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 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
36 #include <sys/types.h>
38 #include <sys/param.h>
44 #ifdef HAVE_SYS_FILE_H
61 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
62 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
65 #if defined(__APPLE__) || defined (BSD)
66 # define MDB_USE_POSIX_SEM 1
67 # define MDB_FDATASYNC fsync
68 #elif defined(ANDROID)
69 # define MDB_FDATASYNC fsync
74 #ifdef MDB_USE_POSIX_SEM
75 #include <semaphore.h>
80 #include <valgrind/memcheck.h>
81 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
82 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
83 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
84 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
85 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
87 #define VGMEMP_CREATE(h,r,z)
88 #define VGMEMP_ALLOC(h,a,s)
89 #define VGMEMP_FREE(h,a)
90 #define VGMEMP_DESTROY(h)
91 #define VGMEMP_DEFINED(a,s)
95 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
96 /* Solaris just defines one or the other */
97 # define LITTLE_ENDIAN 1234
98 # define BIG_ENDIAN 4321
99 # ifdef _LITTLE_ENDIAN
100 # define BYTE_ORDER LITTLE_ENDIAN
102 # define BYTE_ORDER BIG_ENDIAN
105 # define BYTE_ORDER __BYTE_ORDER
109 #ifndef LITTLE_ENDIAN
110 #define LITTLE_ENDIAN __LITTLE_ENDIAN
113 #define BIG_ENDIAN __BIG_ENDIAN
116 #if defined(__i386) || defined(__x86_64)
117 #define MISALIGNED_OK 1
123 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
124 # error "Unknown or unsupported endianness (BYTE_ORDER)"
125 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
126 # error "Two's complement, reasonably sized integer types, please"
129 /** @defgroup internal MDB Internals
132 /** @defgroup compat Windows Compatibility Macros
133 * A bunch of macros to minimize the amount of platform-specific ifdefs
134 * needed throughout the rest of the code. When the features this library
135 * needs are similar enough to POSIX to be hidden in a one-or-two line
136 * replacement, this macro approach is used.
140 #define pthread_t DWORD
141 #define pthread_mutex_t HANDLE
142 #define pthread_key_t DWORD
143 #define pthread_self() GetCurrentThreadId()
144 #define pthread_key_create(x,y) \
145 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
146 #define pthread_key_delete(x) TlsFree(x)
147 #define pthread_getspecific(x) TlsGetValue(x)
148 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
149 #define pthread_mutex_unlock(x) ReleaseMutex(x)
150 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
151 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
152 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
153 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
154 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
155 #define getpid() GetCurrentProcessId()
156 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
157 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
158 #define ErrCode() GetLastError()
159 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
160 #define close(fd) CloseHandle(fd)
161 #define munmap(ptr,len) UnmapViewOfFile(ptr)
164 #ifdef MDB_USE_POSIX_SEM
166 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
167 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
168 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
169 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
172 mdb_sem_wait(sem_t *sem)
175 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
180 /** Lock the reader mutex.
182 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
183 /** Unlock the reader mutex.
185 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
187 /** Lock the writer mutex.
188 * Only a single write transaction is allowed at a time. Other writers
189 * will block waiting for this mutex.
191 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
192 /** Unlock the writer mutex.
194 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
195 #endif /* MDB_USE_POSIX_SEM */
197 /** Get the error code for the last failed system function.
199 #define ErrCode() errno
201 /** An abstraction for a file handle.
202 * On POSIX systems file handles are small integers. On Windows
203 * they're opaque pointers.
207 /** A value for an invalid file handle.
208 * Mainly used to initialize file variables and signify that they are
211 #define INVALID_HANDLE_VALUE (-1)
213 /** Get the size of a memory page for the system.
214 * This is the basic size that the platform's memory manager uses, and is
215 * fundamental to the use of memory-mapped files.
217 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
220 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
223 #define MNAME_LEN (sizeof(pthread_mutex_t))
229 /** A flag for opening a file and requesting synchronous data writes.
230 * This is only used when writing a meta page. It's not strictly needed;
231 * we could just do a normal write and then immediately perform a flush.
232 * But if this flag is available it saves us an extra system call.
234 * @note If O_DSYNC is undefined but exists in /usr/include,
235 * preferably set some compiler flag to get the definition.
236 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
239 # define MDB_DSYNC O_DSYNC
243 /** Function for flushing the data of a file. Define this to fsync
244 * if fdatasync() is not supported.
246 #ifndef MDB_FDATASYNC
247 # define MDB_FDATASYNC fdatasync
251 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
262 /** A page number in the database.
263 * Note that 64 bit page numbers are overkill, since pages themselves
264 * already represent 12-13 bits of addressable memory, and the OS will
265 * always limit applications to a maximum of 63 bits of address space.
267 * @note In the #MDB_node structure, we only store 48 bits of this value,
268 * which thus limits us to only 60 bits of addressable data.
270 typedef MDB_ID pgno_t;
272 /** A transaction ID.
273 * See struct MDB_txn.mt_txnid for details.
275 typedef MDB_ID txnid_t;
277 /** @defgroup debug Debug Macros
281 /** Enable debug output.
282 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
283 * read from and written to the database (used for free space management).
288 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
289 # define DPRINTF (void) /* Vararg macros may be unsupported */
291 static int mdb_debug;
292 static txnid_t mdb_debug_start;
294 /** Print a debug message with printf formatting. */
295 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
296 ((void) ((mdb_debug) && \
297 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
299 # define DPRINTF(fmt, ...) ((void) 0)
300 # define MDB_DEBUG_SKIP
302 /** Print a debug string.
303 * The string is printed literally, with no format processing.
305 #define DPUTS(arg) DPRINTF("%s", arg)
308 /** A default memory page size.
309 * The actual size is platform-dependent, but we use this for
310 * boot-strapping. We probably should not be using this any more.
311 * The #GET_PAGESIZE() macro is used to get the actual size.
313 * Note that we don't currently support Huge pages. On Linux,
314 * regular data files cannot use Huge pages, and in general
315 * Huge pages aren't actually pageable. We rely on the OS
316 * demand-pager to read our data and page it out when memory
317 * pressure from other processes is high. So until OSs have
318 * actual paging support for Huge pages, they're not viable.
320 #define MDB_PAGESIZE 4096
322 /** The minimum number of keys required in a database page.
323 * Setting this to a larger value will place a smaller bound on the
324 * maximum size of a data item. Data items larger than this size will
325 * be pushed into overflow pages instead of being stored directly in
326 * the B-tree node. This value used to default to 4. With a page size
327 * of 4096 bytes that meant that any item larger than 1024 bytes would
328 * go into an overflow page. That also meant that on average 2-3KB of
329 * each overflow page was wasted space. The value cannot be lower than
330 * 2 because then there would no longer be a tree structure. With this
331 * value, items larger than 2KB will go into overflow pages, and on
332 * average only 1KB will be wasted.
334 #define MDB_MINKEYS 2
336 /** A stamp that identifies a file as an MDB file.
337 * There's nothing special about this value other than that it is easily
338 * recognizable, and it will reflect any byte order mismatches.
340 #define MDB_MAGIC 0xBEEFC0DE
342 /** The version number for a database's file format. */
343 #define MDB_VERSION 1
345 /** @brief The maximum size of a key in the database.
347 * We require that keys all fit onto a regular page. This limit
348 * could be raised a bit further if needed; to something just
349 * under #MDB_PAGESIZE / #MDB_MINKEYS.
351 * Note that data items in an #MDB_DUPSORT database are actually keys
352 * of a subDB, so they're also limited to this size.
354 #ifndef MDB_MAXKEYSIZE
355 #define MDB_MAXKEYSIZE 511
358 /** @brief The maximum size of a data item.
360 * We only store a 32 bit value for node sizes.
362 #define MAXDATASIZE 0xffffffffUL
367 * This is used for printing a hex dump of a key's contents.
369 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
370 /** Display a key in hex.
372 * Invoke a function to display a key in hex.
374 #define DKEY(x) mdb_dkey(x, kbuf)
376 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
380 /** An invalid page number.
381 * Mainly used to denote an empty tree.
383 #define P_INVALID (~(pgno_t)0)
385 /** Test if the flags \b f are set in a flag word \b w. */
386 #define F_ISSET(w, f) (((w) & (f)) == (f))
388 /** Used for offsets within a single page.
389 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
392 typedef uint16_t indx_t;
394 /** Default size of memory map.
395 * This is certainly too small for any actual applications. Apps should always set
396 * the size explicitly using #mdb_env_set_mapsize().
398 #define DEFAULT_MAPSIZE 1048576
400 /** @defgroup readers Reader Lock Table
401 * Readers don't acquire any locks for their data access. Instead, they
402 * simply record their transaction ID in the reader table. The reader
403 * mutex is needed just to find an empty slot in the reader table. The
404 * slot's address is saved in thread-specific data so that subsequent read
405 * transactions started by the same thread need no further locking to proceed.
407 * No reader table is used if the database is on a read-only filesystem.
409 * Since the database uses multi-version concurrency control, readers don't
410 * actually need any locking. This table is used to keep track of which
411 * readers are using data from which old transactions, so that we'll know
412 * when a particular old transaction is no longer in use. Old transactions
413 * that have discarded any data pages can then have those pages reclaimed
414 * for use by a later write transaction.
416 * The lock table is constructed such that reader slots are aligned with the
417 * processor's cache line size. Any slot is only ever used by one thread.
418 * This alignment guarantees that there will be no contention or cache
419 * thrashing as threads update their own slot info, and also eliminates
420 * any need for locking when accessing a slot.
422 * A writer thread will scan every slot in the table to determine the oldest
423 * outstanding reader transaction. Any freed pages older than this will be
424 * reclaimed by the writer. The writer doesn't use any locks when scanning
425 * this table. This means that there's no guarantee that the writer will
426 * see the most up-to-date reader info, but that's not required for correct
427 * operation - all we need is to know the upper bound on the oldest reader,
428 * we don't care at all about the newest reader. So the only consequence of
429 * reading stale information here is that old pages might hang around a
430 * while longer before being reclaimed. That's actually good anyway, because
431 * the longer we delay reclaiming old pages, the more likely it is that a
432 * string of contiguous pages can be found after coalescing old pages from
433 * many old transactions together.
435 * @todo We don't actually do such coalescing yet, we grab pages from one
436 * old transaction at a time.
439 /** Number of slots in the reader table.
440 * This value was chosen somewhat arbitrarily. 126 readers plus a
441 * couple mutexes fit exactly into 8KB on my development machine.
442 * Applications should set the table size using #mdb_env_set_maxreaders().
444 #define DEFAULT_READERS 126
446 /** The size of a CPU cache line in bytes. We want our lock structures
447 * aligned to this size to avoid false cache line sharing in the
449 * This value works for most CPUs. For Itanium this should be 128.
455 /** The information we store in a single slot of the reader table.
456 * In addition to a transaction ID, we also record the process and
457 * thread ID that owns a slot, so that we can detect stale information,
458 * e.g. threads or processes that went away without cleaning up.
459 * @note We currently don't check for stale records. We simply re-init
460 * the table when we know that we're the only process opening the
463 typedef struct MDB_rxbody {
464 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
465 * Multiple readers that start at the same time will probably have the
466 * same ID here. Again, it's not important to exclude them from
467 * anything; all we need to know is which version of the DB they
468 * started from so we can avoid overwriting any data used in that
469 * particular version.
472 /** The process ID of the process owning this reader txn. */
474 /** The thread ID of the thread owning this txn. */
478 /** The actual reader record, with cacheline padding. */
479 typedef struct MDB_reader {
482 /** shorthand for mrb_txnid */
483 #define mr_txnid mru.mrx.mrb_txnid
484 #define mr_pid mru.mrx.mrb_pid
485 #define mr_tid mru.mrx.mrb_tid
486 /** cache line alignment */
487 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
491 /** The header for the reader table.
492 * The table resides in a memory-mapped file. (This is a different file
493 * than is used for the main database.)
495 * For POSIX the actual mutexes reside in the shared memory of this
496 * mapped file. On Windows, mutexes are named objects allocated by the
497 * kernel; we store the mutex names in this mapped file so that other
498 * processes can grab them. This same approach is also used on
499 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
500 * process-shared POSIX mutexes. For these cases where a named object
501 * is used, the object name is derived from a 64 bit FNV hash of the
502 * environment pathname. As such, naming collisions are extremely
503 * unlikely. If a collision occurs, the results are unpredictable.
505 typedef struct MDB_txbody {
506 /** Stamp identifying this as an MDB file. It must be set
509 /** Version number of this lock file. Must be set to #MDB_VERSION. */
510 uint32_t mtb_version;
511 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
512 char mtb_rmname[MNAME_LEN];
514 /** Mutex protecting access to this table.
515 * This is the reader lock that #LOCK_MUTEX_R acquires.
517 pthread_mutex_t mtb_mutex;
519 /** The ID of the last transaction committed to the database.
520 * This is recorded here only for convenience; the value can always
521 * be determined by reading the main database meta pages.
524 /** The number of slots that have been used in the reader table.
525 * This always records the maximum count, it is not decremented
526 * when readers release their slots.
528 unsigned mtb_numreaders;
531 /** The actual reader table definition. */
532 typedef struct MDB_txninfo {
535 #define mti_magic mt1.mtb.mtb_magic
536 #define mti_version mt1.mtb.mtb_version
537 #define mti_mutex mt1.mtb.mtb_mutex
538 #define mti_rmname mt1.mtb.mtb_rmname
539 #define mti_txnid mt1.mtb.mtb_txnid
540 #define mti_numreaders mt1.mtb.mtb_numreaders
541 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
544 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
545 char mt2_wmname[MNAME_LEN];
546 #define mti_wmname mt2.mt2_wmname
548 pthread_mutex_t mt2_wmutex;
549 #define mti_wmutex mt2.mt2_wmutex
551 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
553 MDB_reader mti_readers[1];
557 /** Common header for all page types.
558 * Overflow records occupy a number of contiguous pages with no
559 * headers on any page after the first.
561 typedef struct MDB_page {
562 #define mp_pgno mp_p.p_pgno
563 #define mp_next mp_p.p_next
565 pgno_t p_pgno; /**< page number */
566 void * p_next; /**< for in-memory list of freed structs */
569 /** @defgroup mdb_page Page Flags
571 * Flags for the page headers.
574 #define P_BRANCH 0x01 /**< branch page */
575 #define P_LEAF 0x02 /**< leaf page */
576 #define P_OVERFLOW 0x04 /**< overflow page */
577 #define P_META 0x08 /**< meta page */
578 #define P_DIRTY 0x10 /**< dirty page */
579 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
580 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
582 uint16_t mp_flags; /**< @ref mdb_page */
583 #define mp_lower mp_pb.pb.pb_lower
584 #define mp_upper mp_pb.pb.pb_upper
585 #define mp_pages mp_pb.pb_pages
588 indx_t pb_lower; /**< lower bound of free space */
589 indx_t pb_upper; /**< upper bound of free space */
591 uint32_t pb_pages; /**< number of overflow pages */
593 indx_t mp_ptrs[1]; /**< dynamic size */
596 /** Size of the page header, excluding dynamic data at the end */
597 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
599 /** Address of first usable data byte in a page, after the header */
600 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
602 /** Number of nodes on a page */
603 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
605 /** The amount of space remaining in the page */
606 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
608 /** The percentage of space used in the page, in tenths of a percent. */
609 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
610 ((env)->me_psize - PAGEHDRSZ))
611 /** The minimum page fill factor, in tenths of a percent.
612 * Pages emptier than this are candidates for merging.
614 #define FILL_THRESHOLD 250
616 /** Test if a page is a leaf page */
617 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
618 /** Test if a page is a LEAF2 page */
619 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
620 /** Test if a page is a branch page */
621 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
622 /** Test if a page is an overflow page */
623 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
624 /** Test if a page is a sub page */
625 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
627 /** The number of overflow pages needed to store the given size. */
628 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
630 /** Header for a single key/data pair within a page.
631 * We guarantee 2-byte alignment for nodes.
633 typedef struct MDB_node {
634 /** lo and hi are used for data size on leaf nodes and for
635 * child pgno on branch nodes. On 64 bit platforms, flags
636 * is also used for pgno. (Branch nodes have no flags).
637 * They are in host byte order in case that lets some
638 * accesses be optimized into a 32-bit word access.
640 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
641 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
642 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
643 /** @defgroup mdb_node Node Flags
645 * Flags for node headers.
648 #define F_BIGDATA 0x01 /**< data put on overflow page */
649 #define F_SUBDATA 0x02 /**< data is a sub-database */
650 #define F_DUPDATA 0x04 /**< data has duplicates */
652 /** valid flags for #mdb_node_add() */
653 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
656 unsigned short mn_flags; /**< @ref mdb_node */
657 unsigned short mn_ksize; /**< key size */
658 char mn_data[1]; /**< key and data are appended here */
661 /** Size of the node header, excluding dynamic data at the end */
662 #define NODESIZE offsetof(MDB_node, mn_data)
664 /** Bit position of top word in page number, for shifting mn_flags */
665 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
667 /** Size of a node in a branch page with a given key.
668 * This is just the node header plus the key, there is no data.
670 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
672 /** Size of a node in a leaf page with a given key and data.
673 * This is node header plus key plus data size.
675 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
677 /** Address of node \b i in page \b p */
678 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
680 /** Address of the key for the node */
681 #define NODEKEY(node) (void *)((node)->mn_data)
683 /** Address of the data for a node */
684 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
686 /** Get the page number pointed to by a branch node */
687 #define NODEPGNO(node) \
688 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
689 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
690 /** Set the page number in a branch node */
691 #define SETPGNO(node,pgno) do { \
692 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
693 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
695 /** Get the size of the data in a leaf node */
696 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
697 /** Set the size of the data for a leaf node */
698 #define SETDSZ(node,size) do { \
699 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
700 /** The size of a key in a node */
701 #define NODEKSZ(node) ((node)->mn_ksize)
703 /** Copy a page number from src to dst */
705 #define COPY_PGNO(dst,src) dst = src
707 #if SIZE_MAX > 4294967295UL
708 #define COPY_PGNO(dst,src) do { \
709 unsigned short *s, *d; \
710 s = (unsigned short *)&(src); \
711 d = (unsigned short *)&(dst); \
718 #define COPY_PGNO(dst,src) do { \
719 unsigned short *s, *d; \
720 s = (unsigned short *)&(src); \
721 d = (unsigned short *)&(dst); \
727 /** The address of a key in a LEAF2 page.
728 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
729 * There are no node headers, keys are stored contiguously.
731 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
733 /** Set the \b node's key into \b key, if requested. */
734 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
735 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
737 /** Information about a single database in the environment. */
738 typedef struct MDB_db {
739 uint32_t md_pad; /**< also ksize for LEAF2 pages */
740 uint16_t md_flags; /**< @ref mdb_dbi_open */
741 uint16_t md_depth; /**< depth of this tree */
742 pgno_t md_branch_pages; /**< number of internal pages */
743 pgno_t md_leaf_pages; /**< number of leaf pages */
744 pgno_t md_overflow_pages; /**< number of overflow pages */
745 size_t md_entries; /**< number of data items */
746 pgno_t md_root; /**< the root page of this tree */
749 /** Handle for the DB used to track free pages. */
751 /** Handle for the default DB. */
754 /** Meta page content. */
755 typedef struct MDB_meta {
756 /** Stamp identifying this as an MDB file. It must be set
759 /** Version number of this lock file. Must be set to #MDB_VERSION. */
761 void *mm_address; /**< address for fixed mapping */
762 size_t mm_mapsize; /**< size of mmap region */
763 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
764 /** The size of pages used in this DB */
765 #define mm_psize mm_dbs[0].md_pad
766 /** Any persistent environment flags. @ref mdb_env */
767 #define mm_flags mm_dbs[0].md_flags
768 pgno_t mm_last_pg; /**< last used page in file */
769 txnid_t mm_txnid; /**< txnid that committed this page */
772 /** Buffer for a stack-allocated dirty page.
