = newkey->mv_data;
6150 node = NODEPTR(mp, split_indx);
6151 sepkey.mv_size = node->mn_ksize;
6152 sepkey.mv_data = NODEKEY(node);
6156 DPRINTF("separator is [%s]", DKEY(&sepkey));
6158 /* Copy separator key to the parent.
6160 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6164 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6167 if (mn.mc_snum == mc->mc_snum) {
6168 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6169 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6170 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6171 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6176 /* Right page might now have changed parent.
6177 * Check if left page also changed parent.
6179 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6180 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6181 for (i=0; i<ptop; i++) {
6182 mc->mc_pg[i] = mn.mc_pg[i];
6183 mc->mc_ki[i] = mn.mc_ki[i];
6185 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6186 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6190 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6193 mc->mc_flags ^= C_SPLITTING;
6194 if (rc != MDB_SUCCESS) {
6197 if (nflags & MDB_APPEND) {
6198 mc->mc_pg[mc->mc_top] = rp;
6199 mc->mc_ki[mc->mc_top] = 0;
6200 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6203 for (i=0; i<mc->mc_top; i++)
6204 mc->mc_ki[i] = mn.mc_ki[i];
6211 /* Move half of the keys to the right sibling. */
6213 /* grab a page to hold a temporary copy */
6214 copy = mdb_page_malloc(mc);
6218 copy->mp_pgno = mp->mp_pgno;
6219 copy->mp_flags = mp->mp_flags;
6220 copy->mp_lower = PAGEHDRSZ;
6221 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6222 mc->mc_pg[mc->mc_top] = copy;
6223 for (i = j = 0; i <= nkeys; j++) {
6224 if (i == split_indx) {
6225 /* Insert in right sibling. */
6226 /* Reset insert index for right sibling. */
6227 if (i != newindx || (newpos ^ ins_new)) {
6229 mc->mc_pg[mc->mc_top] = rp;
6233 if (i == newindx && !ins_new) {
6234 /* Insert the original entry that caused the split. */
6235 rkey.mv_data = newkey->mv_data;
6236 rkey.mv_size = newkey->mv_size;
6245 /* Update index for the new key. */
6246 mc->mc_ki[mc->mc_top] = j;
6247 } else if (i == nkeys) {
6250 node = NODEPTR(mp, i);
6251 rkey.mv_data = NODEKEY(node);
6252 rkey.mv_size = node->mn_ksize;
6254 xdata.mv_data = NODEDATA(node);
6255 xdata.mv_size = NODEDSZ(node);
6258 pgno = NODEPGNO(node);
6259 flags = node->mn_flags;
6264 if (!IS_LEAF(mp) && j == 0) {
6265 /* First branch index doesn't need key data. */
6269 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6273 nkeys = NUMKEYS(copy);
6274 for (i=0; i<nkeys; i++)
6275 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6276 mp->mp_lower = copy->mp_lower;
6277 mp->mp_upper = copy->mp_upper;
6278 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6279 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6281 /* reset back to original page */
6282 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6283 mc->mc_pg[mc->mc_top] = mp;
6284 if (nflags & MDB_RESERVE) {
6285 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6286 if (!(node->mn_flags & F_BIGDATA))
6287 newdata->mv_data = NODEDATA(node);
6293 /* return tmp page to freelist */
6294 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
6295 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
6296 mc->mc_txn->mt_env->me_dpages = copy;
6299 /* Adjust other cursors pointing to mp */
6300 MDB_cursor *m2, *m3;
6301 MDB_dbi dbi = mc->mc_dbi;
6302 int fixup = NUMKEYS(mp);
6304 if (mc->mc_flags & C_SUB)
6307 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6308 if (m2 == mc) continue;
6309 if (mc->mc_flags & C_SUB)
6310 m3 = &m2->mc_xcursor->mx_cursor;
6313 if (!(m3->mc_flags & C_INITIALIZED))
6315 if (m3->mc_flags & C_SPLITTING)
6320 for (k=m3->mc_top; k>=0; k--) {
6321 m3->mc_ki[k+1] = m3->mc_ki[k];
6322 m3->mc_pg[k+1] = m3->mc_pg[k];
6324 if (m3->mc_ki[0] >= split_indx) {
6329 m3->mc_pg[0] = mc->mc_pg[0];
6333 if (m3->mc_pg[mc->mc_top] == mp) {
6334 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6335 m3->mc_ki[mc->mc_top]++;
6336 if (m3->mc_ki[mc->mc_top] >= fixup) {
6337 m3->mc_pg[mc->mc_top] = rp;
6338 m3->mc_ki[mc->mc_top] -= fixup;
6339 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6341 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6342 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6351 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6352 MDB_val *key, MDB_val *data, unsigned int flags)
6357 assert(key != NULL);
6358 assert(data != NULL);
6360 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6363 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6367 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6371 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6374 mdb_cursor_init(&mc, txn, dbi, &mx);
6375 return mdb_cursor_put(&mc, key, data, flags);
6378 /** Only a subset of the @ref mdb_env flags can be changed
6379 * at runtime. Changing other flags requires closing the environment
6380 * and re-opening it with the new flags.
6382 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
6384 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6386 if ((flag & CHANGEABLE) != flag)
6389 env->me_flags |= flag;
6391 env->me_flags &= ~flag;
6396 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6401 *arg = env->me_flags;
6406 mdb_env_get_path(MDB_env *env, const char **arg)
6411 *arg = env->me_path;
6415 /** Common code for #mdb_stat() and #mdb_env_stat().
6416 * @param[in] env the environment to operate in.