773 * The members define size and alignment, and silence type
774 * aliasing warnings. They are not used directly; that could
775 * mean incorrectly using several union members in parallel.
777 typedef union MDB_pagebuf {
778 char mb_raw[MDB_PAGESIZE];
781 char mm_pad[PAGEHDRSZ];
786 /** Auxiliary DB info.
787 * The information here is mostly static/read-only. There is
788 * only a single copy of this record in the environment.
790 typedef struct MDB_dbx {
791 MDB_val md_name; /**< name of the database */
792 MDB_cmp_func *md_cmp; /**< function for comparing keys */
793 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
794 MDB_rel_func *md_rel; /**< user relocate function */
795 void *md_relctx; /**< user-provided context for md_rel */
798 /** A database transaction.
799 * Every operation requires a transaction handle.
802 MDB_txn *mt_parent; /**< parent of a nested txn */
803 MDB_txn *mt_child; /**< nested txn under this txn */
804 pgno_t mt_next_pgno; /**< next unallocated page */
805 /** The ID of this transaction. IDs are integers incrementing from 1.
806 * Only committed write transactions increment the ID. If a transaction
807 * aborts, the ID may be re-used by the next writer.
810 MDB_env *mt_env; /**< the DB environment */
811 /** The list of pages that became unused during this transaction.
815 MDB_ID2L dirty_list; /**< for write txns: modified pages */
816 MDB_reader *reader; /**< this thread's reader table slot or NULL */
818 /** Array of records for each DB known in the environment. */
820 /** Array of MDB_db records for each known DB */
822 /** @defgroup mt_dbflag Transaction DB Flags
826 #define DB_DIRTY 0x01 /**< DB was written in this txn */
827 #define DB_STALE 0x02 /**< DB record is older than txnID */
829 /** In write txns, array of cursors for each DB */
830 MDB_cursor **mt_cursors;
831 /** Array of flags for each DB */
832 unsigned char *mt_dbflags;
833 /** Number of DB records in use. This number only ever increments;
834 * we don't decrement it when individual DB handles are closed.
838 /** @defgroup mdb_txn Transaction Flags
842 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
843 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
844 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
846 unsigned int mt_flags; /**< @ref mdb_txn */
847 /** Tracks which of the two meta pages was used at the start
848 * of this transaction.
850 unsigned int mt_toggle;
853 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
854 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
855 * raise this on a 64 bit machine.
857 #define CURSOR_STACK 32
861 /** Cursors are used for all DB operations */
863 /** Next cursor on this DB in this txn */
865 /** Original cursor if this is a shadow */
867 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
868 struct MDB_xcursor *mc_xcursor;
869 /** The transaction that owns this cursor */
871 /** The database handle this cursor operates on */
873 /** The database record for this cursor */
875 /** The database auxiliary record for this cursor */
877 /** The @ref mt_dbflag for this database */
878 unsigned char *mc_dbflag;
879 unsigned short mc_snum; /**< number of pushed pages */
880 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
881 /** @defgroup mdb_cursor Cursor Flags
883 * Cursor state flags.
886 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
887 #define C_EOF 0x02 /**< No more data */
888 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
889 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
890 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
891 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
893 unsigned int mc_flags; /**< @ref mdb_cursor */
894 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
895 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
898 /** Context for sorted-dup records.
899 * We could have gone to a fully recursive design, with arbitrarily
900 * deep nesting of sub-databases. But for now we only handle these
901 * levels - main DB, optional sub-DB, sorted-duplicate DB.
903 typedef struct MDB_xcursor {
904 /** A sub-cursor for traversing the Dup DB */
905 MDB_cursor mx_cursor;
906 /** The database record for this Dup DB */
908 /** The auxiliary DB record for this Dup DB */
910 /** The @ref mt_dbflag for this Dup DB */
911 unsigned char mx_dbflag;
914 /** State of FreeDB old pages, stored in the MDB_env */
915 typedef struct MDB_pgstate {
916 txnid_t mf_pglast; /**< ID of last old page record we used */
917 pgno_t *mf_pghead; /**< old pages reclaimed from freelist */
918 pgno_t *mf_pgfree; /**< memory to free when dropping me_pghead */
921 /** The database environment. */
923 HANDLE me_fd; /**< The main data file */
924 HANDLE me_lfd; /**< The lock file */
925 HANDLE me_mfd; /**< just for writing the meta pages */
926 /** Failed to update the meta page. Probably an I/O error. */
927 #define MDB_FATAL_ERROR 0x80000000U
928 /** Read-only Filesystem. Allow read access, no locking. */
929 #define MDB_ROFS 0x40000000U
930 /** Some fields are initialized. */
931 #define MDB_ENV_ACTIVE 0x20000000U
932 uint32_t me_flags; /**< @ref mdb_env */
933 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
934 unsigned int me_maxreaders; /**< size of the reader table */
935 unsigned int me_numreaders; /**< max numreaders set by this env */
936 MDB_dbi me_numdbs; /**< number of DBs opened */
937 MDB_dbi me_maxdbs; /**< size of the DB table */
938 pid_t me_pid; /**< process ID of this env */
939 char *me_path; /**< path to the DB files */
940 char *me_map; /**< the memory map of the data file */
941 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
942 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
943 MDB_txn *me_txn; /**< current write transaction */
944 size_t me_mapsize; /**< size of the data memory map */
945 off_t me_size; /**< current file size */
946 pgno_t me_maxpg; /**< me_mapsize / me_psize */
947 MDB_dbx *me_dbxs; /**< array of static DB info */
948 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
949 pthread_key_t me_txkey; /**< thread-key for readers */
950 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
951 # define me_pglast me_pgstate.mf_pglast
952 # define me_pghead me_pgstate.mf_pghead
953 # define me_pgfree me_pgstate.mf_pgfree
954 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
955 /** IDL of pages that became unused in a write txn */
957 /** ID2L of pages that were written during a write txn */
958 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
959 /** Max number of freelist items that can fit in a single overflow page */
960 unsigned int me_maxfree_1pg;
962 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
964 #elif defined(MDB_USE_POSIX_SEM)
965 sem_t *me_rmutex; /* Shared mutexes are not supported */
970 /** Nested transaction */
971 typedef struct MDB_ntxn {
972 MDB_txn mnt_txn; /* the transaction */
973 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
976 /** max number of pages to commit in one writev() call */
977 #define MDB_COMMIT_PAGES 64
978 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
979 #undef MDB_COMMIT_PAGES
980 #define MDB_COMMIT_PAGES IOV_MAX
983 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
984 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
985 static int mdb_page_touch(MDB_cursor *mc);
987 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
988 static int mdb_page_search_root(MDB_cursor *mc,
989 MDB_val *key, int modify);
990 #define MDB_PS_MODIFY 1
991 #define MDB_PS_ROOTONLY 2
992 static int mdb_page_search(MDB_cursor *mc,
993 MDB_val *key, int flags);
994 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
996 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
997 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
998 pgno_t newpgno, unsigned int nflags);
1000 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1001 static int mdb_env_pick_meta(const MDB_env *env);
1002 static int mdb_env_write_meta(MDB_txn *txn);
1003 static void mdb_env_close0(MDB_env *env, int excl);
1005 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1006 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1007 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1008 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1009 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1010 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1011 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1012 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1013 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1015 static int mdb_rebalance(MDB_cursor *mc);
1016 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1018 static void mdb_cursor_pop(MDB_cursor *mc);
1019 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1021 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1022 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1023 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1024 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1025 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1027 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1028 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1030 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1031 static void mdb_xcursor_init0(MDB_cursor *mc);
1032 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1034 static int mdb_drop0(MDB_cursor *mc, int subs);
1035 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1038 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1042 static SECURITY_DESCRIPTOR mdb_null_sd;
1043 static SECURITY_ATTRIBUTES mdb_all_sa;
1044 static int mdb_sec_inited;
1047 /** Return the library version info. */
1049 mdb_version(int *major, int *minor, int *patch)
1051 if (major) *major = MDB_VERSION_MAJOR;
1052 if (minor) *minor = MDB_VERSION_MINOR;
1053 if (patch) *patch = MDB_VERSION_PATCH;
1054 return MDB_VERSION_STRING;
1057 /** Table of descriptions for MDB @ref errors */
1058 static char *const mdb_errstr[] = {
1059 "MDB_KEYEXIST: Key/data pair already exists",
1060 "MDB_NOTFOUND: No matching key/data pair found",
1061 "MDB_PAGE_NOTFOUND: Requested page not found",
1062 "MDB_CORRUPTED: Located page was wrong type",
1063 "MDB_PANIC: Update of meta page failed",
1064 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1065 "MDB_INVALID: File is not an MDB file",
1066 "MDB_MAP_FULL: Environment mapsize limit reached",
1067 "MDB_DBS_FULL: Environment maxdbs limit reached",
1068 "MDB_READERS_FULL: Environment maxreaders limit reached",
1069 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1070 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1071 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1072 "MDB_PAGE_FULL: Internal error - page has no more space",
1073 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1074 "MDB_INCOMPATIBLE: Operation is incompatible with database",
1078 mdb_strerror(int err)
1082 return ("Successful return: 0");
1084 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1085 i = err - MDB_KEYEXIST;
1086 return mdb_errstr[i];
1089 return strerror(err);
1093 /** Display a key in hexadecimal and return the address of the result.
1094 * @param[in] key the key to display
1095 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1096 * @return The key in hexadecimal form.
1099 mdb_dkey(MDB_val *key, char *buf)
1102 unsigned char *c = key->mv_data;
1108 if (key->mv_size > MDB_MAXKEYSIZE)
1109 return "MDB_MAXKEYSIZE";
1110 /* may want to make this a dynamic check: if the key is mostly
1111 * printable characters, print it as-is instead of converting to hex.
1115 for (i=0; i<key->mv_size; i++)
1116 ptr += sprintf(ptr, "%02x", *c++);
1118 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1123 /** Display all the keys in the page. */
1125 mdb_page_list(MDB_page *mp)
1128 unsigned int i, nkeys, nsize;
1132 nkeys = NUMKEYS(mp);
1133 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1134 for (i=0; i<nkeys; i++) {
1135 node = NODEPTR(mp, i);
1136 key.mv_size = node->mn_ksize;
1137 key.mv_data = node->mn_data;
1138 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1139 if (IS_BRANCH(mp)) {
1140 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1143 if (F_ISSET(node->mn_flags, F_BIGDATA))
1144 nsize += sizeof(pgno_t);
1146 nsize += NODEDSZ(node);
1147 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1153 mdb_cursor_chk(MDB_cursor *mc)
1159 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1160 for (i=0; i<mc->mc_top; i++) {
1162 node = NODEPTR(mp, mc->mc_ki[i]);
1163 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1166 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1172 /** Count all the pages in each DB and in the freelist
1173 * and make sure it matches the actual number of pages
1176 static void mdb_audit(MDB_txn *txn)
1180 MDB_ID freecount, count;
1185 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1186 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1187 freecount += *(MDB_ID *)data.mv_data;
1190 for (i = 0; i<txn->mt_numdbs; i++) {
1191 MDB_xcursor mx, *mxp;
1192 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1193 mdb_cursor_init(&mc, txn, i, mxp);
1194 if (txn->mt_dbs[i].md_root == P_INVALID)
1196 count += txn->mt_dbs[i].md_branch_pages +
1197 txn->mt_dbs[i].md_leaf_pages +
1198 txn->mt_dbs[i].md_overflow_pages;
1199 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1200 mdb_page_search(&mc, NULL, 0);
1204 mp = mc.mc_pg[mc.mc_top];
1205 for (j=0; j<NUMKEYS(mp); j++) {
1206 MDB_node *leaf = NODEPTR(mp, j);
1207 if (leaf->mn_flags & F_SUBDATA) {
1209 memcpy(&db, NODEDATA(leaf), sizeof(db));
1210 count += db.md_branch_pages + db.md_leaf_pages +
1211 db.md_overflow_pages;
1215 while (mdb_cursor_sibling(&mc, 1) == 0);
1218 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1219 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1220 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1226 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1228 return txn->mt_dbxs[dbi].md_cmp(a, b);
1232 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1234 if (txn->mt_dbxs[dbi].md_dcmp)
1235 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1237 return EINVAL; /* too bad you can't distinguish this from a valid result */
1240 /** Allocate a single page.
1241 * Re-use old malloc'd pages first, otherwise just malloc.
1244 mdb_page_malloc(MDB_cursor *mc) {
1246 size_t sz = mc->mc_txn->mt_env->me_psize;
1247 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1248 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1249 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1250 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1251 } else if ((ret = malloc(sz)) != NULL) {
1252 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1258 mdb_page_free(MDB_env *env, MDB_page *mp)
1260 mp->mp_next = env->me_dpages;
1261 VGMEMP_FREE(env, mp);
1262 env->me_dpages = mp;
1265 /** Allocate pages for writing.
1266 * If there are free pages available from older transactions, they
1267 * will be re-used first. Otherwise a new page will be allocated.
1268 * @param[in] mc cursor A cursor handle identifying the transaction and
1269 * database for which we are allocating.
1270 * @param[in] num the number of pages to allocate.
1271 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1272 * will always be satisfied by a single contiguous chunk of memory.
1273 * @return 0 on success, non-zero on failure.
1276 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1278 MDB_txn *txn = mc->mc_txn;
1280 pgno_t pgno = P_INVALID;
1282 txnid_t oldest = 0, last;
1287 /* If our dirty list is already full, we can't do anything */
1288 if (txn->mt_u.dirty_list[0].mid >= MDB_IDL_UM_MAX)
1289 return MDB_TXN_FULL;
1291 /* The free list won't have any content at all until txn 2 has
1292 * committed. The pages freed by txn 2 will be unreferenced
1293 * after txn 3 commits, and so will be safe to re-use in txn 4.
1295 if (txn->mt_txnid > 3) {
1296 if (!txn->mt_env->me_pghead &&
1297 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1298 /* See if there's anything in the free DB */
1305 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1306 if (!txn->mt_env->me_pglast) {
1307 mdb_page_search(&m2, NULL, 0);
1308 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1309 kptr = (txnid_t *)NODEKEY(leaf);
1314 last = txn->mt_env->me_pglast + 1;
1316 key.mv_data = &last;
1317 key.mv_size = sizeof(last);
1318 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET_RANGE, NULL);
1321 last = *(txnid_t *)key.mv_data;
1327 oldest = txn->mt_txnid - 1;
1328 nr = txn->mt_env->me_txns->mti_numreaders;
1329 r = txn->mt_env->me_txns->mti_readers;
1330 for (i=0; i<nr; i++) {
1331 if (!r[i].mr_pid) continue;
1338 if (oldest > last) {
1339 /* It's usable, grab it.
1343 if (!txn->mt_env->me_pglast) {
1344 mdb_node_read(txn, leaf, &data);
1346 idl = (MDB_ID *) data.mv_data;
1347 /* We might have a zero-length IDL due to freelist growth
1348 * during a prior commit
1351 txn->mt_env->me_pglast = last;
1354 mop = malloc(MDB_IDL_SIZEOF(idl));
1357 txn->mt_env->me_pglast = last;
1358 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop;
1359 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1364 DPRINTF("IDL read txn %zu root %zu num %zu",
1365 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1366 for (i=0; i<idl[0]; i++) {
1367 DPRINTF("IDL %zu", idl[i+1]);
1374 if (txn->mt_env->me_pghead) {
1375 pgno_t *mop = txn->mt_env->me_pghead;
1378 int retry = 1, readit = 0, n2 = num-1;
1379 unsigned int i, j, k;
1381 /* If current list is too short, must fetch more and coalesce */
1382 if (mop[0] < (unsigned)num)
1385 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1387 /* If on freelist, don't try to read more. If what we have
1388 * right now isn't enough just use new pages.
1389 * TODO: get all of this working. Many circular dependencies...
1391 if (mc->mc_dbi == FREE_DBI) {
1399 last = txn->mt_env->me_pglast + 1;
1401 /* We haven't hit the readers list yet? */
1407 oldest = txn->mt_txnid - 1;
1408 nr = txn->mt_env->me_txns->mti_numreaders;
1409 r = txn->mt_env->me_txns->mti_readers;
1410 for (i=0; i<nr; i++) {
1411 if (!r[i].mr_pid) continue;
1418 /* There's nothing we can use on the freelist */
1419 if (oldest - last < 1)
1422 key.mv_data = &last;
1423 key.mv_size = sizeof(last);
1424 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1426 if (rc == MDB_NOTFOUND)
1430 last = *(txnid_t*)key.mv_data;
1433 idl = (MDB_ID *) data.mv_data;
1434 mop2 = malloc(MDB_IDL_SIZEOF(idl) + MDB_IDL_SIZEOF(mop));
1437 /* merge in sorted order */
1438 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1440 while (i>0 || j>0) {
1441 if (i && idl[i] < mop[j])
1442 mop2[k--] = idl[i--];
1444 mop2[k--] = mop[j--];
1446 txn->mt_env->me_pglast = last;
1447 free(txn->mt_env->me_pgfree);
1448 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop2;
1450 /* Keep trying to read until we have enough */
1451 if (mop[0] < (unsigned)num) {
1456 /* current list has enough pages, but are they contiguous? */
1457 for (i=mop[0]; i>=(unsigned)num; i--) {
1458 if (mop[i-n2] == mop[i] + n2) {
1461 /* move any stragglers down */
1462 for (j=i+num; j<=mop[0]; j++)
1469 /* Stop if we succeeded, or no retries */
1470 if (!retry || pgno != P_INVALID)
1476 /* peel pages off tail, so we only have to truncate the list */
1477 pgno = MDB_IDL_LAST(mop);
1480 if (MDB_IDL_IS_ZERO(mop)) {
1481 free(txn->mt_env->me_pgfree);
1482 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1487 if (pgno == P_INVALID) {
1488 /* DB size is maxed out */
1489 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1490 DPUTS("DB size maxed out");
1491 return MDB_MAP_FULL;
1494 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1495 if (pgno == P_INVALID) {
1496 pgno = txn->mt_next_pgno;
1497 txn->mt_next_pgno += num;
1499 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1502 if (txn->mt_env->me_dpages && num == 1) {
1503 np = txn->mt_env->me_dpages;
1504 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1505 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1506 txn->mt_env->me_dpages = np->mp_next;
1508 size_t sz = txn->mt_env->me_psize * num;
1509 if ((np = malloc(sz)) == NULL)
1511 VGMEMP_ALLOC(txn->mt_env, np, sz);
1513 if (pgno == P_INVALID) {
1514 np->mp_pgno = txn->mt_next_pgno;
1515 txn->mt_next_pgno += num;
1520 mid.mid = np->mp_pgno;
1522 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1523 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1525 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1532 /** Copy a page: avoid copying unused portions of the page.
1533 * @param[in] dst page to copy into
1534 * @param[in] src page to copy from
1537 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1539 dst->mp_flags = src->mp_flags | P_DIRTY;
1540 dst->mp_pages = src->mp_pages;
1542 if (IS_LEAF2(src)) {
1543 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1545 unsigned int i, nkeys = NUMKEYS(src);
1546 for (i=0; i<nkeys; i++)
1547 dst->mp_ptrs[i] = src->mp_ptrs[i];
1548 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1549 psize - src->mp_upper);
1553 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1554 * @param[in] mc cursor pointing to the page to be touched
1555 * @return 0 on success, non-zero on failure.
1558 mdb_page_touch(MDB_cursor *mc)
1560 MDB_page *mp = mc->mc_pg[mc->mc_top];
1564 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1566 if ((rc = mdb_page_alloc(mc, 1, &np)))
1568 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1569 assert(mp->mp_pgno != np->mp_pgno);
1570 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1572 /* If page isn't full, just copy the used portion */
1573 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1576 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1578 np->mp_flags |= P_DIRTY;
1583 /* Adjust other cursors pointing to mp */
1584 if (mc->mc_flags & C_SUB) {
1585 MDB_cursor *m2, *m3;
1586 MDB_dbi dbi = mc->mc_dbi-1;
1588 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1589 if (m2 == mc) continue;
1590 m3 = &m2->mc_xcursor->mx_cursor;
1591 if (m3->mc_snum < mc->mc_snum) continue;
1592 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1593 m3->mc_pg[mc->mc_top] = mp;
1599 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1600 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1601 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1602 m2->mc_pg[mc->mc_top] = mp;
1606 mc->mc_pg[mc->mc_top] = mp;
1607 /** If this page has a parent, update the parent to point to
1611 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1613 mc->mc_db->md_root = mp->mp_pgno;
1614 } else if (mc->mc_txn->mt_parent) {
1617 /* If txn has a parent, make sure the page is in our
1620 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1621 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1622 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1623 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1624 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1625 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1626 mc->mc_pg[mc->mc_top] = mp;
1631 if (mc->mc_txn->mt_u.dirty_list[0].mid >= MDB_IDL_UM_MAX)
1632 return MDB_TXN_FULL;
1634 np = mdb_page_malloc(mc);
1637 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1638 mid.mid = np->mp_pgno;
1640 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1648 mdb_env_sync(MDB_env *env, int force)
1651 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1652 if (env->me_flags & MDB_WRITEMAP) {
1653 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1654 ? MS_ASYNC : MS_SYNC;
1655 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1658 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1662 if (MDB_FDATASYNC(env->me_fd))
1669 /** Make shadow copies of all of parent txn's cursors */
1671 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1673 MDB_cursor *mc, *m2;
1674 unsigned int i, j, size;
1676 for (i=0;i<src->mt_numdbs; i++) {
1677 if (src->mt_cursors[i]) {
1678 size = sizeof(MDB_cursor);
1679 if (src->mt_cursors[i]->mc_xcursor)
1680 size += sizeof(MDB_xcursor);
1681 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1688 mc->mc_db = &dst->mt_dbs[i];
1689 mc->mc_dbx = m2->mc_dbx;
1690 mc->mc_dbflag = &dst->mt_dbflags[i];
1691 mc->mc_snum = m2->mc_snum;
1692 mc->mc_top = m2->mc_top;
1693 mc->mc_flags = m2->mc_flags | C_SHADOW;
1694 for (j=0; j<mc->mc_snum; j++) {
1695 mc->mc_pg[j] = m2->mc_pg[j];
1696 mc->mc_ki[j] = m2->mc_ki[j];
1698 if (m2->mc_xcursor) {
1699 MDB_xcursor *mx, *mx2;
1700 mx = (MDB_xcursor *)(mc+1);
1701 mc->mc_xcursor = mx;
1702 mx2 = m2->mc_xcursor;
1703 mx->mx_db = mx2->mx_db;
1704 mx->mx_dbx = mx2->mx_dbx;
1705 mx->mx_dbflag = mx2->mx_dbflag;
1706 mx->mx_cursor.mc_txn = dst;
1707 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1708 mx->mx_cursor.mc_db = &mx->mx_db;
1709 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1710 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1711 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1712 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1713 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1714 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1715 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1716 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1719 mc->mc_xcursor = NULL;
1721 mc->mc_next = dst->mt_cursors[i];
1722 dst->mt_cursors[i] = mc;
1729 /** Merge shadow cursors back into parent's */
1731 mdb_cursor_merge(MDB_txn *txn)
1734 for (i=0; i<txn->mt_numdbs; i++) {
1735 if (txn->mt_cursors[i]) {
1737 while ((mc = txn->mt_cursors[i])) {
1738 txn->mt_cursors[i] = mc->mc_next;
1739 if (mc->mc_flags & C_SHADOW) {
1740 MDB_cursor *m2 = mc->mc_orig;
1742 m2->mc_snum = mc->mc_snum;
1743 m2->mc_top = mc->mc_top;
1744 for (j=0; j<mc->mc_snum; j++) {
1745 m2->mc_pg[j] = mc->mc_pg[j];
1746 m2->mc_ki[j] = mc->mc_ki[j];
1749 if (mc->mc_flags & C_ALLOCD)
1757 mdb_txn_reset0(MDB_txn *txn);
1759 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1760 * @param[in] txn the transaction handle to initialize
1761 * @return 0 on success, non-zero on failure. This can only
1762 * fail for read-only transactions, and then only if the
1763 * reader table is full.
1766 mdb_txn_renew0(MDB_txn *txn)
1768 MDB_env *env = txn->mt_env;
1773 txn->mt_numdbs = env->me_numdbs;
1774 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1776 if (txn->mt_flags & MDB_TXN_RDONLY) {
1777 if (env->me_flags & MDB_ROFS) {
1778 i = mdb_env_pick_meta(env);
1779 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1780 txn->mt_u.reader = NULL;
1782 MDB_reader *r = pthread_getspecific(env->me_txkey);
1784 pid_t pid = env->me_pid;
1785 pthread_t tid = pthread_self();
1788 for (i=0; i<env->me_txns->mti_numreaders; i++)
1789 if (env->me_txns->mti_readers[i].mr_pid == 0)
1791 if (i == env->me_maxreaders) {
1792 UNLOCK_MUTEX_R(env);
1793 return MDB_READERS_FULL;
1795 env->me_txns->mti_readers[i].mr_pid = pid;
1796 env->me_txns->mti_readers[i].mr_tid = tid;
1797 if (i >= env->me_txns->mti_numreaders)
1798 env->me_txns->mti_numreaders = i+1;
1799 /* Save numreaders for un-mutexed mdb_env_close() */
1800 env->me_numreaders = env->me_txns->mti_numreaders;
1801 UNLOCK_MUTEX_R(env);
1802 r = &env->me_txns->mti_readers[i];
1803 if ((rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1804 env->me_txns->mti_readers[i].mr_pid = 0;
1808 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1809 txn->mt_u.reader = r;
1811 txn->mt_toggle = txn->mt_txnid & 1;
1812 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1816 txn->mt_txnid = env->me_txns->mti_txnid;
1817 txn->mt_toggle = txn->mt_txnid & 1;
1818 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1821 if (txn->mt_txnid == mdb_debug_start)
1824 txn->mt_u.dirty_list = env->me_dirty_list;
1825 txn->mt_u.dirty_list[0].mid = 0;
1826 txn->mt_free_pgs = env->me_free_pgs;
1827 txn->mt_free_pgs[0] = 0;
1831 /* Copy the DB info and flags */
1832 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1833 for (i=2; i<txn->mt_numdbs; i++)
1834 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1835 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1836 if (txn->mt_numdbs > 2)
1837 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1839 if (env->me_maxpg < txn->mt_next_pgno) {
1840 mdb_txn_reset0(txn);
1841 return MDB_MAP_RESIZED;
1848 mdb_txn_renew(MDB_txn *txn)
1852 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1855 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1856 DPUTS("environment had fatal error, must shutdown!");
1860 rc = mdb_txn_renew0(txn);
1861 if (rc == MDB_SUCCESS) {
1862 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1863 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1864 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1870 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1874 int rc, size, tsize = sizeof(MDB_txn);
1876 if (env->me_flags & MDB_FATAL_ERROR) {
1877 DPUTS("environment had fatal error, must shutdown!");
1880 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1883 /* Nested transactions: Max 1 child, write txns only, no writemap */
1884 if (parent->mt_child ||
1885 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1886 (env->me_flags & MDB_WRITEMAP))
1890 tsize = sizeof(MDB_ntxn);
1892 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1893 if (!(flags & MDB_RDONLY))
1894 size += env->me_maxdbs * sizeof(MDB_cursor *);
1896 if ((txn = calloc(1, size)) == NULL) {
1897 DPRINTF("calloc: %s", strerror(ErrCode()));
1900 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1901 if (flags & MDB_RDONLY) {
1902 txn->mt_flags |= MDB_TXN_RDONLY;
1903 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1905 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1906 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1911 txn->mt_free_pgs = mdb_midl_alloc();
1912 if (!txn->mt_free_pgs) {
1916 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1917 if (!txn->mt_u.dirty_list) {
1918 free(txn->mt_free_pgs);
1922 txn->mt_txnid = parent->mt_txnid;
1923 txn->mt_toggle = parent->mt_toggle;
1924 txn->mt_u.dirty_list[0].mid = 0;
1925 txn->mt_free_pgs[0] = 0;
1926 txn->mt_next_pgno = parent->mt_next_pgno;
1927 parent->mt_child = txn;
1928 txn->mt_parent = parent;
1929 txn->mt_numdbs = parent->mt_numdbs;
1930 txn->mt_dbxs = parent->mt_dbxs;
1931 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1932 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1934 ntxn = (MDB_ntxn *)txn;
1935 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1936 if (env->me_pghead) {
1937 size = MDB_IDL_SIZEOF(env->me_pghead);
1938 env->me_pghead = malloc(size);
1940 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1944 env->me_pgfree = env->me_pghead;
1946 rc = mdb_cursor_shadow(parent, txn);
1948 mdb_txn_reset0(txn);
1950 rc = mdb_txn_renew0(txn);
1956 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1957 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1958 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1964 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1965 * @param[in] txn the transaction handle to reset
1968 mdb_txn_reset0(MDB_txn *txn)
1970 MDB_env *env = txn->mt_env;
1972 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1973 if (!(env->me_flags & MDB_ROFS))
1974 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1979 /* close(free) all cursors */
1980 for (i=0; i<txn->mt_numdbs; i++) {
1981 if (txn->mt_cursors[i]) {
1983 while ((mc = txn->mt_cursors[i])) {
1984 txn->mt_cursors[i] = mc->mc_next;
1985 if (mc->mc_flags & C_ALLOCD)
1991 if (!(env->me_flags & MDB_WRITEMAP)) {
1992 /* return all dirty pages to dpage list */
1993 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1994 dp = txn->mt_u.dirty_list[i].mptr;
1995 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1996 mdb_page_free(txn->mt_env, dp);
1998 /* large pages just get freed directly */
1999 VGMEMP_FREE(txn->mt_env, dp);
2005 free(env->me_pgfree);
2007 if (txn->mt_parent) {
2008 txn->mt_parent->mt_child = NULL;
2009 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2010 mdb_midl_free(txn->mt_free_pgs);
2011 free(txn->mt_u.dirty_list);
2014 if (mdb_midl_shrink(&txn->mt_free_pgs))
2015 env->me_free_pgs = txn->mt_free_pgs;
2018 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
2019 txn->mt_env->me_pglast = 0;
2022 /* The writer mutex was locked in mdb_txn_begin. */
2023 UNLOCK_MUTEX_W(env);
2028 mdb_txn_reset(MDB_txn *txn)
2033 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2034 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2035 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2037 mdb_txn_reset0(txn);
2041 mdb_txn_abort(MDB_txn *txn)
2046 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2047 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2048 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2051 mdb_txn_abort(txn->mt_child);
2053 mdb_txn_reset0(txn);
2058 mdb_txn_commit(MDB_txn *txn)
2066 pgno_t next, freecnt;
2067 txnid_t oldpg_txnid, id;
2070 assert(txn != NULL);
2071 assert(txn->mt_env != NULL);
2073 if (txn->mt_child) {
2074 mdb_txn_commit(txn->mt_child);
2075 txn->mt_child = NULL;
2080 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2081 if (txn->mt_numdbs > env->me_numdbs) {
2082 /* update the DB flags */
2084 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2085 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2092 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2093 DPUTS("error flag is set, can't commit");
2095 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2100 if (txn->mt_parent) {
2101 MDB_txn *parent = txn->mt_parent;
2105 parent->mt_next_pgno = txn->mt_next_pgno;
2106 parent->mt_flags = txn->mt_flags;
2108 /* Merge (and close) our cursors with parent's */
2109 mdb_cursor_merge(txn);
2111 /* Update parent's DB table. */
2112 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2113 memcpy(parent->mt_dbflags, txn->mt_dbflags, txn->mt_numdbs);
2114 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2116 /* Append our free list to parent's */
2117 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
2119 mdb_midl_free(txn->mt_free_pgs);
2121 /* Merge our dirty list with parent's */
2122 dst = txn->mt_parent->mt_u.dirty_list;
2123 src = txn->mt_u.dirty_list;
2124 x = mdb_mid2l_search(dst, src[1].mid);
2125 for (y=1; y<=src[0].mid; y++) {
2126 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
2130 dst[x].mptr = src[y].mptr;
2133 for (; y<=src[0].mid; y++) {
2134 if (++x >= MDB_IDL_UM_MAX) {
2136 return MDB_TXN_FULL;
2141 free(txn->mt_u.dirty_list);
2142 txn->mt_parent->mt_child = NULL;
2143 free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pgfree);
2148 if (txn != env->me_txn) {
2149 DPUTS("attempt to commit unknown transaction");
2154 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2157 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2158 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2160 /* Update DB root pointers */
2161 if (txn->mt_numdbs > 2) {
2164 data.mv_size = sizeof(MDB_db);
2166 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2167 for (i = 2; i < txn->mt_numdbs; i++) {
2168 if (txn->mt_dbflags[i] & DB_DIRTY) {
2169 data.mv_data = &txn->mt_dbs[i];
2170 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2177 /* Save the freelist as of this transaction to the freeDB. This
2178 * can change the freelist, so keep trying until it stabilizes.
2180 * env->me_pglast and the length of txn->mt_free_pgs cannot decrease,
2181 * except the code below can decrease env->me_pglast to split pghead.
2182 * Page numbers cannot disappear from txn->mt_free_pgs. New pages
2183 * can only appear in env->me_pghead when env->me_pglast increases.
2184 * Until then, the me_pghead pointer won't move but can become NULL.
2187 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2188 oldpg_txnid = id = 0;
2191 /* should only be one record now */
2192 if (env->me_pghead || env->me_pglast) {
2193 /* make sure first page of freeDB is touched and on freelist */
2194 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2195 if (rc && rc != MDB_NOTFOUND) {
2202 /* Delete IDLs we used from the free list */
2203 if (env->me_pglast) {
2208 rc = mdb_cursor_first(&mc, &key, NULL);
2211 oldpg_txnid = *(txnid_t *)key.mv_data;
2213 assert(oldpg_txnid <= env->me_pglast);
2215 rc = mdb_cursor_del(&mc, 0);
2218 } while (oldpg_txnid < env->me_pglast);
2221 /* Save IDL of pages freed by this txn, to freeDB */
2223 if (freecnt != txn->mt_free_pgs[0]) {
2226 /* make sure last page of freeDB is touched and on freelist */
2227 key.mv_size = MDB_MAXKEYSIZE+1;
2229 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2230 if (rc && rc != MDB_NOTFOUND)
2236 MDB_IDL idl = txn->mt_free_pgs;
2237 mdb_midl_sort(txn->mt_free_pgs);
2238 DPRINTF("IDL write txn %zu root %zu num %zu",
2239 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2240 for (i=1; i<=idl[0]; i++) {
2241 DPRINTF("IDL %zu", idl[i]);
2245 /* write to last page of freeDB */
2246 key.mv_size = sizeof(pgno_t);
2247 key.mv_data = &txn->mt_txnid;
2248 /* The free list can still grow during this call,
2249 * despite the pre-emptive touches above. So retry
2250 * until the reserved space remains big enough.
2253 assert(freecnt < txn->mt_free_pgs[0]);
2254 freecnt = txn->mt_free_pgs[0];
2255 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2256 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2259 } while (freecnt != txn->mt_free_pgs[0]);
2260 mdb_midl_sort(txn->mt_free_pgs);
2261 memcpy(data.mv_data, txn->mt_free_pgs, data.mv_size);
2262 if (oldpg_txnid < env->me_pglast || (!env->me_pghead && id))
2263 goto free_pgfirst; /* used up freeDB[oldpg_txnid] */
2266 /* Put back page numbers we took from freeDB but did not use */
2267 if (env->me_pghead) {
2272 mop = env->me_pghead;
2273 id = env->me_pglast;
2274 key.mv_size = sizeof(id);
2276 /* These steps may grow the freelist again
2277 * due to freed overflow pages...
2282 if (orig > env->me_maxfree_1pg && id > 4)
2283 orig = env->me_maxfree_1pg; /* Do not use more than 1 page */
2284 data.mv_size = (orig + 1) * sizeof(pgno_t);
2285 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2288 assert(!env->me_pghead || env->me_pglast);
2289 /* mop could have been used again here */
2290 if (id != env->me_pglast || env->me_pghead == NULL)
2291 goto again; /* was completely used up */
2292 assert(mop == env->me_pghead);
2293 } while (mop[0] < orig && --i);
2294 memcpy(data.mv_data, mop, data.mv_size);
2297 *(pgno_t *)data.mv_data = orig;
2298 mop[orig] = mop[0] - orig;
2299 env->me_pghead = mop += orig;
2300 /* Save more oldpages at the previous txnid. */
2301 assert(env->me_pglast == id && id == oldpg_txnid);
2302 env->me_pglast = --oldpg_txnid;
2306 /* Check for growth of freelist again */
2307 if (freecnt != txn->mt_free_pgs[0])
2310 free(env->me_pgfree);
2311 env->me_pghead = env->me_pgfree = NULL;
2313 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2314 if (mdb_midl_shrink(&txn->mt_free_pgs))
2315 env->me_free_pgs = txn->mt_free_pgs;
2322 if (env->me_flags & MDB_WRITEMAP) {
2323 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2324 dp = txn->mt_u.dirty_list[i].mptr;
2325 /* clear dirty flag */
2326 dp->mp_flags &= ~P_DIRTY;
2327 txn->mt_u.dirty_list[i].mid = 0;
2329 txn->mt_u.dirty_list[0].mid = 0;
2333 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2339 /* Windows actually supports scatter/gather I/O, but only on
2340 * unbuffered file handles. Since we're relying on the OS page
2341 * cache for all our data, that's self-defeating. So we just
2342 * write pages one at a time. We use the ov structure to set
2343 * the write offset, to at least save the overhead of a Seek
2347 memset(&ov, 0, sizeof(ov));
2348 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2350 dp = txn->mt_u.dirty_list[i].mptr;
2351 DPRINTF("committing page %zu", dp->mp_pgno);
2352 size = dp->mp_pgno * env->me_psize;
2353 ov.Offset = size & 0xffffffff;
2354 ov.OffsetHigh = size >> 16;
2355 ov.OffsetHigh >>= 16;
2356 /* clear dirty flag */
2357 dp->mp_flags &= ~P_DIRTY;
2358 wsize = env->me_psize;
2359 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2360 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2363 DPRINTF("WriteFile: %d", n);
2370 struct iovec iov[MDB_COMMIT_PAGES];
2374 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2375 dp = txn->mt_u.dirty_list[i].mptr;
2376 if (dp->mp_pgno != next) {
2378 rc = writev(env->me_fd, iov, n);
2382 DPUTS("short write, filesystem full?");
2384 DPRINTF("writev: %s", strerror(n));
2391 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2394 DPRINTF("committing page %zu", dp->mp_pgno);
2395 iov[n].iov_len = env->me_psize;
2396 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2397 iov[n].iov_base = (char *)dp;
2398 size += iov[n].iov_len;
2399 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2400 /* clear dirty flag */
2401 dp->mp_flags &= ~P_DIRTY;
2402 if (++n >= MDB_COMMIT_PAGES) {
2412 rc = writev(env->me_fd, iov, n);
2416 DPUTS("short write, filesystem full?");
2418 DPRINTF("writev: %s", strerror(n));
2425 /* Drop the dirty pages.