6417 * @param[in] db the #MDB_db record containing the stats to return.
6418 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6419 * @return 0, this function always succeeds.
6422 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6424 arg->ms_psize = env->me_psize;
6425 arg->ms_depth = db->md_depth;
6426 arg->ms_branch_pages = db->md_branch_pages;
6427 arg->ms_leaf_pages = db->md_leaf_pages;
6428 arg->ms_overflow_pages = db->md_overflow_pages;
6429 arg->ms_entries = db->md_entries;
6434 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6438 if (env == NULL || arg == NULL)
6441 toggle = mdb_env_pick_meta(env);
6443 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6446 /** Set the default comparison functions for a database.
6447 * Called immediately after a database is opened to set the defaults.
6448 * The user can then override them with #mdb_set_compare() or
6449 * #mdb_set_dupsort().
6450 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6451 * @param[in] dbi A database handle returned by #mdb_open()
6454 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6456 uint16_t f = txn->mt_dbs[dbi].md_flags;
6458 txn->mt_dbxs[dbi].md_cmp =
6459 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6460 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6462 txn->mt_dbxs[dbi].md_dcmp =
6463 !(f & MDB_DUPSORT) ? 0 :
6464 ((f & MDB_INTEGERDUP)
6465 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6466 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6469 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6474 int rc, dbflag, exact;
6475 unsigned int unused = 0;
6478 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6479 mdb_default_cmp(txn, FREE_DBI);
6485 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
6486 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
6487 mdb_default_cmp(txn, MAIN_DBI);
6491 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6492 mdb_default_cmp(txn, MAIN_DBI);
6495 /* Is the DB already open? */
6497 for (i=2; i<txn->mt_numdbs; i++) {
6498 if (!txn->mt_dbxs[i].md_name.mv_size) {
6499 /* Remember this free slot */
6500 if (!unused) unused = i;
6503 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6504 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6510 /* If no free slot and max hit, fail */
6511 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
6512 return MDB_DBS_FULL;
6514 /* Find the DB info */
6518 key.mv_data = (void *)name;
6519 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6520 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6521 if (rc == MDB_SUCCESS) {
6522 /* make sure this is actually a DB */
6523 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6524 if (!(node->mn_flags & F_SUBDATA))
6526 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6527 /* Create if requested */
6529 data.mv_size = sizeof(MDB_db);
6530 data.mv_data = &dummy;
6531 memset(&dummy, 0, sizeof(dummy));
6532 dummy.md_root = P_INVALID;
6533 dummy.md_flags = flags & 0xffff;
6534 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6538 /* OK, got info, add to table */
6539 if (rc == MDB_SUCCESS) {
6540 unsigned int slot = unused ? unused : txn->mt_numdbs;
6541 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
6542 txn->mt_dbxs[slot].md_name.mv_size = len;
6543 txn->mt_dbxs[slot].md_rel = NULL;
6544 txn->mt_dbflags[slot] = dbflag;
6545 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
6547 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
6548 mdb_default_cmp(txn, slot);
6551 txn->mt_env->me_numdbs++;
6558 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6560 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6563 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6566 void mdb_close(MDB_env *env, MDB_dbi dbi)
6569 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6571 ptr = env->me_dbxs[dbi].md_name.mv_data;
6572 env->me_dbxs[dbi].md_name.mv_data = NULL;
6573 env->me_dbxs[dbi].md_name.mv_size = 0;
6577 /** Add all the DB's pages to the free list.
6578 * @param[in] mc Cursor on the DB to free.
6579 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6580 * @return 0 on success, non-zero on failure.
6583 mdb_drop0(MDB_cursor *mc, int subs)
6587 rc = mdb_page_search(mc, NULL, 0);
6588 if (rc == MDB_SUCCESS) {
6593 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6594 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6597 mdb_cursor_copy(mc, &mx);
6598 while (mc->mc_snum > 0) {
6599 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6600 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6601 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6602 if (ni->mn_flags & F_SUBDATA) {
6603 mdb_xcursor_init1(mc, ni);
6604 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6610 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6612 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6615 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6620 rc = mdb_cursor_sibling(mc, 1);
6622 /* no more siblings, go back to beginning
6623 * of previous level. (stack was already popped
6624 * by mdb_cursor_sibling)
6626 for (i=1; i<mc->mc_top; i++)
6627 mc->mc_pg[i] = mx.mc_pg[i];
6631 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6632 mc->mc_db->md_root);
6637 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6642 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6645 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6648 rc = mdb_cursor_open(txn, dbi, &mc);
6652 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6656 /* Can't delete the main DB */
6657 if (del && dbi > MAIN_DBI) {
6658 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6660 mdb_close(txn->mt_env, dbi);
6662 txn->mt_dbflags[dbi] |= DB_DIRTY;
6663 txn->mt_dbs[dbi].md_depth = 0;
6664 txn->mt_dbs[dbi].md_branch_pages = 0;
6665 txn->mt_dbs[dbi].md_leaf_pages = 0;
6666 txn->mt_dbs[dbi].md_overflow_pages = 0;
6667 txn->mt_dbs[dbi].md_entries = 0;
6668 txn->mt_dbs[dbi].md_root = P_INVALID;
6671 mdb_cursor_close(mc);
6675 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6677 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6680 txn->mt_dbxs[dbi].md_cmp = cmp;
6684 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6686 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6689 txn->mt_dbxs[dbi].md_dcmp = cmp;
6693 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6695 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6698 txn->mt_dbxs[dbi].md_rel = rel;
6702 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6704 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6707 txn->mt_dbxs[dbi].md_relctx = ctx;