2427 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2428 dp = txn->mt_u.dirty_list[i].mptr;
2429 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2430 mdb_page_free(txn->mt_env, dp);
2432 VGMEMP_FREE(txn->mt_env, dp);
2435 txn->mt_u.dirty_list[i].mid = 0;
2437 txn->mt_u.dirty_list[0].mid = 0;
2440 if ((n = mdb_env_sync(env, 0)) != 0 ||
2441 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2449 if (txn->mt_numdbs > env->me_numdbs) {
2450 /* update the DB flags */
2452 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2453 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2457 UNLOCK_MUTEX_W(env);
2463 /** Read the environment parameters of a DB environment before
2464 * mapping it into memory.
2465 * @param[in] env the environment handle
2466 * @param[out] meta address of where to store the meta information
2467 * @return 0 on success, non-zero on failure.
2470 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2477 /* We don't know the page size yet, so use a minimum value.
2478 * Read both meta pages so we can use the latest one.
2481 for (i=0; i<2; i++) {
2483 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2485 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2490 else if (rc != MDB_PAGESIZE) {
2494 DPRINTF("read: %s", strerror(err));
2498 p = (MDB_page *)&pbuf;
2500 if (!F_ISSET(p->mp_flags, P_META)) {
2501 DPRINTF("page %zu not a meta page", p->mp_pgno);
2506 if (m->mm_magic != MDB_MAGIC) {
2507 DPUTS("meta has invalid magic");
2511 if (m->mm_version != MDB_VERSION) {
2512 DPRINTF("database is version %u, expected version %u",
2513 m->mm_version, MDB_VERSION);
2514 return MDB_VERSION_MISMATCH;
2518 if (m->mm_txnid > meta->mm_txnid)
2519 memcpy(meta, m, sizeof(*m));
2521 memcpy(meta, m, sizeof(*m));
2523 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2525 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2533 /** Write the environment parameters of a freshly created DB environment.
2534 * @param[in] env the environment handle
2535 * @param[out] meta address of where to store the meta information
2536 * @return 0 on success, non-zero on failure.
2539 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2546 DPUTS("writing new meta page");
2548 GET_PAGESIZE(psize);
2550 meta->mm_magic = MDB_MAGIC;
2551 meta->mm_version = MDB_VERSION;
2552 meta->mm_mapsize = env->me_mapsize;
2553 meta->mm_psize = psize;
2554 meta->mm_last_pg = 1;
2555 meta->mm_flags = env->me_flags & 0xffff;
2556 meta->mm_flags |= MDB_INTEGERKEY;
2557 meta->mm_dbs[0].md_root = P_INVALID;
2558 meta->mm_dbs[1].md_root = P_INVALID;
2560 p = calloc(2, psize);
2562 p->mp_flags = P_META;
2565 memcpy(m, meta, sizeof(*meta));
2567 q = (MDB_page *)((char *)p + psize);
2570 q->mp_flags = P_META;
2573 memcpy(m, meta, sizeof(*meta));
2578 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2579 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2580 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2583 lseek(env->me_fd, 0, SEEK_SET);
2584 rc = write(env->me_fd, p, psize * 2);
2585 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2591 /** Update the environment info to commit a transaction.
2592 * @param[in] txn the transaction that's being committed
2593 * @return 0 on success, non-zero on failure.
2596 mdb_env_write_meta(MDB_txn *txn)
2599 MDB_meta meta, metab, *mp;
2601 int rc, len, toggle;
2608 assert(txn != NULL);
2609 assert(txn->mt_env != NULL);
2611 toggle = !txn->mt_toggle;
2612 DPRINTF("writing meta page %d for root page %zu",
2613 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2616 mp = env->me_metas[toggle];
2618 if (env->me_flags & MDB_WRITEMAP) {
2619 /* Persist any increases of mapsize config */
2620 if (env->me_mapsize > mp->mm_mapsize)
2621 mp->mm_mapsize = env->me_mapsize;
2622 mp->mm_dbs[0] = txn->mt_dbs[0];
2623 mp->mm_dbs[1] = txn->mt_dbs[1];
2624 mp->mm_last_pg = txn->mt_next_pgno - 1;
2625 mp->mm_txnid = txn->mt_txnid;
2626 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2627 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2630 ptr += env->me_psize;
2631 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2638 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2639 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2641 ptr = (char *)&meta;
2642 if (env->me_mapsize > mp->mm_mapsize) {
2643 /* Persist any increases of mapsize config */
2644 meta.mm_mapsize = env->me_mapsize;
2645 off = offsetof(MDB_meta, mm_mapsize);
2647 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2649 len = sizeof(MDB_meta) - off;
2652 meta.mm_dbs[0] = txn->mt_dbs[0];
2653 meta.mm_dbs[1] = txn->mt_dbs[1];
2654 meta.mm_last_pg = txn->mt_next_pgno - 1;
2655 meta.mm_txnid = txn->mt_txnid;
2658 off += env->me_psize;
2661 /* Write to the SYNC fd */
2662 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2663 env->me_fd : env->me_mfd;
2666 memset(&ov, 0, sizeof(ov));
2668 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2671 rc = pwrite(mfd, ptr, len, off);
2676 DPUTS("write failed, disk error?");
2677 /* On a failure, the pagecache still contains the new data.
2678 * Write some old data back, to prevent it from being used.
2679 * Use the non-SYNC fd; we know it will fail anyway.
2681 meta.mm_last_pg = metab.mm_last_pg;
2682 meta.mm_txnid = metab.mm_txnid;
2684 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2686 r2 = pwrite(env->me_fd, ptr, len, off);
2689 env->me_flags |= MDB_FATAL_ERROR;
2693 /* Memory ordering issues are irrelevant; since the entire writer
2694 * is wrapped by wmutex, all of these changes will become visible
2695 * after the wmutex is unlocked. Since the DB is multi-version,
2696 * readers will get consistent data regardless of how fresh or
2697 * how stale their view of these values is.
2699 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2704 /** Check both meta pages to see which one is newer.
2705 * @param[in] env the environment handle
2706 * @return meta toggle (0 or 1).
2709 mdb_env_pick_meta(const MDB_env *env)
2711 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2715 mdb_env_create(MDB_env **env)
2719 e = calloc(1, sizeof(MDB_env));
2723 e->me_free_pgs = mdb_midl_alloc();
2724 if (!e->me_free_pgs) {
2728 e->me_maxreaders = DEFAULT_READERS;
2730 e->me_fd = INVALID_HANDLE_VALUE;
2731 e->me_lfd = INVALID_HANDLE_VALUE;
2732 e->me_mfd = INVALID_HANDLE_VALUE;
2733 #ifdef MDB_USE_POSIX_SEM
2734 e->me_rmutex = SEM_FAILED;
2735 e->me_wmutex = SEM_FAILED;
2737 e->me_pid = getpid();
2738 VGMEMP_CREATE(e,0,0);
2744 mdb_env_set_mapsize(MDB_env *env, size_t size)
2748 env->me_mapsize = size;
2750 env->me_maxpg = env->me_mapsize / env->me_psize;
2755 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2759 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2764 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2766 if (env->me_map || readers < 1)
2768 env->me_maxreaders = readers;
2773 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2775 if (!env || !readers)
2777 *readers = env->me_maxreaders;
2781 /** Further setup required for opening an MDB environment
2784 mdb_env_open2(MDB_env *env)
2786 unsigned int flags = env->me_flags;
2787 int i, newenv = 0, prot;
2791 memset(&meta, 0, sizeof(meta));
2793 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2796 DPUTS("new mdbenv");
2800 /* Was a mapsize configured? */
2801 if (!env->me_mapsize) {
2802 /* If this is a new environment, take the default,
2803 * else use the size recorded in the existing env.
2805 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2806 } else if (env->me_mapsize < meta.mm_mapsize) {
2807 /* If the configured size is smaller, make sure it's
2808 * still big enough. Silently round up to minimum if not.
2810 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2811 if (env->me_mapsize < minsize)
2812 env->me_mapsize = minsize;
2818 LONG sizelo, sizehi;
2819 sizelo = env->me_mapsize & 0xffffffff;
2820 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2822 /* Windows won't create mappings for zero length files.
2823 * Just allocate the maxsize right now.
2826 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2827 if (!SetEndOfFile(env->me_fd))
2829 SetFilePointer(env->me_fd, 0, NULL, 0);
2831 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2832 PAGE_READWRITE : PAGE_READONLY,
2833 sizehi, sizelo, NULL);
2836 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2837 FILE_MAP_WRITE : FILE_MAP_READ,
2838 0, 0, env->me_mapsize, meta.mm_address);
2846 if (flags & MDB_WRITEMAP) {
2848 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2851 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2853 if (env->me_map == MAP_FAILED) {
2860 if (flags & MDB_FIXEDMAP)
2861 meta.mm_address = env->me_map;
2862 i = mdb_env_init_meta(env, &meta);
2863 if (i != MDB_SUCCESS) {
2866 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2867 /* Can happen because the address argument to mmap() is just a
2868 * hint. mmap() can pick another, e.g. if the range is in use.
2869 * The MAP_FIXED flag would prevent that, but then mmap could
2870 * instead unmap existing pages to make room for the new map.
2872 return EBUSY; /* TODO: Make a new MDB_* error code? */
2874 env->me_psize = meta.mm_psize;
2875 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2877 env->me_maxpg = env->me_mapsize / env->me_psize;
2879 p = (MDB_page *)env->me_map;
2880 env->me_metas[0] = METADATA(p);
2881 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2885 int toggle = mdb_env_pick_meta(env);
2886 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2888 DPRINTF("opened database version %u, pagesize %u",
2889 env->me_metas[0]->mm_version, env->me_psize);
2890 DPRINTF("using meta page %d", toggle);
2891 DPRINTF("depth: %u", db->md_depth);
2892 DPRINTF("entries: %zu", db->md_entries);
2893 DPRINTF("branch pages: %zu", db->md_branch_pages);
2894 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2895 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2896 DPRINTF("root: %zu", db->md_root);
2904 /** Release a reader thread's slot in the reader lock table.
2905 * This function is called automatically when a thread exits.
2906 * @param[in] ptr This points to the slot in the reader lock table.
2909 mdb_env_reader_dest(void *ptr)
2911 MDB_reader *reader = ptr;
2917 /** Junk for arranging thread-specific callbacks on Windows. This is
2918 * necessarily platform and compiler-specific. Windows supports up
2919 * to 1088 keys. Let's assume nobody opens more than 64 environments
2920 * in a single process, for now. They can override this if needed.
2922 #ifndef MAX_TLS_KEYS
2923 #define MAX_TLS_KEYS 64
2925 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2926 static int mdb_tls_nkeys;
2928 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2932 case DLL_PROCESS_ATTACH: break;
2933 case DLL_THREAD_ATTACH: break;
2934 case DLL_THREAD_DETACH:
2935 for (i=0; i<mdb_tls_nkeys; i++) {
2936 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2937 mdb_env_reader_dest(r);
2940 case DLL_PROCESS_DETACH: break;
2945 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2947 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2951 /* Force some symbol references.
2952 * _tls_used forces the linker to create the TLS directory if not already done
2953 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2955 #pragma comment(linker, "/INCLUDE:_tls_used")
2956 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2957 #pragma const_seg(".CRT$XLB")
2958 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2959 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2962 #pragma comment(linker, "/INCLUDE:__tls_used")
2963 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2964 #pragma data_seg(".CRT$XLB")
2965 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2967 #endif /* WIN 32/64 */
2968 #endif /* !__GNUC__ */
2971 /** Downgrade the exclusive lock on the region back to shared */
2973 mdb_env_share_locks(MDB_env *env, int *excl)
2975 int rc = 0, toggle = mdb_env_pick_meta(env);
2977 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2982 /* First acquire a shared lock. The Unlock will
2983 * then release the existing exclusive lock.
2985 memset(&ov, 0, sizeof(ov));
2986 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2989 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2995 struct flock lock_info;
2996 /* The shared lock replaces the existing lock */
2997 memset((void *)&lock_info, 0, sizeof(lock_info));
2998 lock_info.l_type = F_RDLCK;
2999 lock_info.l_whence = SEEK_SET;
3000 lock_info.l_start = 0;
3001 lock_info.l_len = 1;
3002 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3003 (rc = ErrCode()) == EINTR) ;
3004 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3011 /** Try to get exlusive lock, otherwise shared.
3012 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3015 mdb_env_excl_lock(MDB_env *env, int *excl)
3019 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3023 memset(&ov, 0, sizeof(ov));
3024 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3031 struct flock lock_info;
3032 memset((void *)&lock_info, 0, sizeof(lock_info));
3033 lock_info.l_type = F_WRLCK;
3034 lock_info.l_whence = SEEK_SET;
3035 lock_info.l_start = 0;
3036 lock_info.l_len = 1;
3037 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3038 (rc = ErrCode()) == EINTR) ;
3042 # ifdef MDB_USE_POSIX_SEM
3043 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3046 lock_info.l_type = F_RDLCK;
3047 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3048 (rc = ErrCode()) == EINTR) ;
3056 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3058 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3060 * @(#) $Revision: 5.1 $
3061 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3062 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3064 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3068 * Please do not copyright this code. This code is in the public domain.
3070 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3071 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3072 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3073 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3074 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3075 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3076 * PERFORMANCE OF THIS SOFTWARE.
3079 * chongo <Landon Curt Noll> /\oo/\
3080 * http://www.isthe.com/chongo/
3082 * Share and Enjoy! :-)
3085 typedef unsigned long long mdb_hash_t;
3086 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3088 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3089 * @param[in] str string to hash
3090 * @param[in] hval initial value for hash
3091 * @return 64 bit hash
3093 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3094 * hval arg on the first call.
3097 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3099 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3100 unsigned char *end = s + val->mv_size;
3102 * FNV-1a hash each octet of the string
3105 /* xor the bottom with the current octet */
3106 hval ^= (mdb_hash_t)*s++;
3108 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3109 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3110 (hval << 7) + (hval << 8) + (hval << 40);
3112 /* return our new hash value */
3116 /** Hash the string and output the hash in hex.
3117 * @param[in] str string to hash
3118 * @param[out] hexbuf an array of 17 chars to hold the hash
3121 mdb_hash_hex(MDB_val *val, char *hexbuf)
3124 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3125 for (i=0; i<8; i++) {
3126 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3132 /** Open and/or initialize the lock region for the environment.
3133 * @param[in] env The MDB environment.
3134 * @param[in] lpath The pathname of the file used for the lock region.
3135 * @param[in] mode The Unix permissions for the file, if we create it.
3136 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3137 * @return 0 on success, non-zero on failure.
3140 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3148 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3149 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3150 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
3152 if (rc == ERROR_WRITE_PROTECT && (env->me_flags & MDB_RDONLY)) {
3153 env->me_flags |= MDB_ROFS;
3158 /* Try to get exclusive lock. If we succeed, then
3159 * nobody is using the lock region and we should initialize it.
3161 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3162 size = GetFileSize(env->me_lfd, NULL);
3168 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
3170 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3171 env->me_flags |= MDB_ROFS;
3176 /* Lose record locks when exec*() */
3177 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3178 fcntl(env->me_lfd, F_SETFD, fdflags);
3180 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
3181 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1) {
3183 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3184 env->me_flags |= MDB_ROFS;
3191 /* Try to get exclusive lock. If we succeed, then
3192 * nobody is using the lock region and we should initialize it.
3194 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3196 size = lseek(env->me_lfd, 0, SEEK_END);
3198 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3199 if (size < rsize && *excl > 0) {
3201 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3202 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3204 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3208 size = rsize - sizeof(MDB_txninfo);
3209 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3214 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3216 if (!mh) goto fail_errno;
3217 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3219 if (!env->me_txns) goto fail_errno;
3221 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3223 if (m == MAP_FAILED) goto fail_errno;
3229 BY_HANDLE_FILE_INFORMATION stbuf;
3238 if (!mdb_sec_inited) {
3239 InitializeSecurityDescriptor(&mdb_null_sd,
3240 SECURITY_DESCRIPTOR_REVISION);
3241 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3242 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3243 mdb_all_sa.bInheritHandle = FALSE;
3244 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3247 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3248 idbuf.volume = stbuf.dwVolumeSerialNumber;
3249 idbuf.nhigh = stbuf.nFileIndexHigh;
3250 idbuf.nlow = stbuf.nFileIndexLow;
3251 val.mv_data = &idbuf;
3252 val.mv_size = sizeof(idbuf);
3253 mdb_hash_hex(&val, hexbuf);
3254 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3255 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3256 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3257 if (!env->me_rmutex) goto fail_errno;
3258 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3259 if (!env->me_wmutex) goto fail_errno;
3260 #elif defined(MDB_USE_POSIX_SEM)
3269 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3270 idbuf.dev = stbuf.st_dev;
3271 idbuf.ino = stbuf.st_ino;
3272 val.mv_data = &idbuf;
3273 val.mv_size = sizeof(idbuf);
3274 mdb_hash_hex(&val, hexbuf);
3275 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3276 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3277 /* Clean up after a previous run, if needed: Try to
3278 * remove both semaphores before doing anything else.
3280 sem_unlink(env->me_txns->mti_rmname);
3281 sem_unlink(env->me_txns->mti_wmname);
3282 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3283 O_CREAT|O_EXCL, mode, 1);
3284 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3285 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3286 O_CREAT|O_EXCL, mode, 1);
3287 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3288 #else /* MDB_USE_POSIX_SEM */
3289 pthread_mutexattr_t mattr;
3291 if ((rc = pthread_mutexattr_init(&mattr))
3292 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3293 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3294 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3296 pthread_mutexattr_destroy(&mattr);
3297 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3299 env->me_txns->mti_version = MDB_VERSION;
3300 env->me_txns->mti_magic = MDB_MAGIC;
3301 env->me_txns->mti_txnid = 0;
3302 env->me_txns->mti_numreaders = 0;
3305 if (env->me_txns->mti_magic != MDB_MAGIC) {
3306 DPUTS("lock region has invalid magic");
3310 if (env->me_txns->mti_version != MDB_VERSION) {
3311 DPRINTF("lock region is version %u, expected version %u",
3312 env->me_txns->mti_version, MDB_VERSION);
3313 rc = MDB_VERSION_MISMATCH;
3317 if (rc != EACCES && rc != EAGAIN) {
3321 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3322 if (!env->me_rmutex) goto fail_errno;
3323 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3324 if (!env->me_wmutex) goto fail_errno;
3325 #elif defined(MDB_USE_POSIX_SEM)
3326 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3327 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3328 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3329 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3340 /** The name of the lock file in the DB environment */
3341 #define LOCKNAME "/lock.mdb"
3342 /** The name of the data file in the DB environment */
3343 #define DATANAME "/data.mdb"
3344 /** The suffix of the lock file when no subdir is used */
3345 #define LOCKSUFF "-lock"
3346 /** Only a subset of the @ref mdb_env flags can be changed
3347 * at runtime. Changing other flags requires closing the
3348 * environment and re-opening it with the new flags.
3350 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3351 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP)
3354 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3356 int oflags, rc, len, excl;
3357 char *lpath, *dpath;
3359 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3363 if (flags & MDB_NOSUBDIR) {
3364 rc = len + sizeof(LOCKSUFF) + len + 1;
3366 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3371 if (flags & MDB_NOSUBDIR) {
3372 dpath = lpath + len + sizeof(LOCKSUFF);
3373 sprintf(lpath, "%s" LOCKSUFF, path);
3374 strcpy(dpath, path);
3376 dpath = lpath + len + sizeof(LOCKNAME);
3377 sprintf(lpath, "%s" LOCKNAME, path);
3378 sprintf(dpath, "%s" DATANAME, path);
3381 flags |= env->me_flags;
3382 /* silently ignore WRITEMAP if we're only getting read access */
3383 if (F_ISSET(flags, MDB_RDONLY|MDB_WRITEMAP))
3384 flags ^= MDB_WRITEMAP;
3385 env->me_flags = flags |= MDB_ENV_ACTIVE;
3387 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3392 if (F_ISSET(flags, MDB_RDONLY)) {
3393 oflags = GENERIC_READ;
3394 len = OPEN_EXISTING;
3396 oflags = GENERIC_READ|GENERIC_WRITE;
3399 mode = FILE_ATTRIBUTE_NORMAL;
3400 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3401 NULL, len, mode, NULL);
3403 if (F_ISSET(flags, MDB_RDONLY))
3406 oflags = O_RDWR | O_CREAT;
3408 env->me_fd = open(dpath, oflags, mode);
3410 if (env->me_fd == INVALID_HANDLE_VALUE) {
3415 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3416 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3417 env->me_mfd = env->me_fd;
3419 /* Synchronous fd for meta writes. Needed even with
3420 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3423 env->me_mfd = CreateFile(dpath, oflags,
3424 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3425 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3427 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3429 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3434 DPRINTF("opened dbenv %p", (void *) env);
3435 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3438 env->me_numdbs = 2; /* this notes that me_txkey was set */
3440 /* Windows TLS callbacks need help finding their TLS info. */
3441 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3442 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3449 rc = mdb_env_share_locks(env, &excl);
3453 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3454 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3455 env->me_path = strdup(path);
3456 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3462 mdb_env_close0(env, excl);
3468 /** Destroy resources from mdb_env_open() and clear our readers */
3470 mdb_env_close0(MDB_env *env, int excl)
3474 if (!(env->me_flags & MDB_ENV_ACTIVE))
3477 free(env->me_dbflags);
3481 if (env->me_numdbs) {
3482 pthread_key_delete(env->me_txkey);
3484 /* Delete our key from the global list */
3485 for (i=0; i<mdb_tls_nkeys; i++)
3486 if (mdb_tls_keys[i] == env->me_txkey) {
3487 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3495 munmap(env->me_map, env->me_mapsize);
3497 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3499 if (env->me_fd != INVALID_HANDLE_VALUE)
3502 pid_t pid = env->me_pid;
3503 /* Clearing readers is done in this function because
3504 * me_txkey with its destructor must be disabled first.
3506 for (i = env->me_numreaders; --i >= 0; )
3507 if (env->me_txns->mti_readers[i].mr_pid == pid)
3508 env->me_txns->mti_readers[i].mr_pid = 0;
3510 if (env->me_rmutex) {
3511 CloseHandle(env->me_rmutex);
3512 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3514 /* Windows automatically destroys the mutexes when
3515 * the last handle closes.
3517 #elif defined(MDB_USE_POSIX_SEM)
3518 if (env->me_rmutex != SEM_FAILED) {
3519 sem_close(env->me_rmutex);
3520 if (env->me_wmutex != SEM_FAILED)
3521 sem_close(env->me_wmutex);
3522 /* If we have the filelock: If we are the
3523 * only remaining user, clean up semaphores.
3526 mdb_env_excl_lock(env, &excl);
3528 sem_unlink(env->me_txns->mti_rmname);
3529 sem_unlink(env->me_txns->mti_wmname);
3533 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3535 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3538 /* Unlock the lockfile. Windows would have unlocked it
3539 * after closing anyway, but not necessarily at once.
3541 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3547 env->me_flags &= ~MDB_ENV_ACTIVE;
3551 mdb_env_copy(MDB_env *env, const char *path)
3553 MDB_txn *txn = NULL;
3557 HANDLE newfd = INVALID_HANDLE_VALUE;
3559 if (env->me_flags & MDB_NOSUBDIR) {
3560 lpath = (char *)path;
3563 len += sizeof(DATANAME);
3564 lpath = malloc(len);
3567 sprintf(lpath, "%s" DATANAME, path);
3570 /* The destination path must exist, but the destination file must not.
3571 * We don't want the OS to cache the writes, since the source data is
3572 * already in the OS cache.
3575 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3576 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3578 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3584 if (!(env->me_flags & MDB_NOSUBDIR))
3586 if (newfd == INVALID_HANDLE_VALUE) {
3591 #ifdef F_NOCACHE /* __APPLE__ */
3592 rc = fcntl(newfd, F_NOCACHE, 1);
3599 /* Do the lock/unlock of the reader mutex before starting the
3600 * write txn. Otherwise other read txns could block writers.
3602 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3606 if (!(env->me_flags & MDB_ROFS)) {
3607 /* We must start the actual read txn after blocking writers */
3608 mdb_txn_reset0(txn);
3610 /* Temporarily block writers until we snapshot the meta pages */
3613 rc = mdb_txn_renew0(txn);
3615 UNLOCK_MUTEX_W(env);
3620 wsize = env->me_psize * 2;
3624 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3625 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3628 rc = write(newfd, env->me_map, wsize);
3629 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3631 if (! (env->me_flags & MDB_ROFS))
3632 UNLOCK_MUTEX_W(env);
3637 ptr = env->me_map + wsize;
3638 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3639 #define MAX_WRITE 2147483648U
3643 if (wsize > MAX_WRITE)
3647 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3648 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3657 if (wsize > MAX_WRITE)
3661 wres = write(newfd, ptr, w2);
3662 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3671 if (newfd != INVALID_HANDLE_VALUE)
3678 mdb_env_close(MDB_env *env)
3685 VGMEMP_DESTROY(env);
3686 while ((dp = env->me_dpages) != NULL) {
3687 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3688 env->me_dpages = dp->mp_next;
3692 mdb_env_close0(env, 0);
3693 mdb_midl_free(env->me_free_pgs);
3697 /** Compare two items pointing at aligned size_t's */
3699 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3701 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3702 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3705 /** Compare two items pointing at aligned int's */
3707 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3709 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3710 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3713 /** Compare two items pointing at ints of unknown alignment.
3714 * Nodes and keys are guaranteed to be 2-byte aligned.
3717 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3719 #if BYTE_ORDER == LITTLE_ENDIAN
3720 unsigned short *u, *c;
3723 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3724 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3727 } while(!x && u > (unsigned short *)a->mv_data);
3730 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3734 /** Compare two items lexically */
3736 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3743 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3749 diff = memcmp(a->mv_data, b->mv_data, len);
3750 return diff ? diff : len_diff<0 ? -1 : len_diff;
3753 /** Compare two items in reverse byte order */
3755 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3757 const unsigned char *p1, *p2, *p1_lim;
3761 p1_lim = (const unsigned char *)a->mv_data;
3762 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3763 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3765 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3771 while (p1 > p1_lim) {
3772 diff = *--p1 - *--p2;
3776 return len_diff<0 ? -1 : len_diff;
3779 /** Search for key within a page, using binary search.
3780 * Returns the smallest entry larger or equal to the key.
3781 * If exactp is non-null, stores whether the found entry was an exact match
3782 * in *exactp (1 or 0).
3783 * Updates the cursor index with the index of the found entry.
3784 * If no entry larger or equal to the key is found, returns NULL.
3787 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3789 unsigned int i = 0, nkeys;
3792 MDB_page *mp = mc->mc_pg[mc->mc_top];
3793 MDB_node *node = NULL;
3798 nkeys = NUMKEYS(mp);
3803 COPY_PGNO(pgno, mp->mp_pgno);
3804 DPRINTF("searching %u keys in %s %spage %zu",
3805 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3812 low = IS_LEAF(mp) ? 0 : 1;
3814 cmp = mc->mc_dbx->md_cmp;
3816 /* Branch pages have no data, so if using integer keys,
3817 * alignment is guaranteed. Use faster mdb_cmp_int.
3819 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3820 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3827 nodekey.mv_size = mc->mc_db->md_pad;
3828 node = NODEPTR(mp, 0); /* fake */
3829 while (low <= high) {
3830 i = (low + high) >> 1;
3831 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3832 rc = cmp(key, &nodekey);
3833 DPRINTF("found leaf index %u [%s], rc = %i",
3834 i, DKEY(&nodekey), rc);
3843 while (low <= high) {
3844 i = (low + high) >> 1;
3846 node = NODEPTR(mp, i);
3847 nodekey.mv_size = NODEKSZ(node);
3848 nodekey.mv_data = NODEKEY(node);
3850 rc = cmp(key, &nodekey);
3853 DPRINTF("found leaf index %u [%s], rc = %i",
3854 i, DKEY(&nodekey), rc);
3856 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3857 i, DKEY(&nodekey), NODEPGNO(node), rc);
3868 if (rc > 0) { /* Found entry is less than the key. */
3869 i++; /* Skip to get the smallest entry larger than key. */
3871 node = NODEPTR(mp, i);
3874 *exactp = (rc == 0);
3875 /* store the key index */
3876 mc->mc_ki[mc->mc_top] = i;
3878 /* There is no entry larger or equal to the key. */
3881 /* nodeptr is fake for LEAF2 */
3887 mdb_cursor_adjust(MDB_cursor *mc, func)
3891 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3892 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3899 /** Pop a page off the top of the cursor's stack. */
3901 mdb_cursor_pop(MDB_cursor *mc)
3904 #ifndef MDB_DEBUG_SKIP
3905 MDB_page *top = mc->mc_pg[mc->mc_top];
3911 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3912 mc->mc_dbi, (void *) mc);
3916 /** Push a page onto the top of the cursor's stack. */
3918 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3920 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3921 mc->mc_dbi, (void *) mc);
3923 if (mc->mc_snum >= CURSOR_STACK) {
3924 assert(mc->mc_snum < CURSOR_STACK);
3925 return MDB_CURSOR_FULL;
3928 mc->mc_top = mc->mc_snum++;
3929 mc->mc_pg[mc->mc_top] = mp;
3930 mc->mc_ki[mc->mc_top] = 0;
3935 /** Find the address of the page corresponding to a given page number.
3936 * @param[in] txn the transaction for this access.
3937 * @param[in] pgno the page number for the page to retrieve.
3938 * @param[out] ret address of a pointer where the page's address will be stored.
3939 * @return 0 on success, non-zero on failure.
3942 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3946 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
3947 (txn->mt_env->me_flags & MDB_WRITEMAP)))
3951 MDB_ID2L dl = tx2->mt_u.dirty_list;
3953 unsigned x = mdb_mid2l_search(dl, pgno);
3954 if (x <= dl[0].mid && dl[x].mid == pgno) {
3959 } while ((tx2 = tx2->mt_parent) != NULL);
3962 if (pgno < txn->mt_next_pgno) {
3963 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3965 DPRINTF("page %zu not found", pgno);
3971 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3974 /** Search for the page a given key should be in.
3975 * Pushes parent pages on the cursor stack. This function continues a
3976 * search on a cursor that has already been initialized. (Usually by
3977 * #mdb_page_search() but also by #mdb_node_move().)
3978 * @param[in,out] mc the cursor for this operation.
3979 * @param[in] key the key to search for. If NULL, search for the lowest
3980 * page. (This is used by #mdb_cursor_first().)
3981 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3982 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3983 * @return 0 on success, non-zero on failure.
3986 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3988 MDB_page *mp = mc->mc_pg[mc->mc_top];
3993 while (IS_BRANCH(mp)) {
3997 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3998 assert(NUMKEYS(mp) > 1);
3999 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4001 if (key == NULL) /* Initialize cursor to first page. */
4003 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4004 /* cursor to last page */
4008 node = mdb_node_search(mc, key, &exact);
4010 i = NUMKEYS(mp) - 1;
4012 i = mc->mc_ki[mc->mc_top];
4021 DPRINTF("following index %u for key [%s]",
4023 assert(i < NUMKEYS(mp));
4024 node = NODEPTR(mp, i);
4026 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
4029 mc->mc_ki[mc->mc_top] = i;
4030 if ((rc = mdb_cursor_push(mc, mp)))
4034 if ((rc = mdb_page_touch(mc)) != 0)
4036 mp = mc->mc_pg[mc->mc_top];
4041 DPRINTF("internal error, index points to a %02X page!?",
4043 return MDB_CORRUPTED;
4046 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4047 key ? DKEY(key) : NULL);
4052 /** Search for the page a given key should be in.
4053 * Pushes parent pages on the cursor stack. This function just sets up
4054 * the search; it finds the root page for \b mc's database and sets this
4055 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4056 * called to complete the search.
4057 * @param[in,out] mc the cursor for this operation.
4058 * @param[in] key the key to search for. If NULL, search for the lowest
4059 * page. (This is used by #mdb_cursor_first().)
4060 * @param[in] modify If true, visited pages are updated with new page numbers.
4061 * @return 0 on success, non-zero on failure.
4064 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4069 /* Make sure the txn is still viable, then find the root from
4070 * the txn's db table.
4072 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4073 DPUTS("transaction has failed, must abort");
4076 /* Make sure we're using an up-to-date root */
4077 if (mc->mc_dbi > MAIN_DBI) {
4078 if ((*mc->mc_dbflag & DB_STALE) ||
4079 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4081 unsigned char dbflag = 0;
4082 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4083 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4086 if (*mc->mc_dbflag & DB_STALE) {
4089 MDB_node *leaf = mdb_node_search(&mc2,
4090 &mc->mc_dbx->md_name, &exact);
4092 return MDB_NOTFOUND;
4093 mdb_node_read(mc->mc_txn, leaf, &data);
4094 /* The txn may not know this DBI, or another process may
4095 * have dropped and recreated the DB with other flags.
4097 if (mc->mc_db->md_flags != *(uint16_t *)
4098 ((char *) data.mv_data + offsetof(MDB_db, md_flags)))
4099 return MDB_INCOMPATIBLE;
4100 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4102 if (flags & MDB_PS_MODIFY)
4104 *mc->mc_dbflag = dbflag;
4107 root = mc->mc_db->md_root;
4109 if (root == P_INVALID) { /* Tree is empty. */
4110 DPUTS("tree is empty");
4111 return MDB_NOTFOUND;
4116 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4117 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
4123 DPRINTF("db %u root page %zu has flags 0x%X",
4124 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4126 if (flags & MDB_PS_MODIFY) {
4127 if ((rc = mdb_page_touch(mc)))
4131 if (flags & MDB_PS_ROOTONLY)
4134 return mdb_page_search_root(mc, key, flags);
4137 /** Return the data associated with a given node.
4138 * @param[in] txn The transaction for this operation.
4139 * @param[in] leaf The node being read.
4140 * @param[out] data Updated to point to the node's data.
4141 * @return 0 on success, non-zero on failure.
4144 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4146 MDB_page *omp; /* overflow page */
4150 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4151 data->mv_size = NODEDSZ(leaf);
4152 data->mv_data = NODEDATA(leaf);
4156 /* Read overflow data.
4158 data->mv_size = NODEDSZ(leaf);
4159 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4160 if ((rc = mdb_page_get(txn, pgno, &omp))) {
4161 DPRINTF("read overflow page %zu failed", pgno);
4164 data->mv_data = METADATA(omp);
4170 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4171 MDB_val *key, MDB_val *data)
4180 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4182 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4185 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4189 mdb_cursor_init(&mc, txn, dbi, &mx);
4190 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4193 /** Find a sibling for a page.
4194 * Replaces the page at the top of the cursor's stack with the
4195 * specified sibling, if one exists.
4196 * @param[in] mc The cursor for this operation.
4197 * @param[in] move_right Non-zero if the right sibling is requested,
4198 * otherwise the left sibling.
4199 * @return 0 on success, non-zero on failure.
4202 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4208 if (mc->mc_snum < 2) {
4209 return MDB_NOTFOUND; /* root has no siblings */
4213 DPRINTF("parent page is page %zu, index %u",
4214 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4216 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4217 : (mc->mc_ki[mc->mc_top] == 0)) {
4218 DPRINTF("no more keys left, moving to %s sibling",
4219 move_right ? "right" : "left");
4220 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4221 /* undo cursor_pop before returning */
4228 mc->mc_ki[mc->mc_top]++;
4230 mc->mc_ki[mc->mc_top]--;
4231 DPRINTF("just moving to %s index key %u",
4232 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4234 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4236 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4237 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4240 mdb_cursor_push(mc, mp);
4245 /** Move the cursor to the next data item. */
4247 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4253 if (mc->mc_flags & C_EOF) {
4254 return MDB_NOTFOUND;
4257 assert(mc->mc_flags & C_INITIALIZED);
4259 mp = mc->mc_pg[mc->mc_top];
4261 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4262 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4263 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4264 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4265 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4266 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4270 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4271 if (op == MDB_NEXT_DUP)
4272 return MDB_NOTFOUND;
4276 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4278 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4279 DPUTS("=====> move to next sibling page");
4280 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4281 mc->mc_flags |= C_EOF;
4282 mc->mc_flags &= ~C_INITIALIZED;
4283 return MDB_NOTFOUND;
4285 mp = mc->mc_pg[mc->mc_top];
4286 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4288 mc->mc_ki[mc->mc_top]++;
4290 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4291 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4294 key->mv_size = mc->mc_db->md_pad;
4295 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4299 assert(IS_LEAF(mp));
4300 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4302 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4303 mdb_xcursor_init1(mc, leaf);
4306 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4309 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4310 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4311 if (rc != MDB_SUCCESS)
4316 MDB_GET_KEY(leaf, key);
4320 /** Move the cursor to the previous data item. */
4322 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4328 assert(mc->mc_flags & C_INITIALIZED);
4330 mp = mc->mc_pg[mc->mc_top];
4332 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4333 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4334 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4335 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4336 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4337 if (op != MDB_PREV || rc == MDB_SUCCESS)
4340 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4341 if (op == MDB_PREV_DUP)
4342 return MDB_NOTFOUND;
4347 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4349 if (mc->mc_ki[mc->mc_top] == 0) {
4350 DPUTS("=====> move to prev sibling page");
4351 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4352 mc->mc_flags &= ~C_INITIALIZED;
4353 return MDB_NOTFOUND;
4355 mp = mc->mc_pg[mc->mc_top];
4356 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4357 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4359 mc->mc_ki[mc->mc_top]--;
4361 mc->mc_flags &= ~C_EOF;
4363 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4364 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4367 key->mv_size = mc->mc_db->md_pad;
4368 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4372 assert(IS_LEAF(mp));
4373 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4375 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4376 mdb_xcursor_init1(mc, leaf);
4379 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4382 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4383 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4384 if (rc != MDB_SUCCESS)
4389 MDB_GET_KEY(leaf, key);
4393 /** Set the cursor on a specific data item. */
4395 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4396 MDB_cursor_op op, int *exactp)
4400 MDB_node *leaf = NULL;
4405 assert(key->mv_size > 0);
4407 /* See if we're already on the right page */
4408 if (mc->mc_flags & C_INITIALIZED) {
4411 mp = mc->mc_pg[mc->mc_top];
4413 mc->mc_ki[mc->mc_top] = 0;
4414 return MDB_NOTFOUND;
4416 if (mp->mp_flags & P_LEAF2) {
4417 nodekey.mv_size = mc->mc_db->md_pad;
4418 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4420 leaf = NODEPTR(mp, 0);
4421 MDB_GET_KEY(leaf, &nodekey);
4423 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4425 /* Probably happens rarely, but first node on the page
4426 * was the one we wanted.
4428 mc->mc_ki[mc->mc_top] = 0;
4435 unsigned int nkeys = NUMKEYS(mp);
4437 if (mp->mp_flags & P_LEAF2) {
4438 nodekey.mv_data = LEAF2KEY(mp,
4439 nkeys-1, nodekey.mv_size);
4441 leaf = NODEPTR(mp, nkeys-1);
4442 MDB_GET_KEY(leaf, &nodekey);
4444 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4446 /* last node was the one we wanted */
4447 mc->mc_ki[mc->mc_top] = nkeys-1;
4453 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4454 /* This is definitely the right page, skip search_page */
4455 if (mp->mp_flags & P_LEAF2) {
4456 nodekey.mv_data = LEAF2KEY(mp,
4457 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4459 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4460 MDB_GET_KEY(leaf, &nodekey);
4462 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4464 /* current node was the one we wanted */
4474 /* If any parents have right-sibs, search.
4475 * Otherwise, there's nothing further.
4477 for (i=0; i<mc->mc_top; i++)
4479 NUMKEYS(mc->mc_pg[i])-1)
4481 if (i == mc->mc_top) {
4482 /* There are no other pages */
4483 mc->mc_ki[mc->mc_top] = nkeys;
4484 return MDB_NOTFOUND;
4488 /* There are no other pages */
4489 mc->mc_ki[mc->mc_top] = 0;
4490 return MDB_NOTFOUND;
4494 rc = mdb_page_search(mc, key, 0);
4495 if (rc != MDB_SUCCESS)
4498 mp = mc->mc_pg[mc->mc_top];
4499 assert(IS_LEAF(mp));
4502 leaf = mdb_node_search(mc, key, exactp);
4503 if (exactp != NULL && !*exactp) {
4504 /* MDB_SET specified and not an exact match. */
4505 return MDB_NOTFOUND;
4509 DPUTS("===> inexact leaf not found, goto sibling");
4510 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4511 return rc; /* no entries matched */
4512 mp = mc->mc_pg[mc->mc_top];
4513 assert(IS_LEAF(mp));
4514 leaf = NODEPTR(mp, 0);
4518 mc->mc_flags |= C_INITIALIZED;
4519 mc->mc_flags &= ~C_EOF;
4522 key->mv_size = mc->mc_db->md_pad;
4523 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4527 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4528 mdb_xcursor_init1(mc, leaf);
4531 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4532 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4533 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4536 if (op == MDB_GET_BOTH) {
4542 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4543 if (rc != MDB_SUCCESS)
4546 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4548 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4550 rc = mc->mc_dbx->md_dcmp(data, &d2);
4552 if (op == MDB_GET_BOTH || rc > 0)
4553 return MDB_NOTFOUND;
4558 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4559 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4564 /* The key already matches in all other cases */
4565 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4566 MDB_GET_KEY(leaf, key);
4567 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4572 /** Move the cursor to the first item in the database. */
4574 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4579 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4580 rc = mdb_page_search(mc, NULL, 0);
4581 if (rc != MDB_SUCCESS)
4584 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4586 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4587 mc->mc_flags |= C_INITIALIZED;
4588 mc->mc_flags &= ~C_EOF;
4590 mc->mc_ki[mc->mc_top] = 0;
4592 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4593 key->mv_size = mc->mc_db->md_pad;
4594 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4599 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4600 mdb_xcursor_init1(mc, leaf);
4601 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4606 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4607 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4611 MDB_GET_KEY(leaf, key);
4615 /** Move the cursor to the last item in the database. */
4617 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4622 if (!(mc->mc_flags & C_EOF)) {
4624 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4627 lkey.mv_size = MDB_MAXKEYSIZE+1;
4628 lkey.mv_data = NULL;
4629 rc = mdb_page_search(mc, &lkey, 0);
4630 if (rc != MDB_SUCCESS)
4633 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4635 mc->mc_flags |= C_INITIALIZED|C_EOF;
4636 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4638 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4640 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4641 key->mv_size = mc->mc_db->md_pad;
4642 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4647 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4648 mdb_xcursor_init1(mc, leaf);
4649 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4654 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4655 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4660 MDB_GET_KEY(leaf, key);
4665 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4674 case MDB_GET_CURRENT:
4675 if (!(mc->mc_flags & C_INITIALIZED)) {
4678 MDB_page *mp = mc->mc_pg[mc->mc_top];
4680 mc->mc_ki[mc->mc_top] = 0;
4686 key->mv_size = mc->mc_db->md_pad;
4687 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4689 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4690 MDB_GET_KEY(leaf, key);
4692 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4693 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4695 rc = mdb_node_read(mc->mc_txn, leaf, data);
4702 case MDB_GET_BOTH_RANGE:
4703 if (data == NULL || mc->mc_xcursor == NULL) {
4711 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4713 } else if (op == MDB_SET_RANGE)
4714 rc = mdb_cursor_set(mc, key, data, op, NULL);
4716 rc = mdb_cursor_set(mc, key, data, op, &exact);
4718 case MDB_GET_MULTIPLE:
4720 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4721 !(mc->mc_flags & C_INITIALIZED)) {
4726 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4727 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4730 case MDB_NEXT_MULTIPLE:
4732 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4736 if (!(mc->mc_flags & C_INITIALIZED))
4737 rc = mdb_cursor_first(mc, key, data);
4739 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4740 if (rc == MDB_SUCCESS) {
4741 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4744 mx = &mc->mc_xcursor->mx_cursor;
4745 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4747 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4748 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4756 case MDB_NEXT_NODUP:
4757 if (!(mc->mc_flags & C_INITIALIZED))
4758 rc = mdb_cursor_first(mc, key, data);
4760 rc = mdb_cursor_next(mc, key, data, op);
4764 case MDB_PREV_NODUP:
4765 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4766 rc = mdb_cursor_last(mc, key, data);
4767 mc->mc_flags |= C_INITIALIZED;
4768 mc->mc_ki[mc->mc_top]++;
4770 rc = mdb_cursor_prev(mc, key, data, op);
4773 rc = mdb_cursor_first(mc, key, data);
4777 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4778 !(mc->mc_flags & C_INITIALIZED) ||
4779 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4783 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4786 rc = mdb_cursor_last(mc, key, data);
4790 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4791 !(mc->mc_flags & C_INITIALIZED) ||
4792 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4796 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4799 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4807 /** Touch all the pages in the cursor stack.
4808 * Makes sure all the pages are writable, before attempting a write operation.
4809 * @param[in] mc The cursor to operate on.
4812 mdb_cursor_touch(MDB_cursor *mc)
4816 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4819 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4820 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4821 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4824 *mc->mc_dbflag = DB_DIRTY;
4826 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4827 rc = mdb_page_touch(mc);
4831 mc->mc_top = mc->mc_snum-1;
4836 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4839 MDB_node *leaf = NULL;
4840 MDB_val xdata, *rdata, dkey;
4843 int do_sub = 0, insert = 0;
4844 unsigned int mcount = 0;
4848 char dbuf[MDB_MAXKEYSIZE+1];
4849 unsigned int nflags;
4852 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4855 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4858 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4861 #if SIZE_MAX > MAXDATASIZE
4862 if (data->mv_size > MAXDATASIZE)
4866 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4867 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4871 if (flags == MDB_CURRENT) {
4872 if (!(mc->mc_flags & C_INITIALIZED))
4875 } else if (mc->mc_db->md_root == P_INVALID) {
4877 /* new database, write a root leaf page */
4878 DPUTS("allocating new root leaf page");
4879 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4883 mdb_cursor_push(mc, np);
4884 mc->mc_db->md_root = np->mp_pgno;
4885 mc->mc_db->md_depth++;
4886 *mc->mc_dbflag = DB_DIRTY;
4887 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4889 np->mp_flags |= P_LEAF2;
4890 mc->mc_flags |= C_INITIALIZED;
4896 if (flags & MDB_APPEND) {
4898 rc = mdb_cursor_last(mc, &k2, &d2);
4900 rc = mc->mc_dbx->md_cmp(key, &k2);
4903 mc->mc_ki[mc->mc_top]++;
4909 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4911 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4912 DPRINTF("duplicate key [%s]", DKEY(key));
4914 return MDB_KEYEXIST;
4916 if (rc && rc != MDB_NOTFOUND)
4920 /* Cursor is positioned, now make sure all pages are writable */
4921 rc2 = mdb_cursor_touch(mc);
4926 /* The key already exists */
4927 if (rc == MDB_SUCCESS) {
4928 /* there's only a key anyway, so this is a no-op */
4929 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4930 unsigned int ksize = mc->mc_db->md_pad;
4931 if (key->mv_size != ksize)
4933 if (flags == MDB_CURRENT) {
4934 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4935 memcpy(ptr, key->mv_data, ksize);
4940 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4943 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4944 /* Was a single item before, must convert now */
4946 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4947 /* Just overwrite the current item */
4948 if (flags == MDB_CURRENT)
4951 dkey.mv_size = NODEDSZ(leaf);
4952 dkey.mv_data = NODEDATA(leaf);
4953 #if UINT_MAX < SIZE_MAX
4954 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4955 #ifdef MISALIGNED_OK
4956 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4958 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4961 /* if data matches, ignore it */
4962 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4963 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4965 /* create a fake page for the dup items */
4966 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4967 dkey.mv_data = dbuf;
4968 fp = (MDB_page *)&pbuf;
4969 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4970 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4971 fp->mp_lower = PAGEHDRSZ;
4972 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4973 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4974 fp->mp_flags |= P_LEAF2;
4975 fp->mp_pad = data->mv_size;
4976 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4978 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4979 (dkey.mv_size & 1) + (data->mv_size & 1);
4981 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4984 xdata.mv_size = fp->mp_upper;
4989 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4990 /* See if we need to convert from fake page to subDB */
4992 unsigned int offset;
4995 fp = NODEDATA(leaf);
4996 if (flags == MDB_CURRENT) {
4998 fp->mp_flags |= P_DIRTY;
4999 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5000 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5004 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5005 offset = fp->mp_pad;
5006 if (SIZELEFT(fp) >= offset)
5008 offset *= 4; /* space for 4 more */
5010 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5012 offset += offset & 1;
5013 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5014 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
5016 /* yes, convert it */
5018 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5019 dummy.md_pad = fp->mp_pad;
5020 dummy.md_flags = MDB_DUPFIXED;
5021 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5022 dummy.md_flags |= MDB_INTEGERKEY;
5025 dummy.md_branch_pages = 0;
5026 dummy.md_leaf_pages = 1;
5027 dummy.md_overflow_pages = 0;
5028 dummy.md_entries = NUMKEYS(fp);
5030 xdata.mv_size = sizeof(MDB_db);
5031 xdata.mv_data = &dummy;
5032 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5034 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5035 flags |= F_DUPDATA|F_SUBDATA;
5036 dummy.md_root = mp->mp_pgno;
5038 /* no, just grow it */
5040 xdata.mv_size = NODEDSZ(leaf) + offset;
5041 xdata.mv_data = &pbuf;
5042 mp = (MDB_page *)&pbuf;
5043 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5046 mp->mp_flags = fp->mp_flags | P_DIRTY;
5047 mp->mp_pad = fp->mp_pad;
5048 mp->mp_lower = fp->mp_lower;
5049 mp->mp_upper = fp->mp_upper + offset;
5051 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5053 nsize = NODEDSZ(leaf) - fp->mp_upper;
5054 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5055 for (i=0; i<NUMKEYS(fp); i++)
5056 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5058 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5062 /* data is on sub-DB, just store it */
5063 flags |= F_DUPDATA|F_SUBDATA;
5067 /* overflow page overwrites need special handling */
5068 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5071 int ovpages, dpages;
5073 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5074 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5075 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5076 mdb_page_get(mc->mc_txn, pg, &omp);
5077 /* Is the ov page writable and large enough? */
5078 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5079 /* yes, overwrite it. Note in this case we don't
5080 * bother to try shrinking the node if the new data
5081 * is smaller than the overflow threshold.
5083 if (F_ISSET(flags, MDB_RESERVE))
5084 data->mv_data = METADATA(omp);
5086 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5089 /* no, free ovpages */
5091 mc->mc_db->md_overflow_pages -= ovpages;
5092 for (i=0; i<ovpages; i++) {
5093 DPRINTF("freed ov page %zu", pg);
5094 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5098 } else if (NODEDSZ(leaf) == data->mv_size) {
5099 /* same size, just replace it. Note that we could
5100 * also reuse this node if the new data is smaller,
5101 * but instead we opt to shrink the node in that case.
5103 if (F_ISSET(flags, MDB_RESERVE))
5104 data->mv_data = NODEDATA(leaf);
5105 else if (data->mv_size)
5106 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5108 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5111 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5112 mc->mc_db->md_entries--;
5114 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5121 nflags = flags & NODE_ADD_FLAGS;
5122 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5123 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5124 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5125 nflags &= ~MDB_APPEND;
5127 nflags |= MDB_SPLIT_REPLACE;
5128 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5130 /* There is room already in this leaf page. */
5131 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5132 if (rc == 0 && !do_sub && insert) {
5133 /* Adjust other cursors pointing to mp */
5134 MDB_cursor *m2, *m3;
5135 MDB_dbi dbi = mc->mc_dbi;
5136 unsigned i = mc->mc_top;
5137 MDB_page *mp = mc->mc_pg[i];
5139 if (mc->mc_flags & C_SUB)
5142 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5143 if (mc->mc_flags & C_SUB)
5144 m3 = &m2->mc_xcursor->mx_cursor;
5147 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5148 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5155 if (rc != MDB_SUCCESS)
5156 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5158 /* Now store the actual data in the child DB. Note that we're
5159 * storing the user data in the keys field, so there are strict
5160 * size limits on dupdata. The actual data fields of the child
5161 * DB are all zero size.
5168 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5169 if (flags & MDB_CURRENT) {
5170 xflags = MDB_CURRENT;
5172 mdb_xcursor_init1(mc, leaf);
5173 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5175 /* converted, write the original data first */
5177 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5181 /* Adjust other cursors pointing to mp */
5183 unsigned i = mc->mc_top;
5184 MDB_page *mp = mc->mc_pg[i];
5186 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5187 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5188 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5189 mdb_xcursor_init1(m2, leaf);
5194 if (flags & MDB_APPENDDUP)
5195 xflags |= MDB_APPEND;
5196 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5197 if (flags & F_SUBDATA) {
5198 void *db = NODEDATA(leaf);
5199 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5202 /* sub-writes might have failed so check rc again.
5203 * Don't increment count if we just replaced an existing item.
5205 if (!rc && !(flags & MDB_CURRENT))
5206 mc->mc_db->md_entries++;
5207 if (flags & MDB_MULTIPLE) {
5209 if (mcount < data[1].mv_size) {
5210 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5211 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5217 /* If we succeeded and the key didn't exist before, make sure
5218 * the cursor is marked valid.
5221 mc->mc_flags |= C_INITIALIZED;
5226 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5231 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5234 if (!(mc->mc_flags & C_INITIALIZED))
5237 rc = mdb_cursor_touch(mc);
5241 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5243 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5244 if (flags != MDB_NODUPDATA) {
5245 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5246 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5248 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5249 /* If sub-DB still has entries, we're done */
5250 if (mc->mc_xcursor->mx_db.md_entries) {
5251 if (leaf->mn_flags & F_SUBDATA) {
5252 /* update subDB info */
5253 void *db = NODEDATA(leaf);
5254 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5256 /* shrink fake page */
5257 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5259 mc->mc_db->md_entries--;
5262 /* otherwise fall thru and delete the sub-DB */
5265 if (leaf->mn_flags & F_SUBDATA) {
5266 /* add all the child DB's pages to the free list */
5267 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5268 if (rc == MDB_SUCCESS) {
5269 mc->mc_db->md_entries -=
5270 mc->mc_xcursor->mx_db.md_entries;
5275 return mdb_cursor_del0(mc, leaf);
5278 /** Allocate and initialize new pages for a database.
5279 * @param[in] mc a cursor on the database being added to.
5280 * @param[in] flags flags defining what type of page is being allocated.
5281 * @param[in] num the number of pages to allocate. This is usually 1,
5282 * unless allocating overflow pages for a large record.
5283 * @param[out] mp Address of a page, or NULL on failure.
5284 * @return 0 on success, non-zero on failure.
5287 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5292 if ((rc = mdb_page_alloc(mc, num, &np)))
5294 DPRINTF("allocated new mpage %zu, page size %u",
5295 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5296 np->mp_flags = flags | P_DIRTY;
5297 np->mp_lower = PAGEHDRSZ;
5298 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5301 mc->mc_db->md_branch_pages++;
5302 else if (IS_LEAF(np))
5303 mc->mc_db->md_leaf_pages++;
5304 else if (IS_OVERFLOW(np)) {
5305 mc->mc_db->md_overflow_pages += num;
5313 /** Calculate the size of a leaf node.
5314 * The size depends on the environment's page size; if a data item
5315 * is too large it will be put onto an overflow page and the node
5316 * size will only include the key and not the data. Sizes are always
5317 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5318 * of the #MDB_node headers.
5319 * @param[in] env The environment handle.
5320 * @param[in] key The key for the node.
5321 * @param[in] data The data for the node.
5322 * @return The number of bytes needed to store the node.
5325 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5329 sz = LEAFSIZE(key, data);
5330 if (sz >= env->me_psize / MDB_MINKEYS) {
5331 /* put on overflow page */
5332 sz -= data->mv_size - sizeof(pgno_t);
5336 return sz + sizeof(indx_t);
5339 /** Calculate the size of a branch node.
5340 * The size should depend on the environment's page size but since
5341 * we currently don't support spilling large keys onto overflow
5342 * pages, it's simply the size of the #MDB_node header plus the
5343 * size of the key. Sizes are always rounded up to an even number
5344 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5345 * @param[in] env The environment handle.
5346 * @param[in] key The key for the node.
5347 * @return The number of bytes needed to store the node.
5350 mdb_branch_size(MDB_env *env, MDB_val *key)
5355 if (sz >= env->me_psize / MDB_MINKEYS) {
5356 /* put on overflow page */
5357 /* not implemented */
5358 /* sz -= key->size - sizeof(pgno_t); */
5361 return sz + sizeof(indx_t);
5364 /** Add a node to the page pointed to by the cursor.
5365 * @param[in] mc The cursor for this operation.
5366 * @param[in] indx The index on the page where the new node should be added.
5367 * @param[in] key The key for the new node.
5368 * @param[in] data The data for the new node, if any.
5369 * @param[in] pgno The page number, if adding a branch node.
5370 * @param[in] flags Flags for the node.
5371 * @return 0 on success, non-zero on failure. Possible errors are:
5373 * <li>ENOMEM - failed to allocate overflow pages for the node.
5374 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5375 * should never happen since all callers already calculate the
5376 * page's free space before calling this function.
5380 mdb_node_add(MDB_cursor *mc, indx_t indx,
5381 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5384 size_t node_size = NODESIZE;
5387 MDB_page *mp = mc->mc_pg[mc->mc_top];
5388 MDB_page *ofp = NULL; /* overflow page */
5391 assert(mp->mp_upper >= mp->mp_lower);
5393 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5394 IS_LEAF(mp) ? "leaf" : "branch",
5395 IS_SUBP(mp) ? "sub-" : "",
5396 mp->mp_pgno, indx, data ? data->mv_size : 0,
5397 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5400 /* Move higher keys up one slot. */
5401 int ksize = mc->mc_db->md_pad, dif;
5402 char *ptr = LEAF2KEY(mp, indx, ksize);
5403 dif = NUMKEYS(mp) - indx;
5405 memmove(ptr+ksize, ptr, dif*ksize);
5406 /* insert new key */
5407 memcpy(ptr, key->mv_data, ksize);
5409 /* Just using these for counting */
5410 mp->mp_lower += sizeof(indx_t);
5411 mp->mp_upper -= ksize - sizeof(indx_t);
5416 node_size += key->mv_size;
5420 if (F_ISSET(flags, F_BIGDATA)) {
5421 /* Data already on overflow page. */
5422 node_size += sizeof(pgno_t);
5423 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5424 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5426 /* Put data on overflow page. */
5427 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5428 data->mv_size, node_size+data->mv_size);
5429 node_size += sizeof(pgno_t);
5430 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5432 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5435 node_size += data->mv_size;
5438 node_size += node_size & 1;
5440 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5441 DPRINTF("not enough room in page %zu, got %u ptrs",
5442 mp->mp_pgno, NUMKEYS(mp));
5443 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5444 mp->mp_upper - mp->mp_lower);
5445 DPRINTF("node size = %zu", node_size);
5446 return MDB_PAGE_FULL;
5449 /* Move higher pointers up one slot. */
5450 for (i = NUMKEYS(mp); i > indx; i--)
5451 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5453 /* Adjust free space offsets. */
5454 ofs = mp->mp_upper - node_size;
5455 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5456 mp->mp_ptrs[indx] = ofs;
5458 mp->mp_lower += sizeof(indx_t);
5460 /* Write the node data. */
5461 node = NODEPTR(mp, indx);
5462 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5463 node->mn_flags = flags;
5465 SETDSZ(node,data->mv_size);
5470 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5475 if (F_ISSET(flags, F_BIGDATA))
5476 memcpy(node->mn_data + key->mv_size, data->mv_data,
5478 else if (F_ISSET(flags, MDB_RESERVE))
5479 data->mv_data = node->mn_data + key->mv_size;
5481 memcpy(node->mn_data + key->mv_size, data->mv_data,
5484 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5486 if (F_ISSET(flags, MDB_RESERVE))
5487 data->mv_data = METADATA(ofp);
5489 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5496 /** Delete the specified node from a page.
5497 * @param[in] mp The page to operate on.
5498 * @param[in] indx The index of the node to delete.
5499 * @param[in] ksize The size of a node. Only used if the page is
5500 * part of a #MDB_DUPFIXED database.
5503 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5506 indx_t i, j, numkeys, ptr;
5513 COPY_PGNO(pgno, mp->mp_pgno);
5514 DPRINTF("delete node %u on %s page %zu", indx,
5515 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5518 assert(indx < NUMKEYS(mp));
5521 int x = NUMKEYS(mp) - 1 - indx;
5522 base = LEAF2KEY(mp, indx, ksize);
5524 memmove(base, base + ksize, x * ksize);
5525 mp->mp_lower -= sizeof(indx_t);
5526 mp->mp_upper += ksize - sizeof(indx_t);
5530 node = NODEPTR(mp, indx);
5531 sz = NODESIZE + node->mn_ksize;
5533 if (F_ISSET(node->mn_flags, F_BIGDATA))
5534 sz += sizeof(pgno_t);
5536 sz += NODEDSZ(node);
5540 ptr = mp->mp_ptrs[indx];
5541 numkeys = NUMKEYS(mp);
5542 for (i = j = 0; i < numkeys; i++) {
5544 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5545 if (mp->mp_ptrs[i] < ptr)
5546 mp->mp_ptrs[j] += sz;
5551 base = (char *)mp + mp->mp_upper;
5552 memmove(base + sz, base, ptr - mp->mp_upper);
5554 mp->mp_lower -= sizeof(indx_t);
5558 /** Compact the main page after deleting a node on a subpage.
5559 * @param[in] mp The main page to operate on.
5560 * @param[in] indx The index of the subpage on the main page.
5563 mdb_node_shrink(MDB_page *mp, indx_t indx)
5570 indx_t i, numkeys, ptr;
5572 node = NODEPTR(mp, indx);
5573 sp = (MDB_page *)NODEDATA(node);
5574 osize = NODEDSZ(node);
5576 delta = sp->mp_upper - sp->mp_lower;
5577 SETDSZ(node, osize - delta);
5578 xp = (MDB_page *)((char *)sp + delta);
5580 /* shift subpage upward */
5582 nsize = NUMKEYS(sp) * sp->mp_pad;
5583 memmove(METADATA(xp), METADATA(sp), nsize);
5586 nsize = osize - sp->mp_upper;
5587 numkeys = NUMKEYS(sp);
5588 for (i=numkeys-1; i>=0; i--)
5589 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5591 xp->mp_upper = sp->mp_lower;
5592 xp->mp_lower = sp->mp_lower;
5593 xp->mp_flags = sp->mp_flags;
5594 xp->mp_pad = sp->mp_pad;
5595 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5597 /* shift lower nodes upward */
5598 ptr = mp->mp_ptrs[indx];
5599 numkeys = NUMKEYS(mp);
5600 for (i = 0; i < numkeys; i++) {
5601 if (mp->mp_ptrs[i] <= ptr)
5602 mp->mp_ptrs[i] += delta;
5605 base = (char *)mp + mp->mp_upper;
5606 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5607 mp->mp_upper += delta;
5610 /** Initial setup of a sorted-dups cursor.
5611 * Sorted duplicates are implemented as a sub-database for the given key.
5612 * The duplicate data items are actually keys of the sub-database.
5613 * Operations on the duplicate data items are performed using a sub-cursor
5614 * initialized when the sub-database is first accessed. This function does
5615 * the preliminary setup of the sub-cursor, filling in the fields that
5616 * depend only on the parent DB.
5617 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5620 mdb_xcursor_init0(MDB_cursor *mc)
5622 MDB_xcursor *mx = mc->mc_xcursor;
5624 mx->mx_cursor.mc_xcursor = NULL;
5625 mx->mx_cursor.mc_txn = mc->mc_txn;
5626 mx->mx_cursor.mc_db = &mx->mx_db;
5627 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5628 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5629 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5630 mx->mx_cursor.mc_snum = 0;
5631 mx->mx_cursor.mc_top = 0;
5632 mx->mx_cursor.mc_flags = C_SUB;
5633 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5634 mx->mx_dbx.md_dcmp = NULL;
5635 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5638 /** Final setup of a sorted-dups cursor.
5639 * Sets up the fields that depend on the data from the main cursor.
5640 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5641 * @param[in] node The data containing the #MDB_db record for the
5642 * sorted-dup database.
5645 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5647 MDB_xcursor *mx = mc->mc_xcursor;
5649 if (node->mn_flags & F_SUBDATA) {
5650 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5651 mx->mx_cursor.mc_pg[0] = 0;
5652 mx->mx_cursor.mc_snum = 0;
5653 mx->mx_cursor.mc_flags = C_SUB;
5655 MDB_page *fp = NODEDATA(node);
5656 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5657 mx->mx_db.md_flags = 0;
5658 mx->mx_db.md_depth = 1;
5659 mx->mx_db.md_branch_pages = 0;
5660 mx->mx_db.md_leaf_pages = 1;
5661 mx->mx_db.md_overflow_pages = 0;
5662 mx->mx_db.md_entries = NUMKEYS(fp);
5663 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5664 mx->mx_cursor.mc_snum = 1;
5665 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5666 mx->mx_cursor.mc_top = 0;
5667 mx->mx_cursor.mc_pg[0] = fp;
5668 mx->mx_cursor.mc_ki[0] = 0;
5669 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5670 mx->mx_db.md_flags = MDB_DUPFIXED;
5671 mx->mx_db.md_pad = fp->mp_pad;
5672 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5673 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5676 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5678 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5680 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5681 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5682 #if UINT_MAX < SIZE_MAX
5683 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5684 #ifdef MISALIGNED_OK
5685 mx->mx_dbx.md_cmp = mdb_cmp_long;
5687 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5692 /** Initialize a cursor for a given transaction and database. */
5694 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5699 mc->mc_db = &txn->mt_dbs[dbi];
5700 mc->mc_dbx = &txn->mt_dbxs[dbi];
5701 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5706 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5708 mc->mc_xcursor = mx;
5709 mdb_xcursor_init0(mc);
5711 mc->mc_xcursor = NULL;
5713 if (*mc->mc_dbflag & DB_STALE) {
5714 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5719 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5722 MDB_xcursor *mx = NULL;
5723 size_t size = sizeof(MDB_cursor);
5725 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5728 /* Allow read access to the freelist */
5729 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5732 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5733 size += sizeof(MDB_xcursor);
5735 if ((mc = malloc(size)) != NULL) {
5736 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5737 mx = (MDB_xcursor *)(mc + 1);
5739 mdb_cursor_init(mc, txn, dbi, mx);
5740 if (txn->mt_cursors) {
5741 mc->mc_next = txn->mt_cursors[dbi];
5742 txn->mt_cursors[dbi] = mc;
5744 mc->mc_flags |= C_ALLOCD;
5755 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5757 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5760 if (txn->mt_cursors)
5763 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5767 /* Return the count of duplicate data items for the current key */
5769 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5773 if (mc == NULL || countp == NULL)
5776 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5779 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5780 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5783 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5786 *countp = mc->mc_xcursor->mx_db.md_entries;
5792 mdb_cursor_close(MDB_cursor *mc)
5795 /* remove from txn, if tracked */
5796 if (mc->mc_txn->mt_cursors) {
5797 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5798 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5800 *prev = mc->mc_next;
5802 if (mc->mc_flags & C_ALLOCD)
5808 mdb_cursor_txn(MDB_cursor *mc)
5810 if (!mc) return NULL;
5815 mdb_cursor_dbi(MDB_cursor *mc)
5821 /** Replace the key for a node with a new key.
5822 * @param[in] mp The page containing the node to operate on.
5823 * @param[in] indx The index of the node to operate on.
5824 * @param[in] key The new key to use.
5825 * @return 0 on success, non-zero on failure.
5828 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5835 indx_t ptr, i, numkeys, indx;
5838 indx = mc->mc_ki[mc->mc_top];
5839 mp = mc->mc_pg[mc->mc_top];
5840 node = NODEPTR(mp, indx);
5841 ptr = mp->mp_ptrs[indx];
5845 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5846 k2.mv_data = NODEKEY(node);
5847 k2.mv_size = node->mn_ksize;
5848 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5850 mdb_dkey(&k2, kbuf2),
5856 delta0 = delta = key->mv_size - node->mn_ksize;
5858 /* Must be 2-byte aligned. If new key is
5859 * shorter by 1, the shift will be skipped.
5861 delta += (delta & 1);
5863 if (delta > 0 && SIZELEFT(mp) < delta) {
5865 /* not enough space left, do a delete and split */
5866 DPRINTF("Not enough room, delta = %d, splitting...", delta);
5867 pgno = NODEPGNO(node);
5868 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5869 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
5872 numkeys = NUMKEYS(mp);
5873 for (i = 0; i < numkeys; i++) {
5874 if (mp->mp_ptrs[i] <= ptr)
5875 mp->mp_ptrs[i] -= delta;
5878 base = (char *)mp + mp->mp_upper;
5879 len = ptr - mp->mp_upper + NODESIZE;
5880 memmove(base - delta, base, len);
5881 mp->mp_upper -= delta;
5883 node = NODEPTR(mp, indx);
5886 /* But even if no shift was needed, update ksize */
5888 node->mn_ksize = key->mv_size;
5891 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5897 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
5899 /** Move a node from csrc to cdst.
5902 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5909 unsigned short flags;
5913 /* Mark src and dst as dirty. */
5914 if ((rc = mdb_page_touch(csrc)) ||
5915 (rc = mdb_page_touch(cdst)))
5918 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5919 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5920 key.mv_size = csrc->mc_db->md_pad;
5921 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5923 data.mv_data = NULL;
5927 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5928 assert(!((long)srcnode&1));
5929 srcpg = NODEPGNO(srcnode);
5930 flags = srcnode->mn_flags;
5931 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5932 unsigned int snum = csrc->mc_snum;
5934 /* must find the lowest key below src */
5935 mdb_page_search_root(csrc, NULL, 0);
5936 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5937 key.mv_size = csrc->mc_db->md_pad;
5938 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5940 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5941 key.mv_size = NODEKSZ(s2);
5942 key.mv_data = NODEKEY(s2);
5944 csrc->mc_snum = snum--;
5945 csrc->mc_top = snum;
5947 key.mv_size = NODEKSZ(srcnode);
5948 key.mv_data = NODEKEY(srcnode);
5950 data.mv_size = NODEDSZ(srcnode);
5951 data.mv_data = NODEDATA(srcnode);
5953 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5954 unsigned int snum = cdst->mc_snum;
5957 /* must find the lowest key below dst */
5958 mdb_page_search_root(cdst, NULL, 0);
5959 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5960 bkey.mv_size = cdst->mc_db->md_pad;
5961 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5963 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5964 bkey.mv_size = NODEKSZ(s2);
5965 bkey.mv_data = NODEKEY(s2);
5967 cdst->mc_snum = snum--;
5968 cdst->mc_top = snum;
5969 mdb_cursor_copy(cdst, &mn);
5971 rc = mdb_update_key(&mn, &bkey);
5976 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5977 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5978 csrc->mc_ki[csrc->mc_top],
5980 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5981 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5983 /* Add the node to the destination page.
5985 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5986 if (rc != MDB_SUCCESS)
5989 /* Delete the node from the source page.
5991 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5994 /* Adjust other cursors pointing to mp */
5995 MDB_cursor *m2, *m3;
5996 MDB_dbi dbi = csrc->mc_dbi;
5997 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5999 if (csrc->mc_flags & C_SUB)
6002 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6003 if (m2 == csrc) continue;
6004 if (csrc->mc_flags & C_SUB)
6005 m3 = &m2->mc_xcursor->mx_cursor;
6008 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6009 csrc->mc_ki[csrc->mc_top]) {
6010 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6011 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6016 /* Update the parent separators.
6018 if (csrc->mc_ki[csrc->mc_top] == 0) {
6019 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6020 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6021 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6023 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6024 key.mv_size = NODEKSZ(srcnode);
6025 key.mv_data = NODEKEY(srcnode);
6027 DPRINTF("update separator for source page %zu to [%s]",
6028 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6029 mdb_cursor_copy(csrc, &mn);
6032 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6035 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6037 indx_t ix = csrc->mc_ki[csrc->mc_top];
6038 nullkey.mv_size = 0;
6039 csrc->mc_ki[csrc->mc_top] = 0;
6040 rc = mdb_update_key(csrc, &nullkey);
6041 csrc->mc_ki[csrc->mc_top] = ix;
6042 assert(rc == MDB_SUCCESS);
6046 if (cdst->mc_ki[cdst->mc_top] == 0) {
6047 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6048 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6049 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6051 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6052 key.mv_size = NODEKSZ(srcnode);
6053 key.mv_data = NODEKEY(srcnode);
6055 DPRINTF("update separator for destination page %zu to [%s]",
6056 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6057 mdb_cursor_copy(cdst, &mn);
6060 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6063 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6065 indx_t ix = cdst->mc_ki[cdst->mc_top];
6066 nullkey.mv_size = 0;
6067 cdst->mc_ki[cdst->mc_top] = 0;
6068 rc = mdb_update_key(cdst, &nullkey);
6069 cdst->mc_ki[cdst->mc_top] = ix;
6070 assert(rc == MDB_SUCCESS);
6077 /** Merge one page into another.
6078 * The nodes from the page pointed to by \b csrc will
6079 * be copied to the page pointed to by \b cdst and then
6080 * the \b csrc page will be freed.
6081 * @param[in] csrc Cursor pointing to the source page.
6082 * @param[in] cdst Cursor pointing to the destination page.
6085 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6093 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6094 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6096 assert(csrc->mc_snum > 1); /* can't merge root page */
6097 assert(cdst->mc_snum > 1);
6099 /* Mark dst as dirty. */
6100 if ((rc = mdb_page_touch(cdst)))
6103 /* Move all nodes from src to dst.
6105 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6106 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6107 key.mv_size = csrc->mc_db->md_pad;
6108 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6109 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6110 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6111 if (rc != MDB_SUCCESS)
6113 key.mv_data = (char *)key.mv_data + key.mv_size;
6116 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6117 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6118 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6119 unsigned int snum = csrc->mc_snum;
6121 /* must find the lowest key below src */
6122 mdb_page_search_root(csrc, NULL, 0);
6123 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6124 key.mv_size = csrc->mc_db->md_pad;
6125 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6127 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6128 key.mv_size = NODEKSZ(s2);
6129 key.mv_data = NODEKEY(s2);
6131 csrc->mc_snum = snum--;
6132 csrc->mc_top = snum;
6134 key.mv_size = srcnode->mn_ksize;
6135 key.mv_data = NODEKEY(srcnode);
6138 data.mv_size = NODEDSZ(srcnode);
6139 data.mv_data = NODEDATA(srcnode);
6140 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6141 if (rc != MDB_SUCCESS)
6146 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6147 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);
6149 /* Unlink the src page from parent and add to free list.
6151 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6152 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6155 rc = mdb_update_key(csrc, &key);
6161 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6162 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6163 csrc->mc_db->md_leaf_pages--;
6165 csrc->mc_db->md_branch_pages--;
6167 /* Adjust other cursors pointing to mp */
6168 MDB_cursor *m2, *m3;
6169 MDB_dbi dbi = csrc->mc_dbi;
6170 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6172 if (csrc->mc_flags & C_SUB)
6175 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6176 if (csrc->mc_flags & C_SUB)
6177 m3 = &m2->mc_xcursor->mx_cursor;
6180 if (m3 == csrc) continue;
6181 if (m3->mc_snum < csrc->mc_snum) continue;
6182 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6183 m3->mc_pg[csrc->mc_top] = mp;
6184 m3->mc_ki[csrc->mc_top] += nkeys;
6188 mdb_cursor_pop(csrc);
6190 return mdb_rebalance(csrc);
6193 /** Copy the contents of a cursor.
6194 * @param[in] csrc The cursor to copy from.
6195 * @param[out] cdst The cursor to copy to.
6198 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6202 cdst->mc_txn = csrc->mc_txn;
6203 cdst->mc_dbi = csrc->mc_dbi;
6204 cdst->mc_db = csrc->mc_db;
6205 cdst->mc_dbx = csrc->mc_dbx;
6206 cdst->mc_snum = csrc->mc_snum;
6207 cdst->mc_top = csrc->mc_top;
6208 cdst->mc_flags = csrc->mc_flags;
6210 for (i=0; i<csrc->mc_snum; i++) {
6211 cdst->mc_pg[i] = csrc->mc_pg[i];
6212 cdst->mc_ki[i] = csrc->mc_ki[i];
6216 /** Rebalance the tree after a delete operation.
6217 * @param[in] mc Cursor pointing to the page where rebalancing
6219 * @return 0 on success, non-zero on failure.
6222 mdb_rebalance(MDB_cursor *mc)
6232 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6233 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6234 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6235 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6239 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
6242 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6243 DPRINTF("no need to rebalance page %zu, above fill threshold",
6249 if (mc->mc_snum < 2) {
6250 MDB_page *mp = mc->mc_pg[0];
6251 if (NUMKEYS(mp) == 0) {
6252 DPUTS("tree is completely empty");
6253 mc->mc_db->md_root = P_INVALID;
6254 mc->mc_db->md_depth = 0;
6255 mc->mc_db->md_leaf_pages = 0;
6256 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6260 /* Adjust other cursors pointing to mp */
6261 MDB_cursor *m2, *m3;
6262 MDB_dbi dbi = mc->mc_dbi;
6264 if (mc->mc_flags & C_SUB)
6267 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6268 if (m2 == mc) continue;
6269 if (mc->mc_flags & C_SUB)
6270 m3 = &m2->mc_xcursor->mx_cursor;
6273 if (m3->mc_snum < mc->mc_snum) continue;
6274 if (m3->mc_pg[0] == mp) {
6280 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6281 DPUTS("collapsing root page!");
6282 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6283 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6284 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6287 mc->mc_db->md_depth--;
6288 mc->mc_db->md_branch_pages--;
6290 /* Adjust other cursors pointing to mp */
6291 MDB_cursor *m2, *m3;
6292 MDB_dbi dbi = mc->mc_dbi;
6294 if (mc->mc_flags & C_SUB)
6297 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6298 if (m2 == mc) continue;
6299 if (mc->mc_flags & C_SUB)
6300 m3 = &m2->mc_xcursor->mx_cursor;
6303 if (m3->mc_snum < mc->mc_snum) continue;
6304 if (m3->mc_pg[0] == mp) {
6305 m3->mc_pg[0] = mc->mc_pg[0];
6310 DPUTS("root page doesn't need rebalancing");
6314 /* The parent (branch page) must have at least 2 pointers,
6315 * otherwise the tree is invalid.
6317 ptop = mc->mc_top-1;
6318 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6320 /* Leaf page fill factor is below the threshold.
6321 * Try to move keys from left or right neighbor, or
6322 * merge with a neighbor page.
6327 mdb_cursor_copy(mc, &mn);
6328 mn.mc_xcursor = NULL;
6330 if (mc->mc_ki[ptop] == 0) {
6331 /* We're the leftmost leaf in our parent.
6333 DPUTS("reading right neighbor");
6335 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6336 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6338 mn.mc_ki[mn.mc_top] = 0;
6339 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6341 /* There is at least one neighbor to the left.
6343 DPUTS("reading left neighbor");
6345 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6346 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6348 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6349 mc->mc_ki[mc->mc_top] = 0;
6352 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6353 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);
6355 /* If the neighbor page is above threshold and has at least two
6356 * keys, move one key from it.
6358 * Otherwise we should try to merge them.
6360 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6361 return mdb_node_move(&mn, mc);
6363 if (mc->mc_ki[ptop] == 0)
6364 rc = mdb_page_merge(&mn, mc);
6366 rc = mdb_page_merge(mc, &mn);
6367 mc->mc_flags &= ~C_INITIALIZED;
6372 /** Complete a delete operation started by #mdb_cursor_del(). */
6374 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6378 /* add overflow pages to free list */
6379 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6383 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6384 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6385 mc->mc_db->md_overflow_pages -= ovpages;
6386 for (i=0; i<ovpages; i++) {
6387 DPRINTF("freed ov page %zu", pg);
6388 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6392 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6393 mc->mc_db->md_entries--;
6394 rc = mdb_rebalance(mc);
6395 if (rc != MDB_SUCCESS)
6396 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6402 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6403 MDB_val *key, MDB_val *data)
6408 MDB_val rdata, *xdata;
6412 assert(key != NULL);
6414 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6416 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6419 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6423 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6427 mdb_cursor_init(&mc, txn, dbi, &mx);
6438 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6440 /* let mdb_page_split know about this cursor if needed:
6441 * delete will trigger a rebalance; if it needs to move
6442 * a node from one page to another, it will have to
6443 * update the parent's separator key(s). If the new sepkey
6444 * is larger than the current one, the parent page may
6445 * run out of space, triggering a split. We need this
6446 * cursor to be consistent until the end of the rebalance.
6448 mc.mc_next = txn->mt_cursors[dbi];
6449 txn->mt_cursors[dbi] = &mc;
6450 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6451 txn->mt_cursors[dbi] = mc.mc_next;
6456 /** Split a page and insert a new node.
6457 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6458 * The cursor will be updated to point to the actual page and index where
6459 * the node got inserted after the split.
6460 * @param[in] newkey The key for the newly inserted node.
6461 * @param[in] newdata The data for the newly inserted node.
6462 * @param[in] newpgno The page number, if the new node is a branch node.
6463 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6464 * @return 0 on success, non-zero on failure.
6467 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6468 unsigned int nflags)
6471 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6474 unsigned int i, j, split_indx, nkeys, pmax;
6476 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6478 MDB_page *mp, *rp, *pp;
6483 mp = mc->mc_pg[mc->mc_top];
6484 newindx = mc->mc_ki[mc->mc_top];
6486 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6487 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6488 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6490 /* Create a right sibling. */
6491 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6493 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6495 if (mc->mc_snum < 2) {
6496 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6498 /* shift current top to make room for new parent */
6499 mc->mc_pg[1] = mc->mc_pg[0];
6500 mc->mc_ki[1] = mc->mc_ki[0];
6503 mc->mc_db->md_root = pp->mp_pgno;
6504 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6505 mc->mc_db->md_depth++;
6508 /* Add left (implicit) pointer. */
6509 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6510 /* undo the pre-push */
6511 mc->mc_pg[0] = mc->mc_pg[1];
6512 mc->mc_ki[0] = mc->mc_ki[1];
6513 mc->mc_db->md_root = mp->mp_pgno;
6514 mc->mc_db->md_depth--;
6521 ptop = mc->mc_top-1;
6522 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6525 mc->mc_flags |= C_SPLITTING;
6526 mdb_cursor_copy(mc, &mn);
6527 mn.mc_pg[mn.mc_top] = rp;
6528 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6530 if (nflags & MDB_APPEND) {
6531 mn.mc_ki[mn.mc_top] = 0;
6533 split_indx = newindx;
6538 nkeys = NUMKEYS(mp);
6539 split_indx = nkeys / 2;
6540 if (newindx < split_indx)
6546 unsigned int lsize, rsize, ksize;
6547 /* Move half of the keys to the right sibling */
6549 x = mc->mc_ki[mc->mc_top] - split_indx;
6550 ksize = mc->mc_db->md_pad;
6551 split = LEAF2KEY(mp, split_indx, ksize);
6552 rsize = (nkeys - split_indx) * ksize;
6553 lsize = (nkeys - split_indx) * sizeof(indx_t);
6554 mp->mp_lower -= lsize;
6555 rp->mp_lower += lsize;
6556 mp->mp_upper += rsize - lsize;
6557 rp->mp_upper -= rsize - lsize;
6558 sepkey.mv_size = ksize;
6559 if (newindx == split_indx) {
6560 sepkey.mv_data = newkey->mv_data;
6562 sepkey.mv_data = split;
6565 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6566 memcpy(rp->mp_ptrs, split, rsize);
6567 sepkey.mv_data = rp->mp_ptrs;
6568 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6569 memcpy(ins, newkey->mv_data, ksize);
6570 mp->mp_lower += sizeof(indx_t);
6571 mp->mp_upper -= ksize - sizeof(indx_t);
6574 memcpy(rp->mp_ptrs, split, x * ksize);
6575 ins = LEAF2KEY(rp, x, ksize);
6576 memcpy(ins, newkey->mv_data, ksize);
6577 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6578 rp->mp_lower += sizeof(indx_t);
6579 rp->mp_upper -= ksize - sizeof(indx_t);
6580 mc->mc_ki[mc->mc_top] = x;
6581 mc->mc_pg[mc->mc_top] = rp;
6586 /* For leaf pages, check the split point based on what
6587 * fits where, since otherwise mdb_node_add can fail.
6589 * This check is only needed when the data items are
6590 * relatively large, such that being off by one will
6591 * make the difference between success or failure.
6593 * It's also relevant if a page happens to be laid out
6594 * such that one half of its nodes are all "small" and
6595 * the other half of its nodes are "large." If the new
6596 * item is also "large" and falls on the half with
6597 * "large" nodes, it also may not fit.
6600 unsigned int psize, nsize;
6601 /* Maximum free space in an empty page */
6602 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6603 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6604 if ((nkeys < 20) || (nsize > pmax/16)) {
6605 if (newindx <= split_indx) {
6608 for (i=0; i<split_indx; i++) {
6609 node = NODEPTR(mp, i);
6610 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6611 if (F_ISSET(node->mn_flags, F_BIGDATA))
6612 psize += sizeof(pgno_t);
6614 psize += NODEDSZ(node);
6618 split_indx = newindx;
6629 for (i=nkeys-1; i>=split_indx; i--) {
6630 node = NODEPTR(mp, i);
6631 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6632 if (F_ISSET(node->mn_flags, F_BIGDATA))
6633 psize += sizeof(pgno_t);
6635 psize += NODEDSZ(node);
6639 split_indx = newindx;
6650 /* First find the separating key between the split pages.
6651 * The case where newindx == split_indx is ambiguous; the
6652 * new item could go to the new page or stay on the original
6653 * page. If newpos == 1 it goes to the new page.
6655 if (newindx == split_indx && newpos) {
6656 sepkey.mv_size = newkey->mv_size;
6657 sepkey.mv_data = newkey->mv_data;
6659 node = NODEPTR(mp, split_indx);
6660 sepkey.mv_size = node->mn_ksize;
6661 sepkey.mv_data = NODEKEY(node);
6665 DPRINTF("separator is [%s]", DKEY(&sepkey));
6667 /* Copy separator key to the parent.
6669 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6673 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6676 if (mn.mc_snum == mc->mc_snum) {
6677 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6678 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6679 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6680 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6685 /* Right page might now have changed parent.
6686 * Check if left page also changed parent.
6688 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6689 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6690 for (i=0; i<ptop; i++) {
6691 mc->mc_pg[i] = mn.mc_pg[i];
6692 mc->mc_ki[i] = mn.mc_ki[i];
6694 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6695 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6699 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6702 mc->mc_flags ^= C_SPLITTING;
6703 if (rc != MDB_SUCCESS) {
6706 if (nflags & MDB_APPEND) {
6707 mc->mc_pg[mc->mc_top] = rp;
6708 mc->mc_ki[mc->mc_top] = 0;
6709 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6712 for (i=0; i<mc->mc_top; i++)
6713 mc->mc_ki[i] = mn.mc_ki[i];
6720 /* Move half of the keys to the right sibling. */
6722 /* grab a page to hold a temporary copy */
6723 copy = mdb_page_malloc(mc);
6727 copy->mp_pgno = mp->mp_pgno;
6728 copy->mp_flags = mp->mp_flags;
6729 copy->mp_lower = PAGEHDRSZ;
6730 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6731 mc->mc_pg[mc->mc_top] = copy;
6732 for (i = j = 0; i <= nkeys; j++) {
6733 if (i == split_indx) {
6734 /* Insert in right sibling. */
6735 /* Reset insert index for right sibling. */
6736 if (i != newindx || (newpos ^ ins_new)) {
6738 mc->mc_pg[mc->mc_top] = rp;
6742 if (i == newindx && !ins_new) {
6743 /* Insert the original entry that caused the split. */
6744 rkey.mv_data = newkey->mv_data;
6745 rkey.mv_size = newkey->mv_size;
6754 /* Update index for the new key. */
6755 mc->mc_ki[mc->mc_top] = j;
6756 } else if (i == nkeys) {
6759 node = NODEPTR(mp, i);
6760 rkey.mv_data = NODEKEY(node);
6761 rkey.mv_size = node->mn_ksize;
6763 xdata.mv_data = NODEDATA(node);
6764 xdata.mv_size = NODEDSZ(node);
6767 pgno = NODEPGNO(node);
6768 flags = node->mn_flags;
6773 if (!IS_LEAF(mp) && j == 0) {
6774 /* First branch index doesn't need key data. */
6778 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6782 nkeys = NUMKEYS(copy);
6783 for (i=0; i<nkeys; i++)
6784 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6785 mp->mp_lower = copy->mp_lower;
6786 mp->mp_upper = copy->mp_upper;
6787 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6788 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6790 /* reset back to original page */
6791 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6792 mc->mc_pg[mc->mc_top] = mp;
6793 if (nflags & MDB_RESERVE) {
6794 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6795 if (!(node->mn_flags & F_BIGDATA))
6796 newdata->mv_data = NODEDATA(node);
6802 /* return tmp page to freelist */
6803 mdb_page_free(mc->mc_txn->mt_env, copy);
6806 /* Adjust other cursors pointing to mp */
6807 MDB_cursor *m2, *m3;
6808 MDB_dbi dbi = mc->mc_dbi;
6809 int fixup = NUMKEYS(mp);
6811 if (mc->mc_flags & C_SUB)
6814 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6815 if (m2 == mc) continue;
6816 if (mc->mc_flags & C_SUB)
6817 m3 = &m2->mc_xcursor->mx_cursor;
6820 if (!(m3->mc_flags & C_INITIALIZED))
6822 if (m3->mc_flags & C_SPLITTING)
6827 for (k=m3->mc_top; k>=0; k--) {
6828 m3->mc_ki[k+1] = m3->mc_ki[k];
6829 m3->mc_pg[k+1] = m3->mc_pg[k];
6831 if (m3->mc_ki[0] >= split_indx) {
6836 m3->mc_pg[0] = mc->mc_pg[0];
6840 if (m3->mc_pg[mc->mc_top] == mp) {
6841 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6842 m3->mc_ki[mc->mc_top]++;
6843 if (m3->mc_ki[mc->mc_top] >= fixup) {
6844 m3->mc_pg[mc->mc_top] = rp;
6845 m3->mc_ki[mc->mc_top] -= fixup;
6846 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6848 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6849 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6858 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6859 MDB_val *key, MDB_val *data, unsigned int flags)
6864 assert(key != NULL);
6865 assert(data != NULL);
6867 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6870 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6874 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6878 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6881 mdb_cursor_init(&mc, txn, dbi, &mx);
6882 return mdb_cursor_put(&mc, key, data, flags);
6886 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6888 if ((flag & CHANGEABLE) != flag)
6891 env->me_flags |= flag;
6893 env->me_flags &= ~flag;
6898 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6903 *arg = env->me_flags;
6908 mdb_env_get_path(MDB_env *env, const char **arg)
6913 *arg = env->me_path;
6917 /** Common code for #mdb_stat() and #mdb_env_stat().
6918 * @param[in] env the environment to operate in.
6919 * @param[in] db the #MDB_db record containing the stats to return.
6920 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6921 * @return 0, this function always succeeds.
6924 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6926 arg->ms_psize = env->me_psize;
6927 arg->ms_depth = db->md_depth;
6928 arg->ms_branch_pages = db->md_branch_pages;
6929 arg->ms_leaf_pages = db->md_leaf_pages;
6930 arg->ms_overflow_pages = db->md_overflow_pages;
6931 arg->ms_entries = db->md_entries;
6936 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6940 if (env == NULL || arg == NULL)
6943 toggle = mdb_env_pick_meta(env);
6945 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6949 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
6953 if (env == NULL || arg == NULL)
6956 toggle = mdb_env_pick_meta(env);
6957 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
6958 arg->me_mapsize = env->me_mapsize;
6959 arg->me_maxreaders = env->me_maxreaders;
6960 arg->me_numreaders = env->me_numreaders;
6961 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
6962 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
6966 /** Set the default comparison functions for a database.
6967 * Called immediately after a database is opened to set the defaults.
6968 * The user can then override them with #mdb_set_compare() or
6969 * #mdb_set_dupsort().
6970 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6971 * @param[in] dbi A database handle returned by #mdb_dbi_open()
6974 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6976 uint16_t f = txn->mt_dbs[dbi].md_flags;
6978 txn->mt_dbxs[dbi].md_cmp =
6979 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6980 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6982 txn->mt_dbxs[dbi].md_dcmp =
6983 !(f & MDB_DUPSORT) ? 0 :
6984 ((f & MDB_INTEGERDUP)
6985 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6986 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6989 #define PERSISTENT_FLAGS 0xffff
6990 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
6991 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
6992 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6998 int rc, dbflag, exact;
6999 unsigned int unused = 0;
7002 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7003 mdb_default_cmp(txn, FREE_DBI);
7006 if ((flags & VALID_FLAGS) != flags)
7012 if (flags & PERSISTENT_FLAGS) {
7013 uint16_t f2 = flags & PERSISTENT_FLAGS;
7014 /* make sure flag changes get committed */
7015 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7016 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7017 txn->mt_flags |= MDB_TXN_DIRTY;
7020 mdb_default_cmp(txn, MAIN_DBI);
7024 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7025 mdb_default_cmp(txn, MAIN_DBI);
7028 /* Is the DB already open? */
7030 for (i=2; i<txn->mt_numdbs; i++) {
7031 if (!txn->mt_dbxs[i].md_name.mv_size) {
7032 /* Remember this free slot */
7033 if (!unused) unused = i;
7036 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7037 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7043 /* If no free slot and max hit, fail */
7044 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7045 return MDB_DBS_FULL;
7047 /* Find the DB info */
7051 key.mv_data = (void *)name;
7052 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7053 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7054 if (rc == MDB_SUCCESS) {
7055 /* make sure this is actually a DB */
7056 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7057 if (!(node->mn_flags & F_SUBDATA))
7059 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7060 /* Create if requested */
7062 data.mv_size = sizeof(MDB_db);
7063 data.mv_data = &dummy;
7064 memset(&dummy, 0, sizeof(dummy));
7065 dummy.md_root = P_INVALID;
7066 dummy.md_flags = flags & PERSISTENT_FLAGS;
7067 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7071 /* OK, got info, add to table */
7072 if (rc == MDB_SUCCESS) {
7073 unsigned int slot = unused ? unused : txn->mt_numdbs;
7074 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7075 txn->mt_dbxs[slot].md_name.mv_size = len;
7076 txn->mt_dbxs[slot].md_rel = NULL;
7077 txn->mt_dbflags[slot] = dbflag;
7078 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7080 txn->mt_env->me_dbflags[slot] = mdflags = txn->mt_dbs[slot].md_flags;
7081 mdb_default_cmp(txn, slot);
7084 txn->mt_env->me_numdbs++;
7086 /* Open the DB in parent txns as well */
7087 while ((txn = txn->mt_parent) != NULL) {
7088 txn->mt_dbflags[slot] = DB_STALE;
7089 txn->mt_dbs[slot].md_flags = mdflags;
7098 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7100 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7103 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7106 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7109 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
7111 ptr = env->me_dbxs[dbi].md_name.mv_data;
7112 env->me_dbxs[dbi].md_name.mv_data = NULL;
7113 env->me_dbxs[dbi].md_name.mv_size = 0;
7117 /** Add all the DB's pages to the free list.
7118 * @param[in] mc Cursor on the DB to free.
7119 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7120 * @return 0 on success, non-zero on failure.
7123 mdb_drop0(MDB_cursor *mc, int subs)
7127 rc = mdb_page_search(mc, NULL, 0);
7128 if (rc == MDB_SUCCESS) {
7133 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7134 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
7137 mdb_cursor_copy(mc, &mx);
7138 while (mc->mc_snum > 0) {
7139 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7140 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7141 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7142 if (ni->mn_flags & F_SUBDATA) {
7143 mdb_xcursor_init1(mc, ni);
7144 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7150 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7152 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7155 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7160 rc = mdb_cursor_sibling(mc, 1);
7162 /* no more siblings, go back to beginning
7163 * of previous level.
7166 for (i=1; i<mc->mc_top; i++)
7167 mc->mc_pg[i] = mx.mc_pg[i];
7171 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7172 mc->mc_db->md_root);
7177 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7182 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1)
7185 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7188 rc = mdb_cursor_open(txn, dbi, &mc);
7192 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7196 /* Can't delete the main DB */
7197 if (del && dbi > MAIN_DBI) {
7198 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7200 txn->mt_dbflags[dbi] = DB_STALE;
7201 mdb_dbi_close(txn->mt_env, dbi);
7204 /* reset the DB record, mark it dirty */
7205 txn->mt_dbflags[dbi] |= DB_DIRTY;
7206 txn->mt_dbs[dbi].md_depth = 0;
7207 txn->mt_dbs[dbi].md_branch_pages = 0;
7208 txn->mt_dbs[dbi].md_leaf_pages = 0;
7209 txn->mt_dbs[dbi].md_overflow_pages = 0;
7210 txn->mt_dbs[dbi].md_entries = 0;
7211 txn->mt_dbs[dbi].md_root = P_INVALID;
7213 if (!txn->mt_u.dirty_list[0].mid) {
7216 /* make sure we have at least one dirty page in this txn
7217 * otherwise these changes will be ignored.
7219 key.mv_size = sizeof(txnid_t);
7220 key.mv_data = &txn->mt_txnid;
7221 data.mv_size = sizeof(MDB_ID);
7222 data.mv_data = txn->mt_free_pgs;
7223 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
7224 rc = mdb_cursor_put(&m2, &key, &data, 0);
7228 mdb_cursor_close(mc);
7232 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7234 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7237 txn->mt_dbxs[dbi].md_cmp = cmp;
7241 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7243 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7246 txn->mt_dbxs[dbi].md_dcmp = cmp;
7250 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7252 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7255 txn->mt_dbxs[dbi].md_rel = rel;
7259 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7261 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7264 txn->mt_dbxs[dbi].md_relctx = ctx